W.145Nvidia’s Vera Rubin trained 1.4 to 2 times as fast as Blackwell Ultra in MLPerf
MLCommons published MLPerf Training v6.1 on November 16, the first round with Nvidia’s Vera Rubin. Nvidia’s own table puts Rubin’s gain per GPU over its Blackwell Ultra system at 2 times on one of seven benchmarks, DeepSeek-V3 pretraining, where 1,024 GPUs took 4.14 minutes against 8.03, and at 1.4 to 1.6 times on the other six. Blackwell’s first round against the generation before it showed 2.2 to 2.6 times. One analyst estimate puts a Rubin rack at about 1.95 times the price of a Blackwell Ultra rack. In a separate inference benchmark, Nvidia’s claim of more than 10 times the throughput per kilowatt rests on a response-speed floor; at equal load the ratio from its own figures is 1.8.
What was published
The round is the first with Rubin hardware. MLCommons said 24 organizations submitted 250 results on 153 systems, the largest with 10,240 accelerators. It added two benchmarks: DLRMv4, a recommendation model, and Qwen3.5-397B GRPO, a reinforcement-learning task run after pretraining. The accelerators came from Nvidia, AMD and Google, TechArena reported, and every result was in the closed division and the available category.
MLPerf Training measures time to a quality target. A result is the minutes a system takes to train a model to a set accuracy from a defined starting point. Google noted that its runs start from the reference checkpoint. A time of a few minutes is not the time to train the model from scratch.
Nvidia’s headline is “up to 2x.” Its blog says Vera Rubin NVL72 delivered up to twice the performance of a GB300 NVL72, its Blackwell Ultra rack, and gives a table of seven benchmarks at equal GPU counts. Both are 72-GPU rack designs, so at equal GPU counts a gain per GPU is roughly a gain per rack; Nvidia’s largest runs used 64 GPUs a rack.
Rubin’s gain reaches 2 times on one benchmark and is 1.4 to 1.6 times on the other six
Vera Rubin NVL72’s training performance per GPU as a multiple of GB300 NVL72 at the same GPU count, as Nvidia lists it for MLPerf Training v6.1, closed division. Pink: DeepSeek-V3. Grey: the other six benchmarks. The line marks 1.95 times, the ratio of one analyst estimate of the two racks’ prices ($7.8 million and $4 million, Morgan Stanley as reported by MarketScale in July 2026); Nvidia has not published prices. Where Nvidia gives times, the wide chart shows them.
Nvidia’s table, with our division
| Benchmark | GPUs | Vera Rubin, min | GB300, min | Our division | Nvidia lists |
|---|---|---|---|---|---|
| DeepSeek-V3 671B | 1,024 | 4.14 | 8.03 | 1.94× | 2.0× |
| gpt-oss 20B | 64 | 11.213 | 18.24 | 1.63× | 1.6× |
| Llama 3.1 8B | 72 | 7.162 | 11.541 | 1.61× | 1.6× |
| Llama 2 70B LoRA | 72 | not given | not given | 1.5× | |
| FLUX.1 | 72 | not given | not given | 1.5× | |
| Qwen3.5 397B GRPO | 64 | not given | not given | 1.4× | |
| DLRMv4 | 128 | 2.829 | 3.871 | 1.37× | 1.4× |
The size of the gain
The 2 times is one benchmark at one scale. On DeepSeek-V3 pretraining with 1,024 GPUs, Rubin took 4.14 minutes and GB300 took 8.03, which is 1.94 times by our division. Nvidia’s text calls it close to 2 times per GPU.
The other six benchmarks show 1.4 to 1.6 times. Nvidia lists 1.6 times for gpt-oss 20B and Llama 3.1 8B, 1.5 times for Llama 2 70B LoRA and FLUX.1, and 1.4 times for Qwen3.5 GRPO and DLRMv4. The plain average of the seven listed figures is 1.57 times. Nvidia lists Qwen3.5 GRPO at 64 GPUs. At 1,024 GPUs its text gives 10.44 minutes against 13.78, which is 1.32 times by our division, a second result at the scale of the DeepSeek-V3 run.
