MLPerf Inference Rules

Table of Contents

1. Overview
- 1.1. Definitions (read this section carefully)
2. General rules
3. Scenarios
4. Benchmarks
- 4.1. Benchmarks
5. Load Generator
- 5.1. LoadGen Operation
6. Divisions
7. Data Sets
- 7.1. Pre- and post-processing
- 7.2. Test Data Traversal Order
8. Model
- 8.1. Weight Definition and Quantization
- 8.2. Model Equivalence
9. FAQ
Appendix A: Early Stopping Criterion
Appendix B: Datacenter Bandwidth Requirements
- B.1. Ingress Bandwidth
- B.2. Egress Bandwidth

1. Overview

This document describes how to implement one or more benchmarks in the MLPerf Inference Suite and how to use those implementations to measure the performance of an ML system performing inference.

There are separate rules for the submission, review, and publication process for all MLPerf benchmarks here.

The MLPerf name and logo are trademarks. In order to refer to a result using the MLPerf name, the result must conform to the letter and spirit of the rules specified in this document. The MLPerf organization reserves the right to solely determine if a use of its name or logo is acceptable.

1.1. Definitions (read this section carefully)

The following definitions are used throughout this document:

A sample is the unit on which inference is run. E.g., an image, or a sentence.

A query is a set of N samples that are issued to an inference system together. N is a positive integer. For example, a single query contains 8 images.

Quality always refers to a model’s ability to produce “correct” outputs.

A system under test consists of a defined set of hardware and software resources that will be measured for performance. The hardware resources may include processors, accelerators, memories, disks, and interconnect. The software resources may include an operating system, compilers, libraries, and drivers that significantly influences the running time of a benchmark.

A reference implementation is a specific implementation of a benchmark provided by the MLPerf organization. The reference implementation is the canonical implementation of a benchmark. All valid submissions of a benchmark must be equivalent to the reference implementation.

A run is a complete execution of a benchmark implementation on a system under the control of the load generator that consists of completing a set of inference queries, including data pre- and post-processing, meeting a latency requirement and a quality requirement in accordance with a scenario.

A run result consists of the scenario-specific metric.

2. General rules

The following rules apply to all benchmark implementations.

2.1. Strive to be fair

Benchmarking should be conducted to measure the framework and system performance as fairly as possible. Ethics and reputation matter.

2.2. System and framework must be consistent

The same system and framework must be used for a suite result or set of benchmark results reported in a single context.

2.3. Benchmark implementations must be shared

Source code used for the benchmark implementations must be open-sourced under a license that permits a commercial entity to freely use the implementation for benchmarking. The code must be available as long as the results are actively used.

2.4. Non-determinism is restricted

The only forms of acceptable non-determinism are:

Floating point operation order
Random traversal of the inputs
Rounding

All random numbers must be based on fixed random seeds and a deterministic random number generator. The deterministic random number generator is the Mersenne Twister 19937 generator (std::mt19937). The random seeds will be announced four weeks before the benchmark submission deadline.

2.5. Benchmark detection is not allowed

The framework and system should not detect and behave differently for benchmarks.

2.6. Input-based optimization is not allowed

The implementation should not encode any information about the content of the input dataset in any form.

2.7. Replicability is mandatory

Results that cannot be replicated are not valid results.

2.8. Audit Process

For audit process guidelines see [MLPerf Audit Guidelines.](MLPerf_Audit_Guidelines.adoc)

In each round, up to two submissions will be audited: one at random from all submissions, and either zero or one selected by the review committee. A "submission" for audit purposes shall denote a combination of a submitter and a platform (equivalent to a line in the results table). Only Available submissions in Closed division are auditable.

The process of random selection is in two stages: first a submitter is randomly chosen from all submitters with auditable submissions, then one of those submissions is randomly chosen. A submission is not a candidate for the randomly chosen audit if the system is equivalent to a system audited in the previous round. For the purposes of this rule, equivalent systems have the same CPU, NIC, accelerator, and accelerator count, with the same configuration of those components as per the system configuration JSON. For LoadGen Over Network submission the Networking must be the same. The review committee may determine that additional systems are equivalent to those audited in a previous round and exempt them from random audit. As a guidance for this exemption, if an accelerator is audited in one of the previous rounds, then the systems using the same accelerator can be excluded from random audit, if the aggregate system performance and the performance per accelerator are not more than 10% from those submitted during last audit time. For systems with power metrics, in addition to the performance, power efficiency must also be within 10% from the last audit time to be eligible for an exclusion from random audit. If any new result like a new model, an additional non-inferred scenario measurement or a new power measurement is submitted from the last audit time, then the exclusion is not applicable unless the review committee decides otherwise.

If a submitter chosen for an audit finds it unfair, they can appeal to the MLCommons Executive Director to ensure fairness.

During the review process, a github issue shall be opened where submitters can nominate systems for audit. Each nomination shall contain a reason, such as new HW or SW, unusual or interesting features, performance outside of expectations, etc. Review committee chairs evaluate the nominations and compile a list of systems at the end of the review period. Any systems with new accelerators are added to the list by the chairs if not nominated. The review committee will select a submission for audit by ranked choice voting using a simple majority. An option "No Selected Audit This Round" may be added if requested by a majority of the review committee.

An auditor shall be chosen by the review committee who has no conflict of interest with the submitter. The process of auditor selection will take no more than 28 days from selection of the submitter.

The burden is on the submitter to provide sufficient materials to demonstrate that the submission is compliant with the rules. Any such materials, including software, documentation, testing results and machine access will be provided to the auditor under NDA.

The submitter shall provide two days of hardware access, at a time mutually agreed with the auditor. The first day will be used to run a pre-agreed list of tests, and to verify other system parameters if needed. The second day will allow the auditor to run additional tests based on outcome of the first day.

The auditor shall write a report describing the work that was performed, a list of unresolved issues, and a recommendation on whether the submission is compliant.

The submitter will provide the auditor an NDA within seven days of the auditor’s selection. The auditor and submitter will negotiate and execute the NDA within 14 days of the auditor’s selection.

The auditor will submit their report to the submitter no more than thirty days after executing all relevant NDAs. The submitter will make any necessary redactions due to NDAs and forward the finalized report to the review committee within seven days. The auditor will confirm the accuracy of the forwarded report.

Submissions that fail the audit at a material level will be moved to open or removed, by review committee decision. If a submission failed an audit that was delayed past publication, then any published material concerning the invalidated result is subject to the MLCommons rules for Violation Determination, Remedies and Penalties for remedial action.

MLCommons shall retain a library of past audit reports and send copies to MLCommons members, auditors, and potential auditors by request. Audit reports will not be further distributed without permission from the audited submitter.

An audit is expected to be completed within a 90 day period. Audits failing to meet this timeline can be requested to be invalidated by the auditee. The final decision to accept such a request will be taken by the Working Group.

3. Scenarios

In order to enable representative testing of a wide variety of inference platforms and use cases, MLPerf has defined four different scenarios as described in the table below.

