No neutral source ranked AI hardware by "top picks" in 2022, but the industry benchmark MLPerf Training v2.0, released by MLCommons on June 29, 2022, measured how long whole computer systems took to train standard models. It reported over 250 performance results from 21 submitters and up to 1.8 times the performance of earlier results. Those figures describe complete systems, often built from many accelerators, so they say little about which single device a student or small lab should buy. The suite has moved on since: MLPerf Training v6.0, published June 16, 2026, drew 95 systems from 24 organizations.
Key takeaways
- MLPerf Training v2.0 (June 29, 2022) included over 250 results from 21 submitters and added an object detection test using the RetinaNet model on the Open Images dataset.
- The latest round as of October 2026, MLPerf Training v6.0 (June 16, 2026), had 95 systems from 24 organizations and added two Mixture-of-Experts benchmarks.
- MLPerf measures wall-clock time to reach a fixed quality target, and results come from full systems, not single chips.
- Closed-division results use the same model and are the most directly comparable; Open-division results may use different models.
- Learners can reach large-scale computing through public programs, such as the US NAIRR Pilot and Canada's national research computing systems, without buying hardware.
What MLPerf Training measures
MLCommons describes MLPerf Training as an open-source, peer-reviewed suite of full-system tests that measure software, hardware, and models across machine learning applications. Each benchmark pairs a dataset with a quality target and a reference model, and the score is the wall-clock time needed to reach that target. A system that trains faster is not necessarily more accurate, because every system is asked to hit the same bar. Submitters in the June 2022 round included Azure, Baidu, Dell, Google, Intel-HabanaLabs, NVIDIA, and Supermicro, and six organizations (ASUSTeK, CASIA, H3C, HazyResearch, Krai, and MosaicML) submitted for the first time.
How to read a result without being misled
The current MLPerf Training page lists several rules that limit what a headline number can mean. Whether each applied in exactly this form in 2022 was not checked, so the table describes the benchmark as documented in 2026.
| Term | Meaning | Effect on comparison |
|---|---|---|
| Closed division | Same model as the reference implementation | Fairest comparison of hardware and frameworks |
| Open division | Different models or retraining allowed | Not directly comparable with Closed results |
| Available system | Built only from parts that can be bought or rented in the cloud | Closest to what an organization could obtain |
| Preview and research systems | Not yet available, or experimental or internal | May never reach buyers in that form |
| Run-to-run variance | About 2.5 percent for imaging benchmarks and 5 percent for others | Small gaps between results may be noise |
Results can also be modified or invalidated after publication, and the rules document is the official source of truth. A reader citing a 2022 number should check that the entry still stands on the MLCommons announcement and the results tables it links.
What the object detection addition shows
The 2022 round added a benchmark that trains the RetinaNet reference model on the Open Images dataset, which MLCommons describes as larger and more diverse than the data used before. It targets uses such as vehicle collision avoidance and robotics. The change illustrates a point about benchmarks generally: the suite is revised as workloads change, so a result from one round is not a lasting ranking. The current suite lists large language model pretraining and image generation tasks, and the June 2026 round added models with hundreds of billions of parameters.
What changed since 2022: MLPerf Training v6.0
MLCommons published MLPerf Training v6.0 on June 16, 2026, the most recent round at the time of writing. It reports 95 unique systems from 24 submitting organizations, using 13 different types of hardware accelerators and 19 types of host processors. Sixty percent of the submissions were multi-node systems. Four organizations (Inventec, Netweb Technologies India, TTA, and Vultr) submitted for the first time.
The two new tests are both Mixture-of-Experts models. DeepSeek V3 has 671 billion total parameters, with 37 billion active per token, and is the largest benchmark in the suite. GPT-OSS 20B has 21 billion total parameters, with 3.6 billion active per token, and is small enough to run on one 8-GPU node. MLCommons also notes more than double the cloud systems compared with the v5.1 round six months earlier. For comparison, the 2022 round counted 21 submitters and added a single object detection test.
Several vendors have published their own readings of v6.0 results. Those are vendor statements, and this article does not repeat their speedup claims. Readers comparing a 2022 and a 2026 figure should remember that the models differ, so the times cannot be set side by side.
Access to computing without buying it
Most students never need the systems in the benchmark tables. In the United States, the National Science Foundation leads the NAIRR Pilot, which connects researchers and educators to computing, data, and training resources. The allocations page states that applicants must be US-based researchers and educators, that graduate students can apply with a faculty advisor's support letter, that an institutional email address is required, and that awards run for 12 months. The pilot's site reports 105 NAIRR Classroom awards, and an NSF award for a NAIRR Operations Center was announced on September 1, 2026.
In Canada, a University of Windsor page describes national research computing systems run through the Digital Research Alliance of Canada as free of charge for eligible faculty and the students and postdocs they sponsor, with larger needs handled through an annual allocation competition. Eligibility and capacity change, and each program's own pages govern.
Questions to ask before choosing hardware
- Is the target workload training or running a finished model? The two have very different memory and speed needs.
- Is the benchmark entry an Available system, and was it measured on the same model the project will use?
- Does a free or institutional option cover the course or project?
- Is the figure from a system with many accelerators that cannot be matched on a single machine?
This article gives general information and does not recommend a product. Prices and availability change, and purchase decisions belong with the buyer and the vendor's current documentation.
FAQ
Was there an official list of the best AI hardware in 2022?
No neutral body published one. MLPerf Training v2.0 gave measured training times for submitted systems, and each result depends on the model, the system size, and the software.
Do MLPerf results tell a student which GPU to buy?
Not directly. Entries are complete systems, often far larger than a workstation, so they show the state of the technology rather than a purchase guide.
Why do Open and Closed results differ?
The Closed division requires the reference model so that hardware can be compared fairly, while the Open division allows changes to the model and so cannot be compared with it on equal terms.
Can undergraduates use the NAIRR Pilot?
The allocations page names graduate students with an advisor's letter and educators. It does not address undergraduates, so an instructor or the program office should be asked.
Related reading
- AI and machine learning compared
- How AI training is implemented
- Deep learning applications and advances
- Traditional and modern AI methods
More in AI in education.
Last verified: October 2026
Sources
- MLPerf Results Highlight More Capable ML Training, MLCommons, June 29, 2022
- MLPerf Training benchmark and rules page, MLCommons, 2026
- MLPerf Training v6.0 Results: New MoE Benchmarks and Record System Diversity, MLCommons, June 16, 2026
- NAIRR Pilot home page and allocations page, National Science Foundation NAIRR Pilot, 2026
- National research computing at no cost for eligible faculty and sponsored users, University of Windsor Research, 2026