Large language models and the applications they power enable unprecedented opportunities for organizations to get deeper insights from their data reservoirs and to build entirely new classes of applications.
But with opportunities often come challenges.
Both on premises and in the cloud, applications that are expected to run in real time place significant demands on data center infrastructure to simultaneously deliver high throughput and low latency with one platform investment.
To drive continuous performance improvements and improve the return on infrastructure investments, NVIDIA regularly optimizes the state-of-the-art community models, including Meta's Llama, Google's Gemma, Microsoft's Phi and our own NVLM-D-72B, released just a few weeks ago.
Relentless Improvements Performance improvements let our customers and partners serve more complex models and reduce the needed infrastructure to host them. NVIDIA optimizes performance at every layer of the technology stack, including TensorRT-LLM, a purpose-built library to deliver state-of-the-art performance on the latest LLMs. With improvements to the open-source Llama 70B model, which delivers very high accuracy, we've already improved minimum latency performance by 3.5x in less than a year.
We're constantly improving our platform performance and regularly publish performance updates. Each week, improvements to NVIDIA software libraries are published, allowing customers to get more from the very same GPUs. For example, in just a few months' time, we've improved our low-latency Llama 70B performance by 3.5x.
NVIDIA has increased performance on the Llama 70B model by 3.5x. In the most recent round of MLPerf Inference 4.1, we made our first-ever submission with the Blackwell platform. It delivered 4x more performance than the previous generation.
This submission was also the first-ever MLPerf submission to use FP4 precision. Narrower precision formats, like FP4, reduces memory footprint and memory traffic, and also boost computational throughput. The process takes advantage of Blackwell's second-generation Transformer Engine, and with advanced quantization techniques that are part of TensorRT Model Optimizer, the Blackwell submission met the strict accuracy targets of the MLPerf benchmark.
Blackwell B200 delivers up to 4x more performance versus previous generation on MLPerf Inference v4.1's Llama 2 70B workload. Improvements in Blackwell haven't stopped the continued acceleration of Hopper. In the last year, Hopper performance has increased 3.4x in MLPerf on H100 thanks to regular software advancements. This means that NVIDIA's peak performance today, on Blackwell, is 10x faster than it was just one year ago on Hopper.
These results track progress on the MLPerf Inference Llama 2 70B Offline scenario over the past year. Our ongoing work is incorporated into TensorRT-LLM, a purpose-built library to accelerate LLMs that contain state-of-the-art optimizations to perform inference efficiently on NVIDIA GPUs. TensorRT-LLM is built on top of the TensorRT Deep Learning Inference library and leverages much of TensorRT's deep learning optimizations with additional LLM-specific improvements.
Improving Llama in Leaps and Bounds More recently, we've continued optimizing variants of Meta's Llama models, including versions 3.1 and 3.2 as well as model sizes 70B and the biggest model, 405B. These optimizations include custom quantization recipes, as well as efficient use of parallelization techniques to more efficiently split the model across multiple GPUs, leveraging NVIDIA NVLink and NVSwitch interconnect technologies. Cutting-edge LLMs like Llama 3.1 405B are very demanding and require the combined performance of multiple state-of-the-art GPUs for fast responses.
Parallelism techniques require a hardware platform with a robust GPU-to-GPU interconnect fabric to get maximum performance and avoid communication bottlenecks. Each NVIDIA H200 Tensor Core GPU features fourth-generation NVLink, which provides a whopping 900GB/s of GPU-to-GPU bandwidth. Every eight-GPU HGX H200 platform also ships with four NVLink Switches, enabling every H200 GPU to communicate with any other H200 GPU at 900GB/s, simultaneously.
Many LLM deployments use parallelism over choosing to keep the workload on a single GPU, which can have compute bottlenecks. LLMs seek to balance low latency and high throughput, with the optimal parallelization technique depending on application requirements.
For instance, if lowest latency is the priority, tensor parallelism is critical, as the combined compute performance of multiple GPUs can be used to serve tokens to users more quickly. However, for use cases where peak throughput across all users is prioritized, pipeline parallelism can efficiently boost overall server throughput.
The table below shows that tensor parallelism can deliver over 5x more throughput in minimum latency scenarios, whereas pipeline parallelism brings 50% more performance for maximum throughput use cases.
For production deployments that seek to maximize throughput within a given latency budget, a platform needs to provide the ability to effectively combine both techniques like in TensorRT-LLM.
Read the technical blog on boosting Llama 3.1 405B throughput to learn more about these techniques.
Different scenarios have different requirements, and parallelism techniques bring optimal performance for each of these scenarios. The Virtuous Cycle Over the lifecycle of our architectures, we deliver significant performance gains from ongoing software tuning and optimization. These improvements translate into additional value for customers who train and deploy on our platforms. They're able to create more capable models and applications and deploy their existing models using less infrastructure, enhancing th
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...
10/06/2026
DAZN-owned digital-media company launches three fan-first series leaning into cr...
10/06/2026
Clear-Com has announced the appointment of Jason Dino as Southwest Regional Sales Manager USA, covering Southern California and the Southwest region. Dino joins...
10/06/2026
An 11% decrease in number of global broadcast deals reflects the organization...
10/06/2026
The Women Without Boundaries Awards recognize women whose work is advancing the future of media, broadcast, AV, workplace technology, digital experience, and re...
10/06/2026
Today is match day minus two for FIFA and HBS. On Thursday, there will be two ma...
10/06/2026
SES is supporting broadcast distribution of the world's biggest football tou...
10/06/2026
NDI has announced that BirdDog has become the first hardware manufacturer to achieve full NDI 6.3 compatibility across its complete lineup of cameras, encoders,...