Scenario	Query Generation	Duration	Samples/query	Latency Constraint	Tail Latency	Performance Metric
Single stream	LoadGen sends next query as soon as SUT completes the previous query	600 seconds	1	None	90%*	90%-ile early-stopping latency estimate
Server/Interactive	LoadGen sends new queries to the SUT according to a Poisson distribution	600 seconds	1	Benchmark specific	99%*	Maximum Poisson throughput parameter supported
Offline	LoadGen sends all samples to the SUT at start in a single query	1 query and 600 seconds	At least 24,576**	None	N/A	Measured throughput
Multistream	Loadgen sends next query, as soon as SUT completes the previous query	600 seconds	8	None	99%*	99%-ile early-stopping latency estimate

- If the dataset used for the accuracy run of the benchmark task is of size less than 24,576 say N, then the Offline scenario query only needs to have at least N samples.

An early stopping criterion (described in more detail in Early Stopping Criterion) allows for runs with a relatively small number of processed queries to be valid, with the penalty that the effective computed percentile will be slightly higher. This penalty counteracts the increased variance inherent to runs with few queries, where there is a higher probability that a particular run will, by chance, report a lower latency than the system should reliably support.

In the above table, tail latency percentiles with an asterisk represent the theoretical lower limit of measured percentile for runs processing a very large number of queries. Submitters may opt to run for longer than the time listed in the "Duration" column, in order to decrease the effect of the early stopping penalty.

For all large language models (LLMs), submitters are NOT allowed to sort samples beyond the boundary of a dataset, regardless of the selected "duration" or "samples/query".

See the following table for a suggested starting point for how to set the minimum number of queries.

Tail Latency Percentile	Confidence Interval	Margin-of-Error	Inferences	Rounded Inferences
90%	99%	0.50%	23,886	3*2^13 = 24,576
95%	99%	0.25%	50,425	7*2^13 = 57,344
97%	99%	0.15%	85,811	11*2^13 = 90,112
99%	99%	0.05%	262,742	33*2^13 = 270,336

A submission may comprise any combination of benchmark and scenario results.

The number of runs required for each scenario is defined below:

Single Stream: 1
Server: 1
Offline: 1
Multistream: 1

Each sample has the following definition:

Model	definition of one sample
Resnet50-v1.5	one image
Retinanet	one image
3D UNET	one image
BERT	one sequence
DLRMv2	up to 700 user-item pairs (more details in FAQ)
SDXL	A pair of postive and negative prompts
Llama2-70b	one sequence
Mixtral-8x7B	one sequence
RGAT	one node id
Llama3.1-405B	one sequence
PointPainting	five images and one lidar point cloud
Llama3.1-8B	one sequence
DeepSeek-r1	one sequence
Whisper	one audio

4. Benchmarks

The MLPerf organization provides a reference implementation of each benchmark, which includes the following elements: Code that implements the model in a framework. A plain text “README.md” file that describes:

Problem
- Dataset/Environment
- Publication/Attribution
- Data pre- and post-processing
- Performance, accuracy, and calibration data sets
- Test data traversal order (CHECK)
Model
- Publication/Attribution
- List of layers
- Weights and biases
Quality and latency
- Quality target
- Latency target(s)
Directions
- Steps to configure machine
- Steps to download and verify data
- Steps to run and time

A “download_dataset” script that downloads the accuracy, speed, and calibration datasets.

A “verify_dataset” script that verifies the dataset against the checksum.

A “run_and_time” script that executes the benchmark and reports the wall-clock time.

4.1. Benchmarks

4.1.1. Constraints for the Closed division

The inference benchmark suite has two categories — Datacenter and Edge (defined herein as non-datacenter) systems. The suite has a carrying capacity of 10 benchmarks i.e at any point in time, the number of benchmarks will not exceed 10 in each category. The minimum requirements for a datacenter system are defined below:

Minimum requirements of a datacenter system

ECC

A Datacenter submission must use ECC in their DRAM and HBM memories, and ECC must be enabled for all performance and accuracy runs. No requirements are imposed on SRAM.

Networking (from the v3.0 round)

A Datacenter system must be equipped with all the necessary networking required by the system architecture described in the LoadGen Operation section. The details of the networking components must be described in the appropriate field of the System JSON. All necessary networking must be populated if power is measured along with performance.

The suites share multiple benchmarks, but characterize them with different requirements. Read the specifications carefully. The Datacenter suite includes the following benchmarks:

Area	Task	Model	Dataset	QSL Size	Quality	Server latency constraint
Vision	Object detection	Retinanet	OpenImages (800x800)	64	99% of FP32 (0.3755 mAP)	100 ms
Vision	Medical image segmentation	3D UNET	KiTS 2019	42	99% of FP32 and 99.9% of FP32 (0.86330 mean DICE score)	N/A
Language	Summarization	Llama3.1-8B	CNN Dailymail (v3.0.0, max_seq_len=2048)	13368	99% of FP32 and 99.9% of FP32 (rouge1=42.9865, rouge2=20.1235, rougeL=29.9881). Additionally, for both cases the total generation length of the texts should be more than 90% of the reference (gen_len=8167644)	Conversational category: TTFT/TPOT: 2000 ms/100 ms. Interactive category: TTFT/TPOT: 500 ms/30 ms. ^[1]
Language	Question Answering	Llama2-70b	OpenOrca (max_seq_len=1024)	24576	99% of FP32 and 99.9% of FP32 (rouge1=44.4312, rouge2=22.0352, rougeL=28.6162). Additionally, for both cases the generation length of the tokens per sample should be more than 90% of the reference (tokens_per_sample=294.45)	Conversational category: TTFT/TPOT: 2000 ms/200 ms. Interactive category: TTFT/TPOT: 450 ms/40 ms. ^[2]
Language	Text Generation	Llama3.1-405B	Subset of LongBench, LongDataCollections, Ruler, GovReport	8313	99% of FP16 ((GovReport + LongDataCollections + 65 Sample from LongBench)rougeL=21.6666, (Remaining samples of the dataset)exact_match=90.1335). Additionally, for both cases tokens per sample should be between than 90% and 110% of the reference (tokens_per_sample=684.68)	TTFT/TPOT: 6000 ms/175 ms Interactive category: TTFT/TPOT: 4500 ms/80 ms. ^[3]: 6000 ms/175 ms
Language	Text Generation (Question Answering, Math and Code Generation)	Mixtral-8x7B	OpenOrca (5k samples, max_seq_len=2048), GSM8K (5k samples of the train split, max_seq_len=2048), MBXP (5k samples, max_seq_len=2048)	15000	99% of FP16 ((OpenOrca)rouge1=45.5989, (OpenOrca)rouge2=23.3526, (OpenOrca)rougeL=30.4608, (gsm8k)Accuracy=73.66, (mbxp)Accuracy=60.16). Additionally, for both cases the tokens per sample should be between than 90% and 110% of the reference (tokens_per_sample=144.84)	TTFT/TPOT^[4]: 2000 ms/200 ms
Language	Reasoning	DeepSeek-r1	mlperf_deepseek_r1	4388	99% of FP16 (exact match 81.9132%).	TTFT/TPOT^[5]: 2000 ms/80 ms
Commerce	Recommendation	DLRMv2	Synthetic Multihot Criteo Dataset	204800	99% of FP32 and 99.9% of FP32 (AUC=80.31%)	60 ms
Generative	Text to image	SDXL	Subset of coco-2014 val	5000	FID range: [23.01085758, 23.95007626] and CLIP range: [31.68631873, 31.81331801]	20 s
Graph	Node classification	RGAT	IGBH	788379	99% of FP32 (72.86%)	N/A
Audio	Speech to Text	Whisper	LibriSpeech	1633	99% of FP32 and 99.9% of FP32 (WER=2.0671%)	N/A

