Google Cloud reported 10 submissions for MLPerf Training v6.1 on November 16, 2026, spanning Ironwood TPUs and NVIDIA Grace Blackwell GPUs. The headline times belong to specific workloads and configurations: Ironwood’s 1.70-minute result was for DeepSeek-V3 (671B), while its Qwen3.5-397B result measured a separate post-training task. The reported scaling efficiency also changed between chip-count steps. (Google Cloud; MLPerf supplemental discussion)

Read each benchmark result alongside its model, workload, accelerator count, and system configuration.

What the benchmark results measured

Google reported these target-quality times for three highlighted MLPerf Training v6.1 runs:

WorkloadSystem and configurationReported time to target quality
DeepSeek-V3 (671B) pre-training8,192 Ironwood chips in one inter-chip interconnect (ICI) domain1.70 minutes
Qwen3.5-397B post-training reinforcement learning640 Ironwood chips18.02 minutes
DeepSeek-V3 pre-training256 NVIDIA GB300 GPUs on Google Cloud A4X Max instances31.96 minutes

Google’s supplemental submission gives the Ironwood DeepSeek-V3 model size and ICI configuration; Google Cloud reported the highlighted run times. The numbers describe different systems running particular workloads, rather than interchangeable tests. (MLPerf supplemental discussion; Google Cloud)

The Qwen3.5-397B benchmark measures post-training with Group Relative Policy Optimization (GRPO) on software-engineering tasks, including tool-calling. MLCommons identifies it as a new workload in v6.1. Google reported an accuracy of 0.765 against a 0.69 goal, with training chips waiting less than 3% of the time. Its setup used a 128-chip trainer slice and 128 four-chip sampler engines. (MLCommons; Google Cloud; MLPerf supplemental discussion)

Ironwood’s DeepSeek-V3 scaling results

For DeepSeek-V3, Google reported target-quality times of 5.03 minutes on 2,048 Ironwood chips, 2.62 minutes on 4,096 chips, and 1.70 minutes on 8,192 chips. Its reported scaling efficiency was about 96% from 2,048 to 4,096 chips, then 77% from 4,096 to 8,192. The 96% figure applies to the first step, not the full scale-up to 8,192 chips. (MLPerf supplemental discussion)

Google also reported effective model FLOPs utilization (EMFU) of 56%, 54%, and 42% at 2,048, 4,096, and 8,192 chips, respectively, measured against Ironwood’s peak BF16 performance. These percentages describe the submitted DeepSeek-V3 runs at those chip counts; they are not the scaling-efficiency figures. (MLPerf supplemental discussion)

Why the TPU and GPU times are not a universal ranking

The highlighted DeepSeek-V3 runs used 8,192 Ironwood chips and 256 GB300 GPUs, respectively. MLCommons describes MLPerf as a full-system benchmark spanning models, software, and hardware, so the reported times apply to those workload-and-system combinations. They do not provide a chip-for-chip comparison or establish which platform would be faster across other workloads.

MLPerf Training v6.1 also added the DLRMv4-HSTU recommendation workload alongside Qwen3.5-397B post-training. Across the round, MLCommons reported 153 systems, 250 test results, and 24 submitting organizations. (MLCommons)