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 ...
09/10/2026
Share
Copy link
Facebook
X
Linkedin
Bluesky
Email...
09/10/2026
COW Jobs: Visual Storyteller - Emeryville, CA
Brie Clayton October 9, 2026
0 Comments
Visual Storyteller
October 7, 2026 COW Jobs: Cinematographer...
09/10/2026
Master 3-Point Lighting in After Effects With an Orbiting Light Rig
Graham Quince October 9, 2026
0 Comments
Ever wondered why your 3D renders in Afte...
09/10/2026
nxtedition will exhibit for the first time at Asias Broadcasting & Infotainment Show (A.B.I.S) in Mumbai, bringing its integrated production platform and latest...
09/10/2026
Today's AI Isn't the Answer
Andy Marken October 8, 2026
0 Comments
Let's say you're telling the truth, and this is all a dream. - Do...
09/10/2026
Ikegami to Introduce New UHK-X700VN Camera at NAB New York, Offering Advanced Cr...
09/10/2026
Share
Copy link
Facebook
X
Linkedin
Bluesky
Email...
09/10/2026
Share
Copy link
Facebook
X
Linkedin
Bluesky
Email...
09/10/2026
Share
Copy link
Facebook
X
Linkedin
Bluesky
Email...
09/10/2026
Witness world firsts, new arrivals and the work that goes on behind the scenes a...
09/10/2026
September 10 2026, 06:00 (PDT) Dolby Expands Dolby OptiView Platform with New Capabilities at IBC 2026
New Sports Intelligence helps providers better unders...
08/10/2026
Modern venues have come a long way since the tinny, echoing audio of the mid 20t...
08/10/2026
ESPN's first Super Bowl production will stream in English and Spanish on Disney+, while the ESPN App carries the full slate of game, ManningCast, studio, an...
08/10/2026
From high-school videoboards to War Eagle Productions, the