Each Datacenter benchmark requires the following scenarios:

Area	Task	Required Scenarios
Vision	Image classification	Server, Offline
Vision	Object detection	Server, Offline
Vision	Medical image segmentation	Offline
Language	Summarization	(Server\| Interactive), Offline
Language	Question Answering	(Server\| Interactive), Offline
Language	Text Generation	(Server\| Interactive), Offline
Language	Reasoning	Server, Offline
Commerce	Recommendation	Server, Offline
Generative	Text to image	Server, Offline
Graph	Node classification	Offline
Audio	Speech to Text	Offline

The Edge suite includes the following benchmarks:

Area	Task	Model	Dataset	QSL Size	Quality
Vision	Image classification	Resnet50-v1.5	ImageNet (224x224)	1024	99% of FP32 (76.46%)
Vision	Object detection	Retinanet	OpenImages (800x800)	64	99% of FP32 (0.3755 mAP)
Vision	Medical image segmentation	3D UNET	KiTS 2019	42	99% of FP32 and 99.9% of FP32 (0.86330 mean DICE score)
Language	Summarization	Llama3.1-8B	CNN Dailymail (v3.0.0, max_seq_len=2048)	5000	99% of FP32 and 99.9% of FP32 (rouge1=42.9865, rouge2=20.1235, rougeL=29.9881). Additionally, for both cases the total generation length of the texts should be more than 90% of the reference (gen_len=8167644)
Language	Language processing	BERT	SQuAD v1.1 (max_seq_len=384)	10833	99% of FP32 and 99.9% of FP32(f1_score=90.874%)
Generative	Text to image	SDXL	Subset of coco-2014 val	5000	FID range: [23.01085758, 23.95007626] and CLIP range: [31.68631873, 31.81331801]
Automotive	3D Object Detection	PointPainting	Waymo Open Dataset	1024	99.9% of FP32 (0.5425 mAP)
Audio	Speech to Text	Whisper	LibriSpeech	1633	99% of FP32 and 99.9% of FP32 (WER=2.0671%)

Each Edge benchmark requires the following scenarios, and sometimes permit an optional scenario:

Area	Task	Required Scenarios
Vision	Image classification	Single Stream, Multistream, Offline
Vision	Object detection	Single Stream, Multistream, Offline
Vision	Medical image segmentation	Single Stream, Offline
Language	Language processing	Single Stream, Offline
Generative	Text to image	Single Stream, Offline
Language	Summarization	Single Stream, Offline
Automotive	3D object detection	Single Stream with 99.9% tail latency
Audio	Speech to Text	Offline

Edge submitters are allowed to infer a multistream result from single stream, and an offline result from either a single stream result or a measured multistream result, according to the following rules:

a multistream result inferred from a single stream result is 8 times the 99th percentile latency reported by loadgen. For example, if the single stream 99%th percentile latency is 25ms, the inferred multistream result is 200ms.
an offline result inferred from a multistream result is 8000 divided by the mean latency in milliseconds. For example, if the multistream result is 200ms, the inferred offline result is 40 img/s.
an offline result inferred from a single stream result is 1000 divided by the mean latency in milliseconds. For example, if the single stream result is 25ms, the inferred offline result is 40 img/s.

The accuracy of an inferred result will be the same as the result from which it was inferred. When inferring a metric for the power table, the measured power used to calculate the metric is the same as for the base result

To simplify automated processing of inferred results, the submitter should create copies of the directories for the inferred results under results/ and measurements/, named according to the inferred result (either offline or multistream).

Accuracy results must be reported to five significant figures with round to even. For example, 98.9995% should be recorded as 99.000%.

For performance runs, the LoadGen will select queries uniformly at random (with replacement) from a test set. The minimum size of the performance test set for each benchmark is listed as 'QSL Size' in the table above. However, the accuracy test must be run with one copy of the MLPerf specified validation dataset.

For 3DUNet, the logical destination for the benchmark output is considered to be the network.

4.1.2. Relaxed constraints for the Open division

An Open benchmark must perform a task matching an existing Closed benchmark, and be substitutable in LoadGen for that benchmark.
The validation dataset must be the same as used in an existing Closed benchmark, or must be pre-approved and added to the following list: ImageNet 2012 validation dataset for Image Classification; COCO 2017 validation dataset for Object Detection. When seeking such pre-approval, it is recommended that a potential submitter convincingly demonstrates the accuracy of the corresponding Closed model on the same validation dataset, which may involve retraining or finetuning the Closed model if required.
An Open benchmark must use all the available samples in the dataset, such that it must be the same as those used in an existing Closed benchmark.
Accuracy constraints are not applicable: instead the submission must report the accuracy obtained.
Latency constraints are not applicable: instead the submission must report the latency constraints under which the reported performance was obtained.
Scenario constraints are not applicable: any combination of scenarios is permitted.
A open submission must be classified as "Available", "Preview", or "Research, Development, or Internal".
The model can be of any origin (trained on any dataset, except the validation dataset; quantized in any way; sparsified in any way).

4.1.3. Additional inference parameters

For each of the following benchmarks it is necessary to use the following inference parameters in the closed division

Benchmark	Parameter	Value	Explanation
Summarization (Llama3.1-8B)	min_new_tokens	1	Minimun number of new tokens to generate
Summarization (Llama3.1-8B)	max_new_tokens	128	Maximum number of new tokens to generate
Summarization (Llama2-70b)	max_new_tokens	1024	Maximum number of new tokens to generate
Text Generation (Llama3.1-405B)	min_new_tokens	2	Minimun number of new tokens to generate
Text Generation (Llama3.1-405B)	max_new_tokens	20000	Maximum number of new tokens to generate
Summarization (Mixtral-8x7B)	min_new_tokens	2	Minimun number of new tokens to generate
Text Generation (Mixtral-8x7B)	max_new_tokens	1024	Maximum number of new tokens to generate
Reasoning (DeepSeek)	min_new_tokens	1	Minimun number of new tokens to generate
Reasoning (DeepSeek)	max_new_tokens	32768	Maximum number of new tokens to generate

5. Load Generator

5.1. LoadGen Operation

The LoadGen is provided in C++ with Python bindings and must be used by all submissions. The LoadGen is responsible for:

Generating the queries according to one of the scenarios.
Tracking the latency of queries.
Validating the accuracy of the results.
Computing final metrics.

Latency is defined as the time from when the LoadGen was scheduled to pass a query to the SUT, to the time it receives a reply.

Single Stream: LoadGen measures the 90th percentile early-stopping latency estimate using a single test run. For the test run, LoadGen sends an initial query then continually sends the next query as soon as the previous query is processed.
Server: LoadGen determines the system throughput using multiple test runs. Each test run evaluates a specific throughput value in queries-per-second (QPS). For a specific throughput value, queries are generated at that QPS using a Poisson distribution. LoadGen will use a binary search to find a candidate value. If a run fails, it will reduce the value by a small delta then try again.
Offline: LoadGen measures throughput using a single test run. For the test run, LoadGen sends all samples at once in a single query.
Multistream: LoadGen measures the 99th percentile early-stopping latency estimate using a single test run. For the test run, LoadGen sends an initial query then continually sends the next query as soon as the previous query is processed.