10/06/2026
Vince Caputo and Scott Carter, winners of the 2026 Sports Emmy for Outstanding Post Produced Audio have been announced as presenters for the 2026 SVG Advanced A...
10/06/2026
FOX One today unveiled a slate of new product features and enhancements designed to elevate the viewing experience for fans on the official streaming platform o...
10/06/2026
Primary production partners Dome Productions and Raycom Sports once again played...
10/06/2026
The Women in Sports Filmmaking Experienceship is an immersive professional devel...
10/06/2026
The league has expanded its HSAN architecture for the NBA Finals to manage more ...
10/06/2026
The inaugural MoonPay X Games League (XGL) Winter Draft will take place Wednesday, September 16, 2026 at Cosm Los Angeles from 7-9 p.m. PT. The event will strea...
10/06/2026
The University of Oklahoma (OU) Athletics Department and Learfield have announce...
10/06/2026
The Video Services Forum (VSF) has released TR-06-4 Part 8, a new specification for RIST Satellite-Hybrid: Out-of-Band Method. The specification creates a mecha...
10/06/2026
Riedel Communications provided the communications infrastructure for the 14th World Live Neurovascular Conference (WLNC) in Lisbon, supporting live medical proc...
10/06/2026
A still from The Doom Generation by Gregg Araki (Courtesy of Sundance Institute)
By Lucy Spicer
Have you checked out our Sundance Film Festival 101 list yet...
10/06/2026
Get Hands-On with Interfaces & Mic Preamp Brands
If youre after a new interface or preamp, then GearExpo UK is the place to be! Well have a whole host of au...
10/06/2026
November 13-14 2026, The Midway, San Francisco
Following their recent rebranding, MONO Music Conference (formerly Music Expo) have officially announced thei...
10/06/2026
Debut instrument free for limited time
deFORM is the debut release from newly founded developer ebbandflow, and it's being offered as a free download fo...
10/06/2026
Alone Australia Season 4: Meet the Cast
10 June, 2026
Media releases
WATCH THE TRAILER
Smash-hit survival series Alone Australia drops its highly anticipa...
10/06/2026
DEADLY THEN, DEADLY NOW, DEADLY ALWAYS: SBS & NITV IGNITE NAIDOC WEEK 2026 WITH ...
10/06/2026
Rohde & Schwarz and TRUMPF advance laser-based drone defense with THORIS LCS Rohde & Schwarz is showcasing THORIS at ILA 2026: A sovereign, end to end counter...
10/06/2026
MAHLE and Rohde & Schwarz develop application for sensor testing of modern drive...
10/06/2026
Production and Development Funding supports the creation of compelling, commercially viable, artistic and culturally relevant South African screen content. Deve...
10/06/2026
Media buyers and sellers can now compare YouTube reach from computer, mobile, an...
10/06/2026
Accedo, Humans not Robots, and the Institution of Engineering and Technology (IET) have announced the launch of Ecoflow X. Formerly an IBC Accelerator project, ...
10/06/2026
Frequency, the engine powering the world's leading streaming television channels, today announced that James Smith has joined the company as General Manager...
10/06/2026
At the 14th World Live Neurovascular Conference (WLNC) in Lisbon, Riedel Communications provided the communications infrastructure for live medical procedures s...
10/06/2026
Globecast, a leading provider of broadcast, media, and entertainment managed services, today announced the launch of its Content Exchange platform powered by Or...
10/06/2026
Klvr will showcase how venues, integrators and production teams are rethinking disposable battery usage at InfoComm 2026 (Las Vegas, June 17-19, booth #N6311). ...
10/06/2026
The Video Services Forum (VSF) has further enhanced the Reliable Internet Streaming Transport (RIST) protocol by incorporating a new feature, RIST Satellite-Hyb...
10/06/2026
Share
Copy link
Facebook
X
Linkedin
Bluesky
Email...
10/06/2026
10 Jun 2026
VEON's Banglalink Brings Every World Cup 2026 Match to Football...
10/06/2026
Wednesday 10 June 2026
How to watch every ICC Women's T20 World Cup 2026 match live on Sky Sports
Where is the ICC Women's T20 World Cup 2026 availabl...
10/06/2026
Wednesday 10 June 2026
P RLA brings first-ever Beautifully Clean Oral Care'...
10/06/2026
Wednesday 10 June 2026
Sky reveals pulse-pounding first teaser trailer for upco...
10/06/2026
Wuppertal June 10, 2026
Riedel Artist at the Heart of the 14th World Live Neurovascular ConferenceAt the 14th World Live Neurovascular Conference (WLNC) in Li...
10/06/2026
Back to All News
Netflix Showcases New Mobile Experience, Collections and Games Vision in APAC
Product
10 June 2026
GlobalSouth KoreaJapan
Link copied to c...
10/06/2026
Back to All News
Ready to Join? The Netflix Fan Club Arrives at Cannes Lions
Business
10 June 2026
Global
Link copied to clipboard
Download all assets
N...
10/06/2026
Chemists snap together complex 3D molecules from highly reactive radicals-without losing their shape Scripps Research team develops stereoretentive radical-radi...
09/06/2026
Kiswe announces an expanded long-term partnership with ONE Championship (ONE), t...
09/06/2026
SiriusXM will broadcast FOX Sports' English-language commentary for all 104 FIFA World Cup 2026 matches from June 11 through July 19, available to subscribe...
09/06/2026
EVS has announced it is changing its corporate name from EVS Broadcast Equipment to EVS, reflecting the company's expanded portfolio beyond broadcast equipm...
09/06/2026
Fox Corporation and the NFL have announced a multi-year agreement to bring NFL c...
Large language models and the applications they power enable unprecedented opportunities for organizations to get deeper insights from their data reservoirs and to build entirely new classes of applications.
