As large language models (LLMs) grow larger, they get smarter, with open models from leading developers now featuring hundreds of billions of parameters. At the same time, today's leading models are also capable of reasoning, which means that they generate many intermediate reasoning tokens before delivering a final response to the user. The combination of these two trends-larger models that think using more tokens-drives the need for significantly higher compute performance.
Delivering the highest performance on production workloads takes a state-of-the-art technology stack-spanning chips, systems, and software-and an expansive developer ecosystem that is constantly building on that stack.
MLPerf Inference v5.1 is the latest version of the MLPerf Inference industry standard benchmark. With benchmark rounds held twice per year, the benchmark features many tests of AI inference performance and is regularly updated with new models and scenarios. This round features:
DeepSeek-R1 - a popular 671-billion parameter mixture-of-experts (MoE) reasoning model, developed by DeepSeek. In the server scenario, the time-to-first-token (TTFT) threshold is 2 seconds with a 12.5 tokens/second/user (TPS/user) target. All TPS/user targets are 99th percentile, meaning that 99% of tokens meet or exceed that TPS/user speed.
Llama 3.1 405B - MLPerf Inference v5.1 adds a new interactive scenario for the largest of the Llama 3.1 series of models, providing a faster 12.5 TPS/user threshold with a shorter 4.5 second TTFT requirement compared to the existing server scenario.
Llama 3.1 8B - an 8-billion parameter member of the Llama 3.1 series of models with offline, server (2 second TTFT, 10 TPS/user), and interactive (0.5 second TTFT, 33 TPS/user) scenarios. This replaces the GPT-J benchmark used in prior rounds.
Whisper - a popular speech recognition model that recently saw nearly 5 million downloads in a month on HuggingFace. This replaces RNN-T, which was featured in prior editions of the MLPerf Inference benchmark suite.
This round, NVIDIA submitted the first results using the new Blackwell Ultra architecture, announced in March. It came just six months after Blackwell made its debut in the available category in MLPerf Inference v5.0, setting new inference performance records. Additionally, the NVIDIA platform set new performance records on all newly added benchmarks this round-DeepSeek-R1, Llama 3.1 405B, Llama 3.1 8B, and Whisper-and continues to hold per-GPU performance records on all other MLPerf inference benchmarks.
MLPerf Inference Per-Accelerator Records
Benchmark Offline Server Interactive
DeepSeek-R1 5,842 tokens/second/GPU 2,907 tokens/second/GPU **
Llama 3.1 405B 224 tokens/second/GPU 170 tokens/second/GPU 138 tokens/second/GPU
Llama 2 70B 99.9% 12,934 tokens/second/GPU 12,701 tokens/second/GPU 7,856 tokens/second/GPU
Llama 2 70B 99% 13,015 tokens/second/GPU 12,701 tokens/second/GPU 7,856 tokens/second/GPU
Llama 3.1 8B 18,370 tokens/second/GPU 16,099 tokens/second/GPU 15,284 tokens/second/GPU
Stable Diffusion XL 4.07 samples/second/GPU 3.59 queries/second/GPU **
Mixtral 8x7B 16,099 tokens/second/GPU 16,131 tokens/second/GPU **
DLRMv2 99% 87,228 samples/second/GPU 80,515 samples/second/GPU **
DLRMv2 99.9% 48,666 samples/second/GPU 46,259 queries/second/GPU **
Whisper 5,667 tokens/second/GPU ** **
R-GAT 81,404 samples/second/GPU ** **
Retinanet 1,875 samples/second/GPU 1,801 queries/second/GPU **
Table 1. Performance records per GPU based on submissions powered by the NVIDIA platform. MLPerf Inference v5.0 and v5.1, Closed Division. Results retrieved from www.mlcommons.org on September 9, 2025. NVIDIA platform results from the following entries: 5.0-0072, 5.1-0007, 5.1-0053, 5.1-0079, 5.1-0028, 5.1-0062, 5.1-0086, 5.1-0073, 5.1-0008, 5.1-0070,5.1-0046, 5.1-0009, 5.1-0060, 5.1-0072. 5.1-0071, 5.1-0069 Per chip performance derived by dividing total throughput by number of reported chips. Per-chip performance is not a primary metric of MLPerf Inference v5.0 or v5.1.The MLPerf name and logo are registered and unregistered trademarks of MLCommons Association in the United States and other countries. All rights reserved. Unauthorized use strictly prohibited. See www.mlcommons.org for more information.
