AI news videos

NVIDIA Says Vera Rubin Hits 3.7x In MLPerf Debut

MLPerf Inference v6.1 results were released on September 16, 2026, and NVIDIA's next generation Vera Rubin NVL72 made its first appearance in the suite as a preview submission.

Watch on YouTube

Transcript

Sources

Every claim in this video comes from the top ranking coverage of this topic. The claims and where each one came from:

• NARRATED UNIT: hook sentence 1, 'NVIDIA published MLPerf benchmark results today for its next generation Vera Rubin rack.' VERBATIM ARTICLE TEXT: the headline 'NVIDIA Vera Rubin NVL72 Delivers Leading Performance in MLPerf Inference v6.1 Debut'; the standfirst 'First preview submission using NVIDIA Vera Rubin NVL72 systems delivers up to 3.7x better throughput than NVIDIA GB300 NVL72'; the body line 'as highlighted by MLPerf Inference v6.1 results released today:'; and the printed dateline 'September 16, 2026'. WORDING CHANGES AND WHY: (1) The article's own phrase 'released today' was kept as the single word 'today' so the Short reads as same day news while never speaking a calendar date aloud, which the date rule forbids. (2) NVIDIA is named as the publisher in the first three words so the viewer hears immediately that the source is the vendor, not an outside tester. (3) 'NVL72' was REMOVED from the spoken product name. It is a glued brand plus digit token and the speech engine turns such tokens into invented words. The full product name is preserved on the scene 1 onscreen card and in titleCandidates. (4) 'v6.1' was likewise NOT spoken, for the same glued token reason, and 'MLPerf benchmark results' is spoken instead. The plain word MLPerf on its own is safe to speak. (5) 'next generation' is written as two unhyphenated words because hyphenated compounds make the engine stutter. It is a characterization, and its basis in the article is that the Vera Rubin rack appears only as a debut preview submission while the article repeatedly treats GB300 NVL72 as the comparison system NVIDIA ships today. (6) The article's idiom 'purpose-built to optimize across all these' was not used anywhere. (7) The article's opening three item comma list 'System performance, efficient infrastructure scaling and continuous software optimization' was dropped wholesale rather than collapsed, because any narrated sentence carrying three or more joined items drops the speech rate from roughly 2.65 to roughly 2.18 words per second and would push this Short past sixty seconds.(FIRST PARTY SOURCE. NVIDIA's own corporate blog, blogs.nvidia.com, article headline 'NVIDIA Vera Rubin NVL72 Delivers Leading Performance in MLPerf Inference v6.1 Debut', by Zhihan Jiang, dated September 16, 2026. This is NVIDIA publishing NVIDIA's own submitted MLPerf Inference v6.1 scores, on NVIDIA's own blog, about hardware NVIDIA designs and sells, and the Vera Rubin numbers are explicitly labelled by the article as a preview submission. Watcher snapshot path: scripts/data/article-cache/a425146e5e224be9.txt. Working copy path: scripts/data/production/nvidia-vera-rubin-mlperf-inference/article.txt (captured 2026-09-16T15:06:48Z). The live URL was NOT independently refetched on this run, so the watcher snapshot is the sole basis for every fact in this file.)