The run procedure is as follows:

LoadGen signals system under test (SUT).
SUT starts up and signals readiness.
LoadGen starts clock and begins generating queries.
LoadGen stops generating queries as soon as the benchmark-specific minimum time has elapsed, and the (optional, submitter-selected) minimum number of queries have been generated.
LoadGen waits for all queries to complete, and errors if all queries fail to complete.
LoadGen computes metrics for the run.

The execution of LoadGen is restricted as follows:

LoadGen must run on the processor that most faithfully simulates queries arriving from the most logical source, which is usually the network or an I/O device such as a camera. For example, if the most logical source is the network and the system is characterized as host - accelerator, then LoadGen should run on the host unless the accelerator incorporates a NIC.
The trace generated by LoadGen must be stored in the DRAM that most faithfully simulates queries arriving from the most logical source, which is usually the network or an I/O device such as a camera. It may be pinned. Similarly, the response provided to Loadgen must be stored in the DRAM that most faithfully simulates transfer to the most logical destination, which is a CPU process unless otherwise specified for the benchmark. From 4.0, submitters must provide with their submission sufficient details of the system architecture and software to show how the I/O bandwidth utilized by each benchmark/scenario combination can be transferred between the memory where the trace is stored and the network or I/O device. Minimum bandwidths for each benchmark can be found in Datacenter Bandwidth Requirements. All components mentioned in the system architecture must be present in the system during the run. A system architecture description must be provided along with the submission, which must include:
- Bandwidth of each NIC and total number of NIC(s)
- Description of the data path from the NIC(s) to the accelerator(s)
- Specifications or measurements indicating that the path from the NIC to the memory in which loadgen data resides can sustain the required bandwidth
Caching values derived from the shapes of input tensors is allowed. Caching of any other queries, query parameters, or intermediate results is prohibited. In particular, caching values derived from activations is prohibited.
The LoadGen must be compiled from a tagged approved revision of the mlperf/inference GitHub repository without alteration. Pull requests addressing portability issues and adding new functionality are welcome.

LoadGen generates queries based on trace. The trace is constructed by uniformly sampling (with replacement) from a library based on a fixed random seed and deterministic generator. The size of the library is listed in as 'QSL Size' in the 'Benchmarks' table above. The trace is usually pre-generated, but may optionally be incrementally generated if it does not fit in memory. LoadGen validates accuracy via a separate test run that use each sample in the test library exactly once but is otherwise identical to the above normal metric run.

One LoadGen validation run is required for each submitted performance result even if two or more performance results share the same source code.

Note: The same code must be run for both the accuracy and performance LoadGen modes. This means the same output should be passed in QuerySampleComplete in both modes.

6. Divisions

There are three divisions of the benchmark suite, the Closed division, the Network division, and the Open division.

6.1. Closed Division

The Closed division requires using pre-processing, post-processing, and model that is equivalent to the reference or alternative implementation. The closed division allows calibration for quantization and does not allow any retraining.

The unqualified name “MLPerf” must be used when referring to a Closed Division suite result, e.g. “a MLPerf result of 4.5.”

6.2. Network Division

The Network division inherits all requirements from the Closed division and imposes further constraints. In the Network division the SUT is connected to the Loadgen system over a network fabric. The Query Dispatch Library (QDL) component is a submitter-implemented SUT proxy that runs on the Loadgen system. The Network division supports only the Datacenter suite. Non-conforming network submission should be submitted to Open category, under the Open category constraints.

6.2.1. QDL Constraints

The QDL is not allowed to do any pre-processing. e.g. changing of precision, or data layout.
The QDL is not allowed to do any post-processing of the responses, e.g. gather, reduction, or ArgMax.
If an SUT compresses its output, the QDL must decompress the output. Decompression is a timed operation. No other post-processing in the QDL is allowed.
The QDL is not allowed to batch queries.
The QDL is not allowed to pad the data in queries.
The QDL is not allowed to cache queries or responses.
The QDL is implementing the network function of the LoadGen Node towards the SUT node and handles the required processing. E.G. padding of the payload as required by the network protocol.
The QDL should reflect a single SUT to the LoadGen. LoadGen operates with a single SUT.
The Name method’s return value must contain the substring "Network SUT".
The Name method’s implementation must include at least one round trip over the network. The Name method must not return until the round trip is complete.
The QDL must query each SUT Node for its name and aggregate the responses in the Name Method. Each SUT Node must have a unique name.

The submission must include source code for the QDL implementation above the level of the OSI session layer (RPC or equivalent), and sufficient documentation of the session layer API that a reader of that code can understand what data is being marshalled and sent over the network for each query.

6.2.2. General Constraints

MLPerf distinguishes between fabric interconnects and bus interconnects. Fabric interconnects are required and bus interconnects are forbidden.

A fabric interconnect must:

Work as out-of-the-box chassis-to-chassis interconnect
Use wireless, copper, or fiber-optics media
Suitable for connecting systems above 10 meters distance
Use switch topology
Be highly scalable, reliable, and fault-tolerant

Currently permitted fabric interconnects are Ethernet, IEEE802.11, Infiniband, and 3GPP.

Examples of forbidden bus interconnect include: PCIe/CXL/CCIX, Hypertransport, NVLink, QPI, UPI, and ICI (interchip interconnect).

Additionally, any interconnect not listed in the permitted list is forbidden unless clearance is first obtained from the MLPerf Inference WG.

System Topology: The SUT and QDL must run on physically separate and distinct systems. The SUT can contain multiple Nodes. The systems can be connected point to point or through network elements like switches.

Fabric and protocol must be reported in the submission metadata. Submission metadata must be sufficient to determine OSI layers one through four of the submission’s network stack.

6.2.3. SUT over the Network and Setup Constraints

SUT parameters and configuration must be uniquely and specifically named in the submission results.
Everything outside the LoadGen node should be considered as part of the SUT, for instance for counting power and latency. As an example, components outside the nodes like a switch or load balancer should be considered part of the SUT.
All queries must be transferred over the network, carrying the inference data, for inference execution at the SUT. All responses must be transferred back over the network, carrying the inference responses.
Caching/Storing of the queries and inference data or responses for further use at the SUT is disallowed. It is allowed to cache/store other data like Neural Network weights or Neural Network executable.
SUT can do the required pre-processing of the data, e.g. Batching, Padding, processing of the requests (precision, data layout), compression, decompression. SUT can do the required post processing functions e.g. gather, reduction or ArgMax.
The report must contain network interface characteristics for both the Loadgen and SUT systems, and every other component through which data passes between Loadgen and SUT. The information must be sufficient for reproducibility.
A system diagram must be included in the submission that shows how the components between the LoadGen node and the SUT nodes are connected, accompanied by any text necessary for another submitter to understand the diagram.
For "Available" submissions, for reproducibility, it is required to specify software version of all components, hardware configurations, software stacks, dockers, and settings of all components and stacks.

6.2.4. Benchmarks and QSL Preprocessing

Data formats for inputs and outputs are allowed to be compressed for network transmission, providing a tradeoff between compute and network bandwidth. Data transferred between the LoadGen system and the SUT can be compressed using one of the options from the following table for each benchmark. Compression is performed by QSL, and is untimed. The compression scheme needs approval by the Working Group, allowing compression schemes that will be suitable for production, so for example, very asymmetric schemes are not expected to be approved.

