AMD’s broadest MLPerf entry yet claims leads over selected NVIDIA results

AMD has filed its broadest MLPerf Inference submission ever, and it has arrived with fighting words attached. The company says its Instinct MI355X GPUs beat selected NVIDIA B200 and B300 submissions on the GPT-OSS-120B language model at eight GPUs, led the 72-GPU GB200 result too, and the brand new MI350P PCIe card won its own benchmark debut. Underneath the trophy talk sits a quieter claim with more bite: AMD says software alone wrung 28% more Offline and 38% more Server throughput from an unchanged eight-GPU MI355X system in a single benchmark cycle.

AMD Watch analysis

What happened

MLPerf Inference is MLCommons' shared yardstick for inference hardware, and AMD's account of its 6.1 entry describes a platform stepping up in scale and breadth. Last round covered three model families; this one spans six, across language, reasoning, text-to-video and recommendation workloads, on Instinct MI355X and MI350X GPUs plus the new MI350P card.

Key findings

• Same chips, more throughput
  AMD credits ROCm software for a 28% Offline and 38% Server uplift on GPT-OSS-120B, and 70% better Wan-2.2 SingleStream, on identical MI355X hardware.
• Seventy-two GPUs beat ninety-four
  This round's 72-GPU MI355X result out-delivered last round's 94-GPU submission on GPT-OSS-120B, a tidy software dividend.
• Scaling that barely leaks
  Growing from eight to 72 GPUs, a ninefold jump, kept 95% scale efficiency in both Offline and Server testing, per AMD.
• Crusoe's record run
  At 512 GPUs, Crusoe's GPT-OSS-120B submission hit 5.75 million Offline tokens per second, which AMD calls the highest aggregate token throughput ever submitted to MLPerf.
• A new card for ordinary racks
  The dual-slot MI350P, with 144GB of HBM3E, led selected NVIDIA RTX PRO 6000 Server Edition and H200 NVL submissions across five Closed workloads in its first round.
• Partners kept pace
  Submissions from Dell, Oracle, HPE, Supermicro, MangoBoost, Crusoe and MiTAC averaged within 4% of AMD's reference results, including a 32-GPU endpoint split between Korea and the United States.

Why it matters

Inference hardware is where AI budgets actually land, and a credible second supplier changes prices for everyone. The comparisons are AMD-selected, so read them as one side of the scorecard. NVIDIA's own headline this round was the Vera Rubin NVL72 debut, with its claimed gains arriving on Qwen3-VL and DeepSeek-R1 rather than GPT-OSS-120B. Both vendors are telling true stories about the workloads they picked.

The more durable signal is the software trajectory. If the same silicon genuinely gains 38% in a cycle, today's benchmark tables have short shelf lives, and fleets keep improving after the invoice is paid. The MI350P matters too: a PCIe path brings this generation of inference to data centres that were never built for exotic accelerator modules.

Our read

Treat the wins as real but curated. "Selected NVIDIA submissions" is doing quiet work in AMD's sentences, and quite right too; everyone curates. What would move this from interesting to important is the combination: partner results tracking AMD within 4%, a record aggregate throughput run at 512 GPUs, and a ROCm stack maturing fast enough to improve hardware people already own.

If you operate inference capacity, the practical step is dull but correct: wait for the full MLCommons charts, then price the MI350P against your existing racks before anyone's marketing deck prices it for you.

What to watch

• The published round.
  MLCommons' full 6.1 results, and how AMD and NVIDIA compare on shared workloads.
• Independent repetition.
  Whether anyone outside Crusoe reproduces the 512-GPU throughput record.
• MI350P availability.
  Pricing and early PCIe deployments are the card's real test.
• The next cycle.
  Whether same-silicon software gains keep arriving at this pace.

Discussion spark: AMD's most striking claim is 38% more throughput from unchanged silicon in a single cycle. If software keeps improving at that pace, are benchmark tables snapshots of chips or of software, and would you still buy on today's numbers?

Independent WittyWires tracker for public updates about AMD. Not affiliated with or endorsed by AMD; this is not an official account.