AI accelerator chip on a circuit board between server racks, circled by a reinforcement learning loop, for MLPerf Training v6.1 results.
Article Icon
Deanna Oothoudt
@
TechArena
Nov 16, 2026

In Record Round, MLPerf Training 6.1 Reaches Into Post-Training

MLCommons released MLPerf Training v6.1 results Monday, the benchmark’s largest round yet and its first to reach beyond pretraining and fine-tuning into post-training. The suite gained an agentic reinforcement learning test and a sequential recommender, and NVIDIA’s Rubin GPU and Google’s Ironwood TPU posted their first results as available systems.

The round drew 250 performance results and one measured-energy result from 104 systems and 24 organizations, up from 229 results in v6.0. Familiar workloads kept getting faster. The best training time for FLUX.1, a text-to-image generation model, fell 2.49x from six months ago, from 17.12 to 6.88 minutes, with the accelerator count unchanged at 512. Every result landed in the closed division’s available category, meaning the systems can be rented or purchased on cloud or on premises today.

New Workloads, New Silicon, Bigger Systems

The v6.1 suite holds seven benchmarks: five carried over from earlier rounds (including FLUX.1) and two new ones. Three (Llama 3.1 8B, GPT-OSS 20B and DeepSeek-V3) measure pretraining, where a model learns to predict the next word. Llama 2 70B LoRA measures fine-tuning, adapting a pretrained model to summarize reports.

The two additions take the suite into new kinds of work. Qwen3.5-397B GRPO, the first post-training benchmark in MLPerf Training, uses reinforcement learning with verifiable rewards to improve a coding agent. DLRMv4-HSTU replaces the long-running DLRMv2 recommender with a model that predicts user engagement from the order and timing of past interactions. The new tests caught on fast, with 29 Qwen3.5 GRPO submissions and 20 for DLRMv4-HSTU.

Twelve accelerator products from NVIDIA, AMD and Google appear. NVIDIA’s Rubin GPU and Google’s TPU7x, known as Ironwood, debuted as available systems, along with a new host processor, Olympus.

The systems running the suite keep getting bigger. Multi-node systems produced 72% of results, up from 58% in v6.0 and 34% in v4.0. Pavan Yalamanchili, MLPerf Training working group co-chair, said models are growing to hundreds of gigabytes or a trillion parameters, and submitters use the benchmarks as proxies for those runs. Three of the largest DeepSeek-V3 submissions used 10,240 GB300 GPUs from NVIDIA, 10,240 GB200 GPUs from Azure and 8,192 Ironwood chips from Google.

Software Improvements Continue to Pay Off

Round-over-round speedups blend new hardware with new software, and the fastest system on a workload can change size between rounds. To control for system size, MLCommons also compared best times at fixed accelerator counts. At eight accelerators, best times improved 2.03x on Llama 2 70B LoRA, 1.97x on GPT-OSS-20B and 1.75x on Llama 3.1 8B.

Holding the accelerator family constant also helps isolate the effects of software and tuning. On that basis, GB300 systems improved 1.85x on FLUX.1 at 512 GPUs, and AMD MI300X, MI350X and MI355X systems gained 1.18x to 1.25x. Yalamanchili said the software gains come from three areas: more efficient kernels, compact activation storage that avoids recomputation, and better overlap of communication with computation.

AI-assisted optimization is new this round, with some submitters using AI tools to help develop and test kernels, the low-level code that runs on the accelerators. KRAI, a UK firm that optimizes and benchmarks AI systems, cut an eight-GPU MI300X Llama 2 70B LoRA run about 20% from its v6.0 result, to 22.83 minutes. Firmus, an Australian builder and operator of AI data centers, cut an eight-GPU GB300 GPT-OSS-20B run about 24% against a baseline it also submitted this round, to 52.90 minutes.

The TechArena Take

Adding post-training is the most consequential change in v6.1. Every other benchmark in the current suite, from language models to FLUX.1’s image generation, learns from a fixed dataset. In reinforcement learning post-training, the model produces its own training data by attempting tasks and being graded on the results. A growing share of the work that turns a base model into a useful product now happens after pretraining, as models are refined this way on tasks whose answers can be checked. Qwen3.5-397B GRPO gives this important part of the training ecosystem a standardized measurement for the first time.

It also changes what training performance means. A reinforcement learning (RL) run combines three jobs: updating the model, generating its attempts and running tests in sandboxes on conventional processors. Each has to keep pace with the others. That makes the benchmark a test of the whole system, not just the accelerator. Between this update and the agentic tests added to MLPerf Inference in September, MLCommons is tracking AI’s shift from answering prompts to completing multi-step tasks.

‍

Subscribe to Our Newsletter

Read the latest in the world of AI, data center, and edge innovation.