• NARRATED UNIT: hook sentence 2, 'The claimed gains are large.' VERBATIM ARTICLE TEXT: 'NVIDIA Vera Rubin NVL72 delivers up to 3.7x better throughput than GB300 NVL72'; 'On DeepSeek-R1, using the NVIDIA TensorRT-LLM library, throughput is up to 2.5x higher than GB300 NVL72'; and 'Vera Rubin NVL72 delivered 30x better performance than GB300 NVL72 in preview testing'. WORDING CHANGES AND WHY: (1) The word 'claimed' is a deliberate hedge. It tells the viewer in the hook itself that these are figures NVIDIA reports about NVIDIA hardware, so nothing in the hook is asserted as independently established. (2) NO figure is spoken in the hook at all. This is intentional: it keeps the spoken figure count to one per scene and puts an entire beat of separation before the first number, which reduces the chance of the transcription checker looping between similar figures. (3) The 30x figure was dropped entirely. It is attributed in the article to a separate third party agentic benchmark whose name is a banned glued token, and it is described as preview testing rather than an MLPerf result, so narrating it would overstate what the MLPerf round actually shows. (4) The 2.5x figure was also not spoken here; see the scene 1 entry for why it is carded rather than narrated. (5) This sentence exists partly to satisfy the hard requirement that the hook card render exactly two beats. A one sentence hook renders the second beat undefined and forces a full voiceover redo.(FIRST PARTY SOURCE. NVIDIA's own corporate blog, blogs.nvidia.com, article headline 'NVIDIA Vera Rubin NVL72 Delivers Leading Performance in MLPerf Inference v6.1 Debut', by Zhihan Jiang, dated September 16, 2026. This is NVIDIA publishing NVIDIA's own submitted MLPerf Inference v6.1 scores, on NVIDIA's own blog, about hardware NVIDIA designs and sells, and the Vera Rubin numbers are explicitly labelled by the article as a preview submission. Watcher snapshot path: scripts/data/article-cache/a425146e5e224be9.txt. Working copy path: scripts/data/production/nvidia-vera-rubin-mlperf-inference/article.txt (captured 2026-09-16T15:06:48Z). The live URL was NOT independently refetched on this run, so the watcher snapshot is the sole basis for every fact in this file.)
• NARRATED UNIT: scene 1, 'NVIDIA says the new rack reached up to three point seven times the throughput of the system it sells now.' VERBATIM ARTICLE TEXT: 'In its first MLPerf Inference preview submission, NVIDIA Vera Rubin NVL72 delivers up to 3.7x better throughput than GB300 NVL72.' and 'Vera Rubin NVL72 delivers up to 3.7x higher throughput than GB300 NVL72 on Qwen3-VL across offline, server and interactive scenarios, using vLLM with the NVIDIA Dynamo open source inference framework.' WORDING CHANGES AND WHY: (1) '3.7x' was written out as 'three point seven times' because every number must be spoken as words and because a digit plus letter token is exactly the shape the speech engine garbles. (2) The qualifier 'up to' was PRESERVED word for word. Dropping it would turn a best case figure into a flat claim and would misstate the source. (3) 'GB300 NVL72' was replaced with the plain description 'the system it sells now'. The token is banned from narration, and the plain phrase also tells a general viewer the useful thing, which is that the comparison baseline is NVIDIA's current shipping rack rather than a rival product. The exact part name is carried on this scene's onscreen card. (4) 'Qwen3-VL' was not spoken. It is a banned hyphenated glued token and the model name carries no meaning for a general audience. It is preserved in this audit and the benchmark comparison stays on the card. (5) The three item list 'across offline, server and interactive scenarios' was DROPPED entirely rather than collapsed, for the speech rate reason stated in the first source entry. (6) 'vLLM' and 'NVIDIA Dynamo' were dropped for word budget and because vLLM is an initialism style token. (7) THE 2.5x DEEPSEEK FIGURE WAS DELIBERATELY DROPPED FROM NARRATION. Three point seven and two point five are similar figures sitting in adjacent sentences of the same article paragraph, and speaking two similar figures close together is the precise pattern that sends the transcription checker into a repetition loop and corrupts the render audit. Only one of the pair is spoken. (8) Attribution is explicit through the opening 'NVIDIA says', so no part of this scene is narrated as independently established fact.(FIRST PARTY SOURCE. NVIDIA's own corporate blog, blogs.nvidia.com, article headline 'NVIDIA Vera Rubin NVL72 Delivers Leading Performance in MLPerf Inference v6.1 Debut', by Zhihan Jiang, dated September 16, 2026. This is NVIDIA publishing NVIDIA's own submitted MLPerf Inference v6.1 scores, on NVIDIA's own blog, about hardware NVIDIA designs and sells, and the Vera Rubin numbers are explicitly labelled by the article as a preview submission. Watcher snapshot path: scripts/data/article-cache/a425146e5e224be9.txt. Working copy path: scripts/data/production/nvidia-vera-rubin-mlperf-inference/article.txt (captured 2026-09-16T15:06:48Z). The live URL was NOT independently refetched on this run, so the watcher snapshot is the sole basis for every fact in this file.)
