MLPerf Inference 6.1 AMD MI355X Vera Rubin Intel Arc Pro Results Explained [2026]
By Indie Kings | September 23, 2026
Updated September 23, 2026: MLPerf Inference v6.1 brings 30 submitters, 120 systems, new agentic and end-to-end benchmarks, and a record 512-accelerator submission, per MLCommons on September 17. Wccftech's relay reports AMD MI355X scale leadership on DeepSeek R1 and GPT-OSS 120B, an early Vera Rubin preview, and Intel Arc Pro B70 plus Xeon submissions.
Image: MLPerf Inference v6.1 analysis graphic. Credit: MLCommons.
What MLPerf 6.1 measures
Per MLCommons, MLPerf Inference v6.1 covers Datacenter and Edge categories with 10 Datacenter benchmarks and 6 Edge benchmarks. New tests include End-to-End RAG and Agentic Edge Inference, while the VLM and GPT-OSS-120B scenarios receive updates, per the official analysis.
Per MLCommons, Closed submissions use a model mathematically equivalent to the reference implementation, holding the model fixed for direct comparison. Open submissions surface optimization techniques and implementation choices, per the benchmark documentation.
Per MLCommons, scenarios can include Offline, Server, Interactive, SingleStream, or MultiStream. The field includes 30 submitters and 120 systems across both divisions, with joint submissions from organizations such as Dell with AMD, Red Hat with Intel, and Red Hat with Supermicro, per the official analysis.
| Benchmark area | v6.1 coverage per MLCommons |
|---|---|
| Datacenter | 10 benchmarks including GPT-OSS-120B, Llama2-70B, Llama3.1-8B, DeepSeek R1, VLM, Whisper, DLRM, and RAG |
| Edge | 6 benchmarks including Agentic Edge, YOLO, Whisper, ResNet, and Llama3.1-8B Edge |
| New emphasis | End-to-End RAG plus Agentic Edge Inference |
| Comparison | Closed and Open divisions with different comparison rules |
Selected results
Per Wccftech's September 16 relay, AMD submitted a 512-GPU MI355X configuration that led the DeepSeek R1 Offline and Server scenarios. NVIDIA submitted GB300, GB200, and an early VR200 Vera Rubin preview, while Intel submitted Arc Pro B70 and Xeon systems, per the report.
Per the same relay, AMD also led the 512-GPU GPT-OSS-120B results, while NVIDIA led selected 72-GPU and 8-GPU configurations on other models. The table below preserves Wccftech's reported figures and does not turn them into a universal ranking.
| System per Wccftech relay | Workload and scenario | Reported result |
|---|---|---|
| AMD MI355X, 512 GPUs | DeepSeek R1 Offline | 2,901,950 tokens per second |
| AMD MI355X, 512 GPUs | DeepSeek R1 Server | 2,405,310 tokens per second |
| AMD MI355X, 512 GPUs | GPT-OSS-120B Offline | 5,749,440 tokens per second |
| AMD MI355X, 512 GPUs | GPT-OSS-120B Server | 5,392,930 tokens per second |
| NVIDIA GB300, 288 GPUs | DeepSeek R1 Offline | 2,705,130 tokens per second |
| Intel Arc Pro B70, 4 GPUs | Llama3.1-8B Offline | 4,977 tokens per second |
Scale and software
Per MLCommons, the largest v6.1 submission used 512 accelerators and set a new aggregate throughput record for GPT-OSS-120B at almost 5.8 million Offline tokens per second. The same analysis reports 16 multi-node submissions, the highest count in the series, per the official article.
Per Wccftech's relay, NVIDIA reported up to 3.7 times better Vera Rubin throughput than GB300 on selected Qwen3-VL scenarios and 99 percent scaling efficiency for a 288-GPU GB300 submission. Those are submitter and relay claims, not independently reproduced results, per the report.
Per the same relay, AMD reported gains from continued ROCm optimization on the same MI355X hardware. Intel reported a 36 percent Server gain and 27 percent Offline gain for GPT-OSS-120B on the same four-B70 system used in v6.0, per the Intel statement quoted by Wccftech.
- Scale: 512 accelerators becomes the headline submission size.
- Software: Same silicon can move between rounds as stacks mature.
- New models: GPT-OSS-120B becomes the most popular Datacenter benchmark.
- New workloads: Agentic and end-to-end tests join the suite.
How to read the numbers
MLPerf is not a gaming benchmark and does not turn one submission into a universal hardware winner. Results depend on model, scenario, precision, software stack, system design, and whether the submission is Closed or Open, per MLCommons methodology.
Per MLCommons, benchmark bars are comparable within a model and scenario, not across unrelated benchmarks. A DeepSeek R1 result cannot be read as a Llama result, and an Offline batch result cannot be treated as interactive latency, per the official explanation.
Use the round for measured direction, not price guidance. This post contains no buying advice, price target, or investment conclusion.
| Question | Correct read |
|---|---|
| Who won v6.1? | No single winner; results depend on workload and scale |
| Is Vera Rubin proven? | No, Wccftech calls it an early preview |
| Does MLPerf measure gaming FPS? | No, it measures inference workloads |
| Can Open and Closed be merged? | No, per MLCommons rules |
FAQ
What is new in MLPerf Inference v6.1?
Per MLCommons, 30 submitters, 120 systems, Agentic Edge Inference, and End-to-End RAG headline the round.
Did AMD lead every benchmark?
No, per Wccftech. AMD led selected 512-GPU DeepSeek R1 and GPT-OSS results, while NVIDIA led other configurations.
Is Vera Rubin fully benchmarked?
No, per Wccftech. Its submission is an early preview with selected results.
What did Intel submit?
Per Wccftech, Intel submitted Arc Pro B70 and Xeon 6 systems across several inference workloads.
What is the difference between Closed and Open?
Per MLCommons, Closed holds the model fixed for direct comparison, while Open exposes optimization techniques.
Is this a gaming benchmark?
No. It measures inference systems and scenarios, not game frame rate.
Bottom Line
MLPerf 6.1 shows a widening field, record scale, and fast-moving software stacks. Read each result inside its model, scenario, and division. The useful signal is trajectory, not a single champion.
Related: NVIDIA MLPerf Training scaling report | Intel Gaudi 3 MLPerf 6.0 analysis