NVIDIA also made extensive use of NVFP4 acceleration across all DeepSeek-R1 and Llama model submissions using the Blackwell and Blackwell Ultra architectures.
In this post, we take a closer look at these performance results and the full-stack technologies that enabled them.
Blackwell Ultra sets reasoning records in MLPerf debut This round, NVIDIA submitted results in the available category using the GB300 NVL72 rack-scale system, the first-ever MLPerf submissions using the Blackwell Ultra architecture. Blackwell Ultra builds upon the many advances in the NVIDIA Blackwell architecture, with several key enhancements:
1.5x higher peak NVFP4 AI compute
2x higher attention-layer compute
1.5x higher HBM3e capacity
Compared to the GB200 NVL72 submission, GB300 NVL72 delivered up to 45% higher performance per GPU, setting the standard on the new DeepSeek-R1 benchmark. And compared to unverified results collected on a Hopper-based system, Blackwell Ultra delivered about 5x higher throughput per GPU-translating into significantly higher AI factory throughput and much lower cost per token.
DeepSeek-R1 Performance
Architecture Offline Server
Hopper 1,253 tokens/second/GPU 556 tokens/second/GPU
Blackwell Ultra 5,842 tokens/second/GPU 2,907 tokens/second/GPU
Blackwell Ultra Advantage 4.7x 5.2x
Table 2. Per-GPU performance on DeepSeek-R1. MLPerf Inference v5.1, Closed. Blackwell Ultra results based on results in entry 5.1-0072. Hopper results not verified by MLCommons Association. Per-GPU performance is not a primary metric of MLPerf Inference v5.1 and is calcu
Most recent headlines
05/01/2027
Worlds first 802.15.4ab-UWB chip verified by Calterah and Rohde & Schwarz to be ...
06/09/2026
June 9 2026, 23:00 (PDT) Dolby and MagentaTV Bring Fans Closer to the FIFA Worl...
04/08/2026
Dalet, a leading technology and service provider for media-rich organizations, t...
04/07/2026
April 7 2026, 19:00 (PDT) Detective Conan: Fallen Angel of the Highway Opens in...
15/06/2026
Detach from Direct-Attached: How Remote Editing with EVO Keeps Creative Teams Mo...
14/06/2026
Library captures 1960s R&B/pop drum sound
Following on from their recent wave of plug-in effects, Iconic Instruments have just launched an all-new virtual d...
14/06/2026
HBO Comedy Rooster Shot with URSA Cine 17K 65
Brie Clayton June 14, 2026
0 Comments
Large format brings viewers intimately close to characters.
Black...
13/06/2026
Latest expansion pack includes 252 presets
Devious Machines have recently introduced another expansion for their powerful multi-effects plug-in, Infiltrator...
13/06/2026
Create custom DAW/plug-in controllers using prompts
MetaGrid have recently introduced an all-new AI Builder function to their touchscreen-based control surf...
13/06/2026
Share
Copy link
Facebook
X
Linkedin
Bluesky
Email...
13/06/2026
Share
Copy link
Facebook
X
Linkedin
Bluesky
Email...
12/06/2026
YES Network and The Gotham Sports App will air seven Athletes Unlimited Softball...
12/06/2026
The United Football League will host its FAST Innovation Suite at the 2026 United Bowl presented by Credit One Bank on Saturday, June 13 at 3:00 p.m. ET at Audi...
12/06/2026
PTZOptics and LayerJot will present live demonstrations at InfoComm 2026 showing how natural-language AI prompting, robotic camera control, and on-device comput...
12/06/2026
MultiDyne Video and Fiber Optic Systems will exhibit at InfoComm 2026, featuring...
12/06/2026
Ateme has announced that Eurovision Services is using Ateme's software-based frame-rate conversion technology for international live event workflows. The de...
12/06/2026
Bitmovin and Simplestream have announced a partnership with Xperi to simplify the launch of OTT streaming services on TiVo OS smart TVs and devices. The collabo...
12/06/2026
Net Insight has announced that a multinational technology company is deploying a...
12/06/2026
MLB Players Inc., the business arm of the MLB Players Association, has announced a partnership with Athletes First to develop and sell brand partnerships across...
12/06/2026
Guntermann and Drunck (G&D) and VuWall have announced the CommandKeyboard-Advanc...