Area	Task	Model	QSL side PreProcessing^(1,2,3)
Vision	Image classification	Resnet50-v1.5	Allow one of the following compression options for pre-processing: 1) No compression 2) Lossless compression 3) The original compression of the dataset (JPEG)
Vision	Object detection (large)	Retinanet	Allow one of the following compression options for pre-processing: 1) No compression 2) Lossless compression 3) The original compression of the dataset (For the Coco dataset JPEG, for Open Images JPEG)
Vision	Medical image segmentation	3D UNET	Allow one of the following compression options: 1) No compression 2) Lossless compression This rule applies both for the QSL pre-processing and for post-processing function allowed in QDL for this benchmark results.
Language	Language processing	BERT-large	Input is either Token IDs, Input Masks and Segment IDs or just the Token IDs (generating the other tensors at the SUT in a timed operation). 1) No compression 2) Lossless compression
Language	Summarization	Llama3.1-8B	Input is either Token IDs, Input Masks and Input Lengths or just the Token IDs (the other tensors are generated at the SUT in a timed operation). No compression allowed.
Language	Question Answering	Llama2-70b	Input is either Token IDs, Input Masks and Input Lengths or just the Token IDs (the other tensors are generated at the SUT in a timed operation). No compression allowed.
Language	Question Answering	Llama3.1-405B	Input is either Token IDs, Input Masks and Input Lengths or just the Token IDs (the other tensors are generated at the SUT in a timed operation). No compression allowed.
Language	Text Generation	Mixtral-8x7B	Input is either Token IDs, Input Masks and Input Lengths or just the Token IDs (the other tensors are generated at the SUT in a timed operation). No compression allowed.
Language	Reasoning	DeepSeek-r1	Input is either Token IDs, Input Masks and Input Lengths or just the Token IDs (the other tensors are generated at the SUT in a timed operation). No compression allowed.
Commerce	Recommendation	DLRMv2	QDL sends query (Batch of samples). Allow one of the following compression options for pre-processing: 1) No compression 2) Lossless compression Allow any lossless compression that will be suitable for production use. In Server mode allow per-Query compression.
Generative	Text to image	SDXL	No compression allowed.
Graph	Node Classification	RGAT	No compression allowed.

Compression scheme needs pre-approval, at least two weeks before a submission deadline.
A compression scheme may use information from the training set, but not the validation set (ex: check index probability).
Only per-Sample compression is allowed, except for DLRMv2 Server mode where per-Query compression is allowed.

6.3. Open Division

The Open division allows using arbitrary pre- or post-processing and model, including retraining. The qualified name “MLPerf Open” must be used when referring to an Open Division suite result, e.g. “a MLPerf Open result of 7.2.”

Restricted retraining rules characterize a subset of Open division retraining possibilities that are expected to be straightforward for customers to use. The restrictions are optional; conformance will be indicated by a tag on the submission.

7. Data Sets

For each benchmark, MLPerf will provide pointers to:

An accuracy data set, to be used to determine whether a submission meets the quality target, and used as a validation set
A speed/performance data set that is a subset of the accuracy data set to be used to measure performance

For each benchmark, MLPerf will provide pointers to:

A calibration data set, to be used for quantization (see quantization section), that is a small subset of the training data set used to generate the weights

Each reference implementation shall include a script to verify the datasets using a checksum. The dataset must be unchanged at the start of each run.

7.1. Pre- and post-processing

As input, before preprocessing:

all imaging benchmarks take uncropped uncompressed bitmap
BERT, Llama2-70b, Llama3.1-405B, Mixtral-8x7B, Llama3.1-8B and DeepSeek-r1 take texts
RNN-T and Whisper takes a waveform
DLRMv2 takes a variable sized set of items, each a sequence of embedding indices

Sample-independent pre-processing that matches the reference model is untimed. However, it must be pre-approved and added to the following list:

May resize to processed size (e.g. SSD-large)
May reorder channels / do arbitrary transpositions
May pad to arbitrary size (don’t be creative)
May do a single, consistent crop
Mean subtraction and normalization provided reference model expect those to be done
May convert data among numerical formats
May convert to token ids from texts using the reference tokenizer

Any other pre- and post-processing time is included in the wall-clock time for a run result.

7.2. Test Data Traversal Order

Test data is determined by the LoadGen. For scenarios where processing multiple samples can occur (i.e., and offline), any ordering is allowed subject to latency requirements.

8. Model

CLOSED: MLPerf provides a reference implementation of each benchmark. The benchmark implementation must use a model that is equivalent, as defined in these rules, to the model used in the reference implementation.

OPEN: The benchmark implementation may use a different model to perform the same task. Retraining is allowed.

8.1. Weight Definition and Quantization

CLOSED: MLPerf will provide trained weights and biases in fp16/fp32 format for both the reference and alternative implementations.

MLPerf will provide a calibration data set for all models. Submitters may do arbitrary purely mathematical, reproducible quantization using only the calibration data and weight and bias tensors from the benchmark owner provided model to any numerical format that achieves the desired quality. The quantization method must be publicly described at a level where it could be reproduced.

To be considered principled, the description of the quantization method must be much much smaller than the non-zero weights it produces.

Calibration is allowed and must only use the calibration data set provided by the benchmark owner. Submitters may choose to use only a subset of the calibration data set.

Additionally, MLPerf may provide an INT8 reference for some models. Model weights and input activations are scaled per tensor, and must preserve the same shape modulo padding. Convolution layers are allowed to be in either NCHW or NHWC format. No other retraining is allowed.

OPEN: Weights and biases must be initialized to the same values for each run, any quantization scheme is allowed that achieves the desired quality.

8.2. Model Equivalence

All implementations are allowed as long as the latency and accuracy bounds are met and the reference weights are used. Reference weights may be modified according to the quantization rules.

Examples of allowed techniques include, but are not limited to:

Arbitrary frameworks and runtimes: TensorFlow, TensorFlow-lite, ONNX, PyTorch, etc, provided they conform to the rest of the rules
Running any given control flow or operations on or off an accelerator
Arbitrary data arrangement
Different in-memory representations of inputs, weights, activations, and outputs
Variation in matrix-multiplication or convolution algorithm provided the algorithm produces asymptotically accurate results when evaluated with asymptotic precision
Mathematically equivalent transformations (e.g. Tanh versus Logistic, ReluX versus ReluY, any linear transformation of an activation function)
Approximations (e.g. replacing a transcendental function with a polynomial)
Processing queries out-of-order within discretion provided by scenario
Replacing dense operations with mathematically equivalent sparse operations
Hand picking different numerical precisions for different operations
Fusing or unfusing operations
Dynamically switching between one or more batch sizes
Different implementations based on scenario (e.g., single stream vs. offline) or dynamically determined batch size or input size
Mixture of experts combining differently quantized weights
Stochastic quantization algorithms with seeds for reproducibility
Reducing ImageNet classifiers with 1001 classes to 1000 classes
Dead code elimination
Sorting samples in a query when it improves performance even when all samples are distinct
Incorporating explicit statistical information about the calibration set (eg. min, max, mean, distribution)
Empirical performance and accuracy tuning based on the performance and accuracy set (eg. selecting batch sizes or numerics experimentally)
Sorting an embedding table based on frequency of access in the training set. (Submitters should include in their submission details of how the ordering was derived.)