Against the original Blackwell rack the gain is larger. Nvidia’s blog gives 12.36 minutes for GB200 NVL72 on the same DeepSeek-V3 run at 1,024 GPUs, which makes Rubin 2.99 times as fast. It gives no GB200 comparison for the other benchmarks.
Software may move the numbers. Nvidia’s blog also shows reruns on a newer software build with gains of up to 2.4 times over the GB300 submissions, and says MLCommons has not verified them. The 2.4 times sets Rubin’s rerun against GB300’s submitted time; with both on the newer build, the three reruns give 1.5 to 1.8 times by our division. TechArena noted that gains between rounds mix hardware and software.
Blackwell’s first round showed 2.2 to 2.6 times over Hopper; Rubin’s shows 1.4 to 2 times over Blackwell Ultra
Training gain per GPU of each Nvidia system over an earlier one in MLPerf Training, from Nvidia’s blogs. Grey: GB200 over Hopper in round 5.0, a range across four benchmarks, three of them at 8 GPUs against Hopper results from an earlier round, and GB300 over GB200 in round 5.1, our division of Nvidia’s index values of 4.2 and 3.2 for one benchmark at 512 GPUs. Pink: Vera Rubin over GB300 in round 6.1, the range across seven benchmarks. Strong grey: Vera Rubin over GB200 on the one benchmark where Nvidia gives both times. Outlined extensions show the top of a range. The benchmarks differ between rounds.
Against earlier generations and earlier claims
The step from Blackwell Ultra to Rubin is smaller than the step from Hopper to Blackwell. In round 5.0, in June 2025, Nvidia reported GB200 at 2.2 to 2.64 times Hopper per GPU. Blackwell Ultra was a refresh within the Blackwell generation: Nvidia’s index for round 5.1, on one benchmark at 512 GPUs, put GB300 at 4.2 and GB200 at 3.2 against an H100 at 1, a step of 1.31 times. Rubin against GB200, the fuller generation comparison, is 2.99 times on the one benchmark with both times.
Nvidia quoted 3.5 times for the chip at launch. At CES in January 2026 it said Rubin chips would deliver up to 5 times the floating-point performance of Blackwell for inference and 3.5 times for training, The Register reported. That is a chip specification, not a time-to-train figure. The nearest measured comparison in this round is 2.99 times against GB200 on DeepSeek-V3.
On one price estimate, only the best result keeps pace with the price. Morgan Stanley estimated a Vera Rubin NVL72 rack at about $7.8 million against about $4 million for GB300, MarketScale reported in July, a ratio of 1.95. A 1.94 times gain at that ratio is about level per dollar of rack; a 1.57 times gain is 0.81 times. Nvidia has not published prices, and a rack’s price is not the whole cost of running it.
Power estimates point the same way with a wider range. SemiAnalysis in February put rack TDP at 180 to 220 kW for Vera Rubin NVL72 against 120 to 140 kW for GB200 and GB300, a ratio of 1.29 to 1.83 at the extremes. The round drew one measured-energy result alongside its 250 performance results, TechArena reported.
At 1,024 GPUs Rubin took 4.14 minutes on DeepSeek-V3; the fastest DeepSeek-V3 time, 1.38 minutes, came from 10,240 Blackwell Ultra GPUs
Minutes to reach the MLPerf quality target on the DeepSeek-V3 benchmark in round 6.1; shorter is faster. Top group, at 1,024 GPUs, from Nvidia’s blog: faint grey GB200 NVL72, grey GB300 NVL72, pink Vera Rubin NVL72. Bottom group, the largest systems: black Google’s seventh-generation TPU, which TechArena calls TPU7x, on 8,192 chips, from Google’s blog; grey GB300 NVL72 on 10,240 GPUs, from Nvidia’s. The bottom two differ in chip count as well as chip, and chips are not equivalent units. Hardwareluxx reports a GB300 run on 8,192 GPUs at 1.53 minutes, which Nvidia’s blog does not give and we do not chart.