• NARRATED UNIT: scene 2, two sentences, 'The results show ninety nine percent scaling efficiency as NVIDIA added more racks.' and 'Throughput grew almost in proportion to the hardware.' VERBATIM ARTICLE TEXT: 'A 288-GPU submission across four GB300 NVL72 racks achieved 99% scaling efficiency, with throughput growing nearly linearly from a single-rack baseline.' and 'NVIDIA's DeepSeek-R1 (DSR1) submission scaled from a single GB300 NVL72 rack (72 GPUs) to four racks (288 GPUs), achieving 99% scaling efficiency in the offline scenario. Throughput grew nearly in proportion to the hardware added.' WORDING CHANGES AND WHY: (1) '99%' was written out as 'ninety nine percent' with no hyphen, because numbers must be spoken as words and hyphenated compounds make the engine stutter. (2) EVERY OTHER FIGURE IN THIS SECTION WAS KEPT OUT OF NARRATION. The rack count of four, the single rack baseline of seventy two GPUs and the total of two hundred and eighty eight GPUs are all numeric figures, and only one figure may be spoken per scene and never two in one sentence. 'four racks' became the plain comparative 'more racks' for that reason alone. The GPU count is carried on this scene's onscreen card instead. (3) '288-GPU' and 'single-rack' are hyphenated compounds and were avoided in spoken form for the same garbling reason. (4) 'GB300 NVL72' was not spoken; the subject is carried implicitly from scene 1's 'the system it sells now' and explicitly on the card. (5) 'throughput growing nearly linearly' is a mathematical characterization that would read as jargon aloud, so the article's own plainer second phrasing was used instead, shortened from 'grew nearly in proportion to the hardware added' to 'grew almost in proportion to the hardware'. This is the article's own literal wording, not an added metaphor. (6) 'in the offline scenario' was dropped for budget; it narrows the claim in the source, and this audit records that the ninety nine percent figure is specifically an offline scenario result. (7) Attribution is carried by the opening 'The results show', which points at the benchmark round rather than asserting the number as independent fact. (8) The article's separate text to video rack scale figures of zero point six five videos per second at five point seven seconds per video, nine times higher throughput and seven point five times lower latency, were all dropped. They are a four figure cluster in one sentence and narrating any of them would breach the one figure per scene rule.(FIRST PARTY SOURCE. NVIDIA's own corporate blog, blogs.nvidia.com, article headline 'NVIDIA Vera Rubin NVL72 Delivers Leading Performance in MLPerf Inference v6.1 Debut', by Zhihan Jiang, dated September 16, 2026. This is NVIDIA publishing NVIDIA's own submitted MLPerf Inference v6.1 scores, on NVIDIA's own blog, about hardware NVIDIA designs and sells, and the Vera Rubin numbers are explicitly labelled by the article as a preview submission. Watcher snapshot path: scripts/data/article-cache/a425146e5e224be9.txt. Working copy path: scripts/data/production/nvidia-vera-rubin-mlperf-inference/article.txt (captured 2026-09-16T15:06:48Z). The live URL was NOT independently refetched on this run, so the watcher snapshot is the sole basis for every fact in this file.)