The following techniques are disallowed:

Wholesale weight replacement or supplements
Discarding non-zero weight elements, including pruning
Caching queries or responses
Coalescing identical queries
Modifying weights during the timed portion of an inference run (no online learning or related techniques)
Weight quantization algorithms that are similar in size to the non-zero weights they produce
Hard coding the total number of queries
Techniques that boost performance for fixed length experiments but are inapplicable to long-running services except in the offline scenario
Using knowledge of the LoadGen implementation to predict upcoming lulls or spikes in the server scenario
Treating beams in a beam search differently. For example, employing different precision for different beams
Changing the number of beams per beam search relative to the reference
Incorporating explicit statistical information about the performance or accuracy sets (eg. min, max, mean, distribution)
Techniques that take advantage of upsampled images. For example, downsampling inputs and kernels for the first convolution.
Techniques that only improve performance when there are identical samples in a query. For example, sorting samples in SSD and R-GAT.
Techniques that only improve performance by artificially creating batches of identical samples. Sorting samples beyond the boundary of a dataset is prohibited for all LLM workloads such as - Llama2-70b, Llama2-70b-Interactive, Llama3.1-405B, DeepSeek-r1, Llama3.1-8B and Mixtral-8x7B.
Speculative decoding for auto-generative language models (i.e. using a smaller model to predict the next token for the reference model).
Modifying the sample duration loaded by a fixed input-length audio model (ex. a shorter audio sample should be padded to match the fixed input-length).

9. FAQ

Q: Do I have to use the reference implementation framework?

A: No, you can use another framework provided that it matches the reference in the required areas.

Q: Do I have to use the reference implementation scripts?

A: No, you don’t have to use the reference scripts. The reference is there to settle conformance questions - with a few exceptions, a submission to the closed division must match what the reference is doing.

Q: Can I submit a single benchmark (e.g., object detection) in a suite (e.g., data center), or do I have to submit all benchmarks?

A: You can submit any of the benchmarks that are interesting, from just one benchmark to the entire set of benchmarks. Keep in mind that submitting one benchmark typically requires running several scenarios as described in Section 4. For example, submitting object detection in the data center suite requires the server and offline scenario and submitting object detection in the edge suite requires the single stream and offline scenarios.

Q: Why does a run require so many individual inference queries?

A: The numbers were selected to be sufficiently large to statistically verify that the system meets the latency requirements.

Q: For my submission, I am going to use a different model format (e.g., ONNX vs TensorFlow Lite). Should the conversion routine/script be included in the submission? Or is it sufficient to submit the converted model?

A: The goal is reproducibility, so you should include the conversion routine/scripts.

Q: Is it permissible to exceed both the minimum number of queries and minimum time duration in a valid test run?

A: Yes. However, for all large language models (LLMs), submitters are NOT allowed to sort samples beyond the boundary of a dataset, regardless of the selected "duration" or "samples/query".

Q: Can we give the driver a hint to preload the image data to somewhere closer to the accelerator?

A: No.

Q: Can we preload image data somewhere closer to the accelerator that is mapped into host memory?

A: No.

Q: Can we preload image data in host memory somewhere that is mapped into accelerator memory?

A: Yes, provided the image data isn’t eventually cached on the device.

Q: For the server scenario, there are 'Scheduled samples per second', 'Completed samples per second', and the user input target QPS. Which one is reported as the final metric?

A: Scheduled samples per second

Q: What can I cache based on the query indices?

A: Query indices are an artifact of using a finite set of samples to represent an infinite set, and would have no counterpart in production scenarios. As such, the system under test should not cache any information associated with query indices.

9.1. SSD

Q: Is non-maximal suppression (NMS) timed?

A: Yes. NMS is a per image operation. NMS is used to make sure that in object detection, a particular object is identified only once. Production systems need NMS to ensure high-quality inference.

Q: Is COCO eval timed?

A: No. COCO eval compares the proposed boxes and classes in all the images against ground truth in COCO dataset. COCO eval is not possible in production.

9.2. Softmax

Q: In classification and segmentation models (ResNet50, 3DUNet) the final softmax does not change the order of class probabilities. Can it be omitted?

A: Yes.

9.3. DLRMv2

Q: For DLRMv2, what’s the distribution of user-item pairs per sample for all scenarios?

A: For all scenarios, the distribution of user-item pairs per sample is specified by dist_quantile.txt. To verify that your sample aggregation trace matches the reference, please follow the steps in dist_trace_verification.txt. Or simply download the reference dlrm_trace_of_aggregated_samples.txt from Zenodo (MD5:3db90209564316f2506c99cc994ad0b2).

Q: What is dist_trace_verification.txt?

The benchmark provides a pre-defined quantile distribution in ./tools/dist_quantile.txt from which the samples will be drawn using the inverse transform algorithm. This algorithm relies on randomly drawn numbers from the interval [0,1) and that depend on the --numpy-rand-seed, which specific value will be provided shortly before MLPerf inference submissions.

Q: What is the rational for the distribution of user-item pairs?

In the case of DLRMv2 we have agreed that we should use multiple samples drawn from a distribution, similar to the one shown on Fig. 5: "Queries for personalized recommendation models" in the DeepRecSys paper.

Q: Generating dlrm_trace_of_aggregated_samples.txt uses a pseudo-random number generator. How can submitters verify their system pseudo-random number generator is compatible?

Submitters can verify their compatibility by using the default --numpy-rand-seed and comparing the trace generated on their system with ./tools/dist_trace_verification.txt using the following command

./run_local.sh pytorch dlrm terabyte cpu --count-samples=100 --scenario Offline --max-ind-range=40000000 --samples-to-aggregate-quantile-file=./tools/dist_quantile.txt --max-batchsize=128

Q: I understand that --samples-to-aggregate-quantile-file=./tools/dist_quantile.txt is the only compliant setting for MLPerf, but what are the alternative settings and what do they do?

The DLRMv2 MLPerf inference code has an option to aggregate multiple consecutive samples together into a single aggregated sample. The number of samples to be aggregated can be selected using either of the following options

fixed [--samples-to-aggregate-fix]
drawn uniformly from interval [--samples-to-aggregate-min, --samples-to-aggregate-max]
drawn from a custom distribution, with its quantile (inverse of CDP) specified in --samples-to-aggregate-quantile-file=./tools/dist_quantile.txt.

9.4. LLM Benchmarks

Q: Are submitters allowed to change the maximum output token length to be less than the benchmark reference? No, submitter implementations must adhere to the specificied maximum output token length as in the benchmark reference. Truncating output tokens to boost performance or meet accuracy is not permitted.

Q: What algorithm is used for the auto-regressive decoding loop?

A: The algorithms used by the benchmarks (greedy search and beam search) are described at a high level here: https://huggingface.co/blog/how-to-generate. Specifically, Llama2-70b, Llama3.1-405B, Mixtral-8x7B, DeepSeek-r1 and Llama3.1-8B use greedy search.

Q: MLPerf disallows caching queries. Is using a KV-cache in decoding allowed?

A: Using a KV-cache is allowed in the same way as it is included in the reference model, where it does not apply across queries. A KV-cache row is a tensor that must handle the execution of a single inference query. When a KV-cache is used, every input query must be computed in its entirety.

Q: Is it allowed to not use a KV-cache or use it partially?