The other submitters
None of the other vendors’ announcements we read reports a DeepSeek-V3 result at 1,024 accelerators. AMD submitted its Instinct MI355X and MI350X on 5 of the seven benchmarks, not including DeepSeek-V3 or Qwen3.5 GRPO. The largest AMD-based system, from its cloud partner Crusoe, had 128 GPUs, and AMD’s own largest had 64. By AMD’s own account the MI355X trails Nvidia’s B300 by 7.4% on Llama 2 70B LoRA and 4.7% on Llama 3.1 8B, both at 8 GPUs. It did not compare itself with Rubin on those.
Google’s seventh-generation TPU appeared for the first time. Google reported DeepSeek-V3 in 1.70 minutes on 8,192 chips and the Qwen3.5 task in 18.02 minutes on 640 chips. Nvidia’s fastest DeepSeek-V3 time was 1.38 minutes on 10,240 GB300 GPUs. Hardwareluxx noted that not every configuration ran every benchmark, so the largest clusters cannot be compared directly, and found the only clean three-way comparison on Llama 2 70B LoRA at eight accelerators.
Rubin did not post the fastest DeepSeek-V3 time. Nvidia says Rubin was fastest on six of the seven benchmarks. Its largest Rubin system had 1,024 GPUs in 16 racks, and its 10,240-GPU run used Blackwell Ultra. CoreWeave also submitted Rubin results, on 72 and 256 GPUs.
The inference claim
Nvidia also claimed more than 10 times the throughput per kilowatt. The claim is for a different benchmark, MLPerf Endpoints v1.0, which measures a served model’s total tokens per second against the speed each user sees. On the Kimi K3 model, Nvidia reports 41,414.2 tokens a second from a Rubin system with 32 accelerators and 100.89 kW of provisioned power, and 2,210.4 from a GB300 system with 48 accelerators and 60 kW. That is 410 against 37 tokens a second per kilowatt, 11.1 times.
The comparison is set by a speed floor. Nvidia compares the systems at a minimum of 50 tokens a second per user. Rubin clears it while serving 1,024 concurrent queries. The GB300 point Nvidia uses is 48 concurrent queries. At 1,024 concurrent queries GB300 produces 13,905.8 tokens a second, but each user gets 16.8. Per kilowatt at that equal load, Rubin’s figure is 1.77 times GB300’s.
The Rubin result is provisional. Nvidia says it is pending verification by MLCommons; the GB300 result is verified. Nvidia calls the power figure provisioned power in its text and operating power in its footnotes and defines neither; it is not presented as measured energy. The systems have 32 and 48 accelerators, fewer than the 72 in a full rack.
Per kilowatt, Rubin’s Kimi K3 result is 11 times GB300’s at the speed floor and 1.8 times at equal load
MLPerf Endpoints v1.0, Kimi K3: tokens per second per kilowatt of provisioned power, our division of the figures in Nvidia’s blog. Pink: Vera Rubin at 1,024 concurrent queries, a provisional result. Faint grey: GB300 at the same 1,024 concurrent queries, where each user gets 16.8 tokens a second, below the 50 Nvidia uses as a floor. Grey: GB300 at 48 concurrent queries, the result Nvidia compares against. Nvidia does not define provisioned power, and it is not presented as energy measured during the run.
The Endpoints figures
| System | Accelerators | Concurrent queries | Tokens/s | Tokens/s per user | kW provisioned | Tokens/s per kW |
|---|---|---|---|---|---|---|
| Vera Rubin, provisional | 32 | 1,024 | 41,414.2 | 50.4 | 100.89 | 410.5 |
| GB300, verified | 48 | 48 | 2,210.4 | 56.2 | 60 | 36.8 |
| GB300 at equal load | 48 | 1,024 | 13,905.8 | 16.8 | 60 | 231.8 |
What this changes on inthesingularity.ai
- Who owns the most Blackwell per dollar? Lead: the first peer-reviewed training numbers for the next generation. A per-GPU gain of 1.4 to 2 times against a rack price estimated at 1.95 times would leave training per dollar of hardware flat or lower; the pack could carry a Rubin row once a price is reported.
- W.143. The forecast cut reported this week lowers assumed Nvidia selling prices for later generations; a measured gain below the chip-level multiple Nvidia quoted at launch bears on the same question.