• NARRATED UNIT: scene 3, 'NVIDIA also reports software updates alone delivered up to one point six times higher performance than the previous round.' VERBATIM ARTICLE TEXT: 'Software optimizations in NVIDIA's MLPerf Inference v6.1 submissions delivered up to 1.6x higher performance over v6.0.' and 'In v6.1, GB300 NVL72 performance on Qwen3-VL improved up to 1.6x over v6.0 results. The gains came through lower KV cache precision, additional kernel fusion, better kernels and disaggregated serving with vLLM and NVIDIA Dynamo.' WORDING CHANGES AND WHY: (1) '1.6x' became 'one point six times' because numbers must be spoken as words. (2) 'v6.1' and 'v6.0' are both banned glued tokens and were replaced by 'the previous round', which conveys the only thing a general viewer needs, namely that this is a comparison against the last MLPerf Inference round of the same benchmark. The exact version pair is preserved on the scene 3 onscreen card and here. (3) 'up to' was preserved word for word so the best case qualifier is not lost. (4) 'Software optimizations' became 'software updates alone'. The word 'alone' is the point of the scene and is supported directly by the article, which attributes this gain to software changes on the same GB300 NVL72 hardware rather than to new silicon. (5) THE FOUR ITEM CAUSE LIST 'lower KV cache precision, additional kernel fusion, better kernels and disaggregated serving' WAS DROPPED WHOLESALE. It is a multi item list, it contains the initialism KV, and narrating it would drop the speech rate and blow the sixty second ceiling. (6) 'Qwen3-VL' and 'GB300 NVL72' were again not spoken, for the banned token reason. (7) Attribution is explicit through 'NVIDIA also reports'. (8) One point six sits two scenes away from three point seven and one scene away from ninety nine percent, which are structurally different figures, so no two similar numbers are spoken close enough together to trigger the transcription repetition loop.(FIRST PARTY SOURCE. NVIDIA's own corporate blog, blogs.nvidia.com, article headline 'NVIDIA Vera Rubin NVL72 Delivers Leading Performance in MLPerf Inference v6.1 Debut', by Zhihan Jiang, dated September 16, 2026. This is NVIDIA publishing NVIDIA's own submitted MLPerf Inference v6.1 scores, on NVIDIA's own blog, about hardware NVIDIA designs and sells, and the Vera Rubin numbers are explicitly labelled by the article as a preview submission. Watcher snapshot path: scripts/data/article-cache/a425146e5e224be9.txt. Working copy path: scripts/data/production/nvidia-vera-rubin-mlperf-inference/article.txt (captured 2026-09-16T15:06:48Z). The live URL was NOT independently refetched on this run, so the watcher snapshot is the sole basis for every fact in this file.)
• NARRATED UNIT: scene 4, the honest framing scene, two sentences, 'These are preview results that NVIDIA submitted itself.' and 'NVIDIA notes some gains are not yet verified by the benchmark group.' VERBATIM ARTICLE TEXT: 'In its first MLPerf Inference preview submission'; 'NVIDIA submitted Vera Rubin NVL72 preview results on two of the most demanding benchmarks in the MLPerf Inference v6.1 suite: DeepSeek-R1 and Qwen3-VL.'; 'These early results showcase NVIDIA's accelerated pace of innovation and how performance will improve with continuous software optimizations.'; and 'Software optimization continued past the v6.1 submission deadline as well. Post-submission results, not yet verified by MLCommons, on GPT-OSS-120B and DLRMv3 show further performance gains.' WORDING CHANGES AND WHY: (1) 'preview results' and 'preview submission' are the article's own words and are spoken back plainly. This is the single most important honest point in the Short: the headline three point seven times figure is a preview submission, not a verified shipping product result. (2) 'that NVIDIA submitted itself' restates the article's own 'NVIDIA submitted' and makes the vendor origin of the scores audible. It is a description of the document, not an accusation, and no word in the scene suggests wrongdoing. (3) 'not yet verified' is taken almost verbatim from 'not yet verified by MLCommons'. The scope was widened from that one sentence to 'some gains' rather than all results, which is accurate: the article applies the caveat specifically to the post submission gains, and 