A: Yes, KV-cache is an optional optimization. It is not required to use a KV-cache, but if you do, your implementation must adhere to the reference implementation. If you do not use a KV-cache, the corresponding values must be rematerialized during the decoding process.

Q: Is it allowed to store continuous keys and values in non-contiguous memory space for the KV-cache, i.e. PagedAttention?

A: Yes, it is allowed as long as the KV-cache block is reused only within the batch of queries. A high level explanation of PagedAttention can be found here: https://blog.vllm.ai/2023/06/20/vllm.html.

Q: How does quantization and pruning apply to the KV-cache?

A: The entries of the KV-cache should be handled in the same way as the activations of a forward pass. They can be quantized according to the quantization rules. However, according to the model equivalence rules, they cannot be pruned (or sparsified). It should be noted that pruning is different from not using a KV-cache (or caching only some entries while rematerializing others); pruning alters the computation and the model’s predictions.

Q: How does query batching affect the KV-cache usage?

A: The size of the KV-cache is determined by the batch size. The KV-cache size can also be cached across queries, in accordance with the rule of allowing caching of sizes and shapes.

Q: Is it allowed to apply continuous batching (or dynamic batching) for auto-generative benchmarks?

A: Yes. Continuous batching is explained at a high level here: https://www.anyscale.com/blog/continuous-batching-llm-inference.

9.5. RGAT

Q: Is loading the node neighbors a timed operation?

A: Yes, this is the main operation of this benchmark

9.6. Whisper

Q: Whisper’s native audio input duration is fixed at 30 seconds. Is it permitted to modify the loaded duration to match the sample’s specific duration?

A: No, it is not permitted to modify the loaded duration even if continuing to meet the model’s accuracy threshold. Samples must be zero-padded to ensure consistent computation and accuracy criteria.

9.7. Audit

Q: What characteristics of my submission will make it more likely to be audited?

A: A submission is more likely to be audited if:

the submission’s performance is not consistent with the known or expected characteristics of the hardware
the review committee lacks insight into how the measured performance was achieved
the hardware and software is not reasonably available to the general public

Q: What should I be expected to provide for audit?

A: You should expect to provide the following:

An explanation of the hardware and software mechanisms required to achieve the measured performance
Hardware access to enable the auditor to replicate submission runs (or partial runs in the case of very long-running submission)
Hardware access to enable performance tests through the APIs used in the submission, to verify that performance-critical elements perform as claimed

The auditor may also request source code access to binary elements of the submission software. Where information or access is not provided, the auditor’s report will list the issues that could not be resolved.

Q: Is it expected that an audit will be concluded during the review period? A: No. We should try to finish the audit before the publication date.

Appendix A: Early Stopping Criterion

The early stopping criterion allows for systems to process a smaller number of queries during their runs than previously allowed. In particular, given a desired tail latency p, tolerance d, and confidence c, we determine the required number of queries to process as a function of the number of seen overlatency queries. If we have processed at least this many queries, we are able to stop processing queries early. See the final section of this appendix for a more detailed description of the algorithm.

A.1. Motivating Example

Processing more queries allows us to better estimate the percentage of the time a system passes a given latency bound, p. However, if p is particularly high, then with fewer queries we will have a larger margin-of-error, but will still be statistically confident that it is above the required threshold. Because the benchmark threshold is what we really care about (and not closely estimating p), early stopping allows submitters to process fewer queries in such cases.

Suppose we have a benchmark that requires that submissions achieve a given latency bound 90% of the time. We have system A which achieves this latency bound 99% of the time, and system B which achieves it 91% of the time. In order to have a 99% confidence interval with a margin-of-error of 0.50%, we must perform 23,886 inferences.

This makes sense for system B (whose underlying probability, 91%, is very close to the required benchmark percentile of 90%). However, assuming we see close to 99% of the queries passing the latency requirement for system A, we will be 99% sure that the underlying probability of success for a query on A will be within 99% 土 0.50%. This range is well above the requested latency percentile of 90%. Therefore, by performing fewer queries for such a system, we could widen the margin-of-error slightly, while still being statistically certain of being above the latency benchmark.

A.2. Combinatorial analysis

Suppose we have a system that meets its latency requirement for each query with probability p. What are the odds that we see at least h underlatency queries and at most t overlatency queries? We can answer this by using the cumulative distribution function for the binomial distribution.

We can think of processing queries as performing n Bernoulli trials, with probability of success for any given trial (i.e., odds of being underlatency) equal to p. The probability of exactly k successes (underlatency queries) is equal to:

f(k; n, p) = P(k successes) = (n choose k) * p^k * (1-p)^(n-k)

For fixed n and p, f(k; n, p) is called the binomial distribution with parameters n and p.

In order to determine how unusual our distribution of latency successes and failures is given the underlying probability of passing the latency bound (p), we compute the probability that we had at most h successes, keeping the total number of queries, n, fixed. This, by definition, involves computing the cumulative density function for our binomial distribution, F(h; n, p):

F(h; n, p) = ∑ f(k; n, p),

with the summation going from k = h to n.

Note that, holding h and n fixed, this probability decreases as p increases. This is because, as p gets larger, the odds that our n queries produced results at least as poor as h successes and t failures decreases. In other words, it is harder to achieve a larger number of failures when the underlying probability of an individual success is higher.

This cumulative distribution function for the binomial distribution, F(k; n, p), can be written in terms of the regularized incomplete beta function. The (unregularized) incomplete beta function is defined as:

B(x; a, b) = ∫t^(a - 1) * (1-t)^(b-1) dt, where the integral goes from 0 to x.

We can regularize this to attain:

I(x; a, b) = B(x; a, b) / B(1; a, b).

Note that this is "regularized" in the sense that I(0; a, b) = 0, and I(1; a, b) = 1.

We have an alternate expression for F(k; n, p) in terms of this function:

F(k; n, p) = I(1 - p; n - k, k + 1).

Since the regularized incomplete beta function can be estimated via a continued fraction or by evaluating the Gaussian hypergeometric function, this provides a method for efficiently computing the cumulative density function, F(k; n, p).

A.3. Determining the Early Stopping Criterion

We can use the computation from the previous section to derive an early stopping condition for performing queries to determine whether a system meets a latency bound. Suppose a benchmark requires that our system meets the latency bound p percent of the time. Given that we have seen t queries which are overlatency, at least how many underlatency queries must we see to be sure—within a certain confidence threshold—that we achieve the desired latency bound?

Fix the following variables:

p = the percentile for our tail latency (the percentage of the time we would like our system to achieve the set latency bound)
c = confidence (1 - (false positive rate for minimally failing system))
d = tolerance (amount below target success rate for minimally failing system)
t = number of overlatency queries seen thus far.

We need to determine the smallest h (number of underlatency queries) so that the likelihood of seeing at most t overlatency queries less than 1-c. This is given by an expression involving the cumulative distribution function from the previous section:

F(t; h + t, 1 - (p - d)) ⇐ 1-c.

The left hand side is the probability that the minimally failing system (i.e. one with underlatency rate p-d) resulted in t or fewer overlatency queries. Intuitively, we want to know the smallest number of underlatency queries required such that the probability of us seeing this good of a result, assuming a minimally failing system, is very low (at most 1-c). In other words, in order for us to have seen such a good result, we should be quite sure that we meet the latency bound.

We substitute in our previous expression for F in terms of the regularized incomplete beta function to obtain:

I(p - d; h, t + 1) ⇐ 1-c.