- Lead: a series of MLPerf per-GPU gains by generation and round, which now has three points for Nvidia and a first for Google’s TPU7x.
- Lead: the verified MLPerf Endpoints result for Rubin, when MLCommons completes its review, and any result that reports measured energy.
Sources and method
| Item | Figure | Source |
|---|---|---|
| Rubin results | Table of Vera Rubin NVL72 against GB300 NVL72 at equal GPU counts, closed division: DeepSeek-V3 671B (1,024 GPUs) 2.0×, 4.14 against 8.03 minutes; gpt-oss 20B (64 GPUs) 1.6×, 11.213 against 18.24 minutes; Llama 3.1 8B (72 GPUs) 1.6×, 7.162 against 11.541 minutes; Llama 2 70B LoRA (72 GPUs) 1.5×; FLUX.1 (72 GPUs) 1.5×; Qwen3.5 397B GRPO (64 GPUs) 1.4×; DLRMv4 (128 GPUs) 1.4×, 2.829 against 3.871 minutes; Qwen3.5 GRPO at 1,024 GPUs: 10.44 against 13.78 minutes; Vera Rubin fastest on six of the seven benchmarks; GB300 baselines are submissions in this round | Nvidia Technical Blog, Nov 16 |
| Rubin results | DeepSeek-V3 at 1,024 GPUs on GB200 NVL72: 12.36 minutes; GB300 NVL72 on 10,240 GPUs in 160 racks: 1.38 minutes; Vera Rubin’s largest system 1,024 GPUs across 16 racks; above 97% efficiency scaling from 256 to 1,024 GPUs; unverified reruns on newer software show gains of up to 2.4×, with reruns of gpt-oss 20B at 8.08 (Vera Rubin) and 13.151 (GB300) minutes, Llama 3.1 8B at 6.14 (Vera Rubin) and 9.211 (GB300) minutes, DLRMv4 at 1.628 (Vera Rubin) and 2.94 (GB300) minutes | Nvidia Technical Blog, Nov 16 |
| The round | 24 submitting organizations, 153 systems, 250 results; largest system 10,240 accelerators; two new benchmarks, DLRMv4-HSTU and Qwen3.5-397B GRPO; first Rubin and first Google TPUv7 (Ironwood) results | MLCommons, Nov 16 |
| The round | Seven benchmarks; all results in the closed division and available category; accelerators from Nvidia, AMD and Google, with Google’s called TPU7x; one measured-energy result alongside the 250 performance results; round-over-round gains mix hardware and software | TechArena, Nov 16 |
| Other submitters | Instinct MI355X and MI350X on 5 of 7 benchmarks, none on DeepSeek-V3 or Qwen3.5 GRPO; largest AMD-based system 128 GPUs, from Crusoe, and AMD’s own largest 64; AMD says MI355X trails Nvidia’s B300 by 7.4% on Llama 2 70B LoRA and 4.7% on Llama 3.1 8B at 8 GPUs | AMD ROCm blog, Nov 16 |
| Other submitters | Seventh-generation Ironwood TPUs: DeepSeek-V3 in 1.70 minutes on 8,192 chips; Qwen3.5-397B in 18.02 minutes on 640 chips | Google blog, Nov 16 |
| Other submitters | CoreWeave submitted Vera Rubin results on 72 and 256 GPUs: Llama 3.1 8B in 7.16 minutes on 72, DeepSeek-V3 in 17.65 minutes on 256 | CoreWeave release, Nov 16 |
| Other submitters | Not every configuration ran every benchmark, so the largest clusters cannot be compared directly; the only clean three-way comparison is Llama 2 70B LoRA at eight accelerators; a GB300 run on 8,192 GPUs at 1.53 minutes on DeepSeek-V3 | Hardwareluxx, Nov 16 |
| Earlier rounds | Round 5.0: GB200 per GPU over Hopper 2.2× on Llama 3.1 405B, up to 2.64× on Stable Diffusion v2 | Nvidia Technical Blog, Jun 4, 2025 |
| Earlier rounds | Round 5.1 index on Llama 3.1 405B at 512 GPUs, H100 in round 5.0 = 1: GB200 3.2, GB300 4.2 | Nvidia Technical Blog, Dec 11, 2025 |