'some' is the honest quantifier. (4) 'MLCommons' was rendered as 'the benchmark group' because the proper noun is unfamiliar to a general viewer and because the plain phrase states its role. The name is recorded here in full. (5) 'GPT-OSS-120B' and 'DLRMv3' were not spoken; both are banned glued brand plus digit tokens. (6) The article's idiom 'accelerated pace of innovation' and its closing figure of speech about edge devices and AI factories were avoided entirely. (7) No figure is spoken in this scene at all, which keeps the whole script to three spoken figures. APPENDIX A, SENTENCE BY SENTENCE AUDIT: hook sentence 1 traces to the headline, the standfirst and 'as highlighted by MLPerf Inference v6.1 results released today'; hook sentence 2 traces to the three reported multiples, 3.7x, 2.5x and 30x, and is deliberately hedged with 'claimed'; scene 1 traces to 'NVIDIA Vera Rubin NVL72 delivers up to 3.7x better throughput than GB300 NVL72'; scene 2 sentence 1 traces to 'A 288-GPU submission across four GB300 NVL72 racks achieved 99% scaling efficiency'; scene 2 sentence 2 traces to 'Throughput grew nearly in proportion to the hardware added'; scene 3 traces to 'Software optimizations in NVIDIA's MLPerf Inference v6.1 submissions delivered up to 1.6x higher performance over v6.0'; scene 4 sentence 1 traces to 'In its first MLPerf Inference preview submission' and 'NVIDIA submitted Vera Rubin NVL72 preview results'; scene 4 sentence 2 traces to 'Post-submission results, not yet verified by MLCommons'; the endcard is house copy about Inboxsmith and makes no claim about the story. APPENDIX B, BUDGET COUNT: hook 18 words in 2 sentences; scene 1 narration 20 words in 1 sentence; scene 2 narration 21 words in 2 sentences; scene 3 narration 19 words in 1 sentence; scene 4 narration 20 words in 2 sentences; endcard narration 17 words in 2 sentences. TOTAL NARRATED WORDS 115, inside the 108 to 118 target band and well under the 132 hard gate. TOTAL NARRATED SENTENCES 10, at or under the 12 sentence ceiling. Hook is exactly 2 sentences and 18 words, under the 20 word ceiling. No scene narration exceeds 30 words; the longest is scene 2 at 21. TOTAL COMMAS ACROSS ALL NARRATION: ZERO. No narrated sentence contains a multi item list of any kind, comma joined or otherwise. FIGURES SPOKEN: exactly three, 'three point seven times' in scene 1, 'ninety nine percent' in scene 2 and 'one point six times' in scene 3, all written as words, never more than one per sentence or per scene, with no figure at all in the hook, in scene 4 or in the endcard. The two similar multiples in the article, 3.7x and 2.5x, are never both spoken; only 3.7 is narrated and 2.5 is carried in this audit. Scene durationS integers are 11, 11, 11 and 11, summing to 44, inside the 30 to 50 quality gate band. No timeline key exists in this file, deliberately, because the story is hours old and a timeline block would age it to the catalog freshness tier. No em dash, en dash or double hyphen appears anywhere in this file. No banned glued token appears in any narration string. APPENDIX C, HONESTY CAVEAT: every substantive claim in the narration is attributed out loud, using the phrases 'NVIDIA published', 'claimed', 'NVIDIA says', 'The results show', 'NVIDIA also reports', 'NVIDIA submitted itself' and 'NVIDIA notes'. Nothing is narrated as independently established fact. The entire Short rests on one document written and published by NVIDIA about NVIDIA hardware, reporting scores NVIDIA itself submitted, and scene 4 exists specifically to tell the viewer that plainly. The Vera Rubin numbers are a preview submission by the article's own description, and the article itself concedes that some post submission gains are not yet verified by MLCommons. The narration makes no accusation of bias or wrongdoing; it states what the document is and what has not yet been independently confirmed. No calendar date is spoken aloud, and no countdown, expiry, deadline or last chance phrasing appears anywhere in this file. APPENDIX D, CORROBORATION STATUS: the transcripts research step captured ZERO of five YouTube transcripts on this run. transcripts.json holds an empty videos array with fetchedAt 2026-09-16T15:42:43Z. That is a known out of date yt-dlp build on this host and carries no signal whatsoever