In practice we solve this (i.e. find the smallest h satisfying the above expression) via binary search, keeping a cache of previously-computed solutions for other values of t.

A.4. Early Stopping Criterion

Putting this together, we have the following algorithm for determining early stopping criteria for the server scenario:

When the minimum run duration is met, find the total number of queries processed, q, and the total number of overlatency queries, t.
Using the equations above, compute a minimum total query count, n, given t.
If q is greater than or equal to n, the run is successful.
Otherwise, run for an additional n - q queries and proceed from step 2.

How many times we must iterate through steps 2-4 (and thus how many queries we must process) before ending at a step 3 depends on how close the system’s percentile latency is to the target latency. Systems with lower percentile latency will need to process fewer queries, and those with higher percentile latency will have to process more. In cases where the system percentile latency is worse than the target, the run will never terminate successfully.

The corresponding early stopping algorithm for single-stream and multi-stream scenarios is:

When the minimum run duration is met, find the total number of queries processed, q.
Using the equations above, compute a maximum overlatency count, t, given q.
If t is zero, continue processing queries until t is at least one.
Discard the t - 1 highest latency queries.
Report the maximum latency of the remaining queries.

For our implementation, we use:

d = 0
c = .99.

Appendix B: Datacenter Bandwidth Requirements

Datacenter systems must satisfy both the ingress and egress bandwidth requirements for each benchmark.

B.1. Ingress Bandwidth

Datacenter systems must provide at least the following bandwidths from the network or I/O device to the location where the trace is stored (e.g. DRAM). The minimum bandwidth is a function of the throughput achieved by the SUT and the input data types. The formulas below assume that the inputs are not pre-processed in any way (e.g. padded). If the inputs are pre-processed, and pre-processing affects the input size, submitters must adjust the formulas below accordingly.

Area	Model	Dataset	Symbolic input size formula	Numeric input size formula	Minimum network bandwidth (bytes/sec)
Vision	Resnet50-v1.5	ImageNet (224x224)	CHWdtype_size*	3224224dtype_size*	throughput150528dtype_size
Vision	Retinanet	OpenImages (800x800)	CHWdtype_size*	3800800dtype_size*	throughput1920000dtype_size
Vision	3D UNET	KiTS 2019	avg(CDHW)dtype_size^[6]	32944795dtype_size*	throughput32944795dtype_size
Language	BERT	SQuAD v1.1 (max_seq_len=384)	num_inputsmax_seq_lendtype_size	3384dtype_size	throughput1152dtype_size
Language	Llama3.1-8B	CNN Dailymail (v3.0.0, max_seq_len=2048)	num_inputsmax_seq_lendtype_size	2048dtype_size*	throughput2048dtype_size
Language	Llama2-70b	OpenOrca (max_seq_len=1024)	num_inputsmax_seq_lendtype_size	1024dtype_size*	throughput1024dtype_size
Language	Llama3.1-405B	Subset of LongBench, LongDataCollections, Ruler, GovReport	num_inputsmax_seq_lendtype_size	20000dtype_size*	throughput20000dtype_size
Language	Mixtral-8x7B	OpenOrca (5k samples, max_seq_len=2048), GSM8K (5k samples of the train split, max_seq_len=2048), MBXP (5k samples, max_seq_len=2048)	num_inputsmax_seq_lendtype_size	2048dtype_size*	throughput2048dtype_size
Language	DeepSeek-r1	mlperf_deepseek_r1	num_inputsmax_seq_lendtype_size	3136dtype_size*	throughput3136dtype_size
Commerce	DLRMv2	1TB Click Logs	avg(num_pairs_per_sample)(num_numerical_inputsdtype_size₁ +num_categorical_inputsdtype_size₂))*^[7]	270(13dtype_size₁+26dtype_size₂)*	throughput270(13dtype_size₁+26dtype_size₂)
Generative	SDXL	Subset of coco-2014 val captions (max_prompt_len=77)	num_inputsmax_prompt_lendtype_size	77dtype_size*	throughput77dtype_size
Graph	RGAT	IGBH	negligible	negligible	> 0

B.2. Egress Bandwidth

Datacenter systems must provide at least the following bandwidths from the output location (e.g. DRAM) to the network or I/O device. The minimum bandwidth is a function of the throughput achieved by the SUT and the output data types. For all models except 3D Unet and SDXL, the output sizes are negligible. Therefore, for those models, the egress bandwidth must simply be greater than 0.

Area	Model	Dataset	Symbolic input size formula	Numeric input size formula	Minimum network bandwidth (bytes/sec)
Vision	Resnet50-v1.5	ImageNet (224x224)	negligible	negligible	> 0
Vision	Retinanet	OpenImages (800x800)	negligible	negligible	> 0
Vision	3D UNET	KiTS 2019	avg(CDHW)dtype_size^[6]	32944795dtype_size*	throughput32944795dtype_size
Language	BERT	SQuAD v1.1 (max_seq_len=384)	negligible	negligible	> 0
Language	Llama3.1-8B	CNN Dailymail (v3.0.0, max_seq_len=2048)	negligible	negligible	> 0
Language	Llama2-70b	OpenOrca (max_seq_len=1024)	max_output_lendtype_size*	1024dtype_size*	throughput1024dtype_size
Language	Llama3.1-405B	Subset of LongBench, LongDataCollections, Ruler, GovReport	max_output_lendtype_size*	20000dtype_size*	throughput20000dtype_size
Language	Mixtral-8x7B	OpenOrca (5k samples, max_seq_len=2048), GSM8K (5k samples of the train split, max_seq_len=2048), MBXP (5k samples, max_seq_len=2048)	max_output_lendtype_size*	2048dtype_size*	throughput2048dtype_size
Language	DeepSeek-r1	mlperf_deepseek_r1	num_inputsmax_seq_lendtype_size	32768dtype_size*	throughput32768dtype_size
Commerce	DLRMv2	Synthetic Multihot Criteo Dataset	negligible	negligible	> 0
Generative	SDXL	Subset of coco-2014 val captions (max_prompt_len=77)	3,145,728dtype_size*	throughput3,145,728dtype_size	> 0
Graph	RGAT	IGBH	negligible	negligible	> 0

1. For Llama2-70b, 2 latency metrics are collected - time to first token (TTFT) which measures the latency of the first token, and time per output token (TPOT) which measures the average interval between all the tokens generated.

2. For Llama2-70b, 2 latency metrics are collected - time to first token (TTFT) which measures the latency of the first token, and time per output token (TPOT) which measures the average interval between all the tokens generated.

3. For Llama3.1-405B, 2 latency metrics are collected - time to first token (TTFT) which measures the latency of the first token, and time per output token (TPOT) which measures the average interval between all the tokens generated.

4. For Mixtral-8x7B, 2 latency metrics are collected - time to first token (TTFT) which measures the latency of the first token, and time per output token (TPOT) which measures the average interval between all the tokens generated.

5. For DeepSeek-r1, 2 latency metrics are collected - time to first token (TTFT) which measures the latency of the first token, and time per output token (TPOT) which measures the average interval between all the tokens generated.

6. The average image size above is the average image size of the inference cases specified in inference_cases.json.

7. Each DLRMv2 sample consists of up to 700 user-item pairs draw from the distribution specified in dist_quantile.txt.

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Uh oh!