| Price and power | Morgan Stanley estimate: about $7.8 million for a Vera Rubin NVL72 rack against about $4 million for a GB300 NVL72 rack | MarketScale, Jul 6 |
| Price and power | Rack TDP of 180–220 kW for Vera Rubin NVL72 against 120–140 kW for GB200 and GB300 | SemiAnalysis, Feb 25 |
| Earlier claims | At CES in January 2026 Nvidia said Rubin chips would deliver up to 5× the floating-point performance of Blackwell for inference and 3.5× for training; 72 Rubin GPUs per rack | The Register, Jan 5 |
| Endpoints | MLPerf Endpoints v1.0, Kimi K3. Vera Rubin, provisional: 32 accelerators, 1,024 concurrent queries, 41,414.2 tokens/s, 50.4 tokens/s per user, 100.89 kW. GB300, verified: 48 accelerators, 48 concurrent queries, 2,210.4 tokens/s, 56.2 tokens/s per user, 60 kW; at 1,024 concurrent queries 13,905.8 tokens/s and 16.8 tokens/s per user. Claim: over 10× the throughput per kilowatt of provisioned power at a minimum of 50 tokens/s per user | Nvidia Technical Blog, Nov 16 |
What we read. MLCommons’ release, Nvidia’s two technical blogs, AMD’s and Google’s blogs, CoreWeave’s release and the reports listed. We did not query the MLCommons results database; the times are as the vendors state them. Arithmetic. Qwen3.5 GRPO at 1,024 GPUs: 13.78 ÷ 10.44 = 1.32. Reruns: 13.151 ÷ 8.08 = 1.63; 9.211 ÷ 6.14 = 1.50; 2.94 ÷ 1.628 = 1.81; 3.871 ÷ 1.628 = 2.38. DeepSeek-V3 671B: 8.03 ÷ 4.14 = 1.94; gpt-oss 20B: 18.24 ÷ 11.213 = 1.63; Llama 3.1 8B: 11.541 ÷ 7.162 = 1.61; DLRMv4: 3.871 ÷ 2.829 = 1.37. Average of the seven listed gains: (2.0 + 1.6 + 1.6 + 1.5 + 1.5 + 1.4 + 1.4) ÷ 7 = 1.57. GB200: 12.36 ÷ 4.14 = 2.99. Round 5.1: 4.2 ÷ 3.2 = 1.31. Price: 7.8 ÷ 4 = 1.95; 1.94 ÷ 1.95 = 0.99; 1.57 ÷ 1.95 = 0.81. Power: 180 ÷ 140 = 1.29; 220 ÷ 120 = 1.83. Endpoints: 41,414.2 ÷ 100.89 = 410.5; 2,210.4 ÷ 60 = 36.8; 410.5 ÷ 36.8 = 11.1; 13,905.8 ÷ 60 = 231.8; 410.5 ÷ 231.8 = 1.77. Total throughput at equal load: 41,414.2 ÷ 13,905.8 = 2.98, from 32 accelerators against 48. Conflicts between sources. Nvidia’s table lists Qwen3.5 GRPO at 64 GPUs, at 1.4 times, and its text gives times at 1,024 GPUs, which divide to 1.32; these are different runs. We chart the listed figures. Nvidia’s index for round 5.1 gives 1.31 times for GB300 over GB200 in the same round, and the same blog cites 1.9 times against GB200’s result a round earlier. TechArena counts 104 systems and MLCommons 153; we use MLCommons. Estimates. The rack prices are a Morgan Stanley bill-of-materials estimate as relayed by MarketScale and the rack power is a SemiAnalysis estimate; Nvidia has published neither. The per-dollar figures divide a benchmark ratio by a price estimate and are illustrations. Not charted. A GB300 time on 8,192 GPUs given by one outlet; CoreWeave’s stated gains, which compare with GB200; Nvidia’s unverified reruns, except to note them; and the results for two other models in MLPerf Endpoints, which Nvidia puts at 9 and 5 times.
Wire reports are written within hours of the news, on a six-hour schedule. Each lists its sources and shows the arithmetic behind its figures. Not investment advice.