about the story. timeline-probes.json holds exactly ONE probe, an unrelated third party video title dated 2025-11-21 comparing NVIDIA GPUs with chips from other companies, marked 'taken down banned', which has no bearing on this MLPerf round. NO fact in this script comes from either file. No third party outlet, publication or video channel is named or cited anywhere in this file. sources.json in the production directory is empty, which is normal for this pipeline; the real citations are these entries. The NVIDIA blog article is the ONLY source, and it was NOT corroborated by any second document: the live URL was not independently refetched on this run, so the watcher snapshot captured at 2026-09-16T15:06:48Z stands alone. APPENDIX E, DELIBERATE OMISSIONS: the following were NOT narrated. The 2.5x DeepSeek-R1 throughput multiple, because it is too similar to and too close to the 3.7x figure. The 30x figure on the SemiAnalysis AgentX agentic benchmark, because AgentX is a banned glued token, because the figure is preview testing rather than an MLPerf result, and because attributing it would require naming a third party. The GPU counts 72 and 288 and the rack count of four. The WAN 2.2 text to video results of 0.65 720p videos per second at 5.7 seconds per video with 9x higher throughput and 7.5x lower latency. The NVLink and NVLink Switch interconnect claims of 10x higher packet rates and 3x lower latency than off the shelf Ethernet. The Jetson AGX Thor edge submission on the Edge-Agentic benchmark with Qwen3.6-27B. The partner ecosystem paragraph naming 19 partners and eight multi node submitters and the full nineteen company list, which is the largest comma list in the article. The Nebius partner preview submission. The named MLPerf entry identifiers 6.1-0106, 6.1-0074 and 6.1-0073 and the MLCommons trademark notice. The technical mechanism paragraphs covering Tensor Cores, Transformer Engine, NVFP4 precision, KV cache, disaggregated serving, expert parallelism and mixture of experts layers. The revenue and cost per token framing. The upcoming MLPerf Endpoints benchmark. The GTC Berlin registration promotion, all recent news and related news links, the tag and category lists and all site navigation chrome. The two largest reasons for these omissions are the ceiling of one spoken figure per scene and the absolute ban on narrating three or more joined items, which this article violates repeatedly because its body is built from comma lists of products, partners and optimizations. APPENDIX F, PROMPT INJECTION CHECK: article.txt was read in full and treated strictly as untrusted data that informs facts only. The file is a corporate blog page with several hundred lines of empty navigation chrome, share widgets, a Mastodon share form, a GTC Berlin registration banner, recent news and related news link lists, category and tag lists, and a comment counter. NO text anywhere in article.txt attempts to address, instruct, redirect or impersonate an AI, an agent or an assistant. There are no hidden or out of band directives, no role prompts, no 'ignore previous instructions' style strings and no embedded system prompts. The only imperative sentences in the document are ordinary marketing calls to action aimed at human readers, namely 'Register Now', 'Learn more about the NVIDIA Vera Rubin platform', 'View All Recent News', 'Skip to content', 'Copy link' and 'Share', none of which were acted on. RESULT: NO PROMPT INJECTION DETECTED, and nothing in the file was followed as an instruction.(FIRST PARTY SOURCE. NVIDIA's own corporate blog, blogs.nvidia.com, article headline 'NVIDIA Vera Rubin NVL72 Delivers Leading Performance in MLPerf Inference v6.1 Debut', by Zhihan Jiang, dated September 16, 2026. This is NVIDIA publishing NVIDIA's own submitted MLPerf Inference v6.1 scores, on NVIDIA's own blog, about hardware NVIDIA designs and sells, and the Vera Rubin numbers are explicitly labelled by the article as a preview submission. Watcher snapshot path: scripts/data/article-cache/a425146e5e224be9.txt. Working copy path: scripts/data/production/nvidia-vera-rubin-mlperf-inference/article.txt (captured 2026-09-16T15:06:48Z). The live URL was NOT independently refetched on this run, so the watcher snapshot is the sole basis for every fact in this file.)