12/06/2026
Comcast Smart Solutions announces a new smart technology deployment with Major L...
12/06/2026
Elevation Worship completed the initial leg of its Elevation Nights 2026 tour ...
12/06/2026
AJA Video Systems has announced KONA IP25 support for Colorfront Transkoder and ...
12/06/2026
Audinate Group Limited (ASX: AD8) will exhibit at InfoComm 2026 (Booth C7321, Ce...
12/06/2026
Pac-12 Commissioner Teresa Gould has announced the appointment of Scott Adametz as Chief Technology Officer. The Pac-12 describes the hire as the first CTO appo...
12/06/2026
Grass Valley has announced AMPP Edge Live, a production system combining Grass Valley hardware, NVIDIA Blackwell GPU acceleration, and AMPP OS in a single platf...
12/06/2026
At one time a trailblazer with the launch of the Longhorn Network, the Universit...
12/06/2026
Ratings Roundup is a rundown of recent rating news and is derived from press rel...
12/06/2026
Chyron has announced PAINT 10.4, an update to its illustrated replay and sports ...
12/06/2026
SVP, Production, Mark Gross: With the new schedule, with not having every Sunday night, it has given us an opportunity to take a step back and reimagine what o...
12/06/2026
For Televisa Technical Engineering Manager Roberto N nez Ibarra and the small team of 12 technicians and two production personnel at the IBC things are already ...
12/06/2026
Simple Steps to Better Acoustics - Taming The Small Room
Most of us mix in spare rooms and small spaces, where the acoustics fight us at every turn. At Gear...
12/06/2026
Latest addition expands vintage-inspired effects palette
Meris' Ottobit pedal range draws its inspiration from vintage gaming consoles, and the latest a...
12/06/2026
Soundbox-based chamber strings series expanded
Sonora Cinematic have just announced the launch of the second instalment in their Soundbox-based chamber stri...
12/06/2026
Share
Copy link
Facebook
X
Linkedin
Bluesky
Email...
12/06/2026
Share
Copy link
Facebook
X
Linkedin
Bluesky
Email...
12/06/2026
AJA Announces KONA IP25 Integration with Colorfront Software
Brie Clayton June 12, 2026
0 Comments
Collaboration enables uncompressed SMPTE ST 2110 I/O ...
12/06/2026
URSA Cine 12K LF Used to Create Visuals for STUTS' K-Arena Concert
Brie Clayton June 12, 2026
0 Comments
Organic visuals projected on a giant scre...
12/06/2026
MTI FILM Acquires Mango New Edit, Expanding its Global Post-Production Services ...
12/06/2026
AI Point Tracking Speeds Up Complex VFX Tracks in Mocha Pro
Jessie Electa Petrov June 12, 2026
0 Comments
The 2026.5 release adds automatic point trac...
12/06/2026
Bitmovin, a provider of video streaming solutions, has partnered with Simplestream, a provider of OTT and broadcast solutions, and technology provider Xperi, to...
12/06/2026
Leostream Corporation, creator of the world-leading Leostream Remote Desktop Access Platform, today announced Jigsaw24, a leading B2B IT solutions provider wit...
12/06/2026
Share
Copy link
Facebook
X
Linkedin
Bluesky
Email...
12/06/2026
Share
Copy link
Facebook
X
Linkedin
Bluesky
Email...
12/06/2026
Share
Copy link
Facebook
X
Linkedin
Bluesky
Email...
12/06/2026
Share
Copy link
Facebook
X
Linkedin
Bluesky
Email...
12/06/2026
Share
Copy link
Facebook
X
Linkedin
Bluesky
Email...
12/06/2026
How Aussie indie games and screen are levelling up with IP 11 June 2026
Ari Harrison, Pro Jank Footy
Head of Games Joey Egger and Ari Harrison of Umbrella sha...
12/06/2026
AgentPerf from Artificial Analysis, the industry's first agentic AI benchmark, gives developers, enterprises and infrastructure providers a clear way to com...
As large language models (LLMs) grow larger, they get smarter, with open models from leading developers now featuring hundreds of billions of parameters. At the same time, today's leading models are also capable of reasoning, which means that they generate many intermediate reasoning tokens before delivering a final response to the user. The combination of these two trends-larger models that think using more tokens-drives the need for significantly higher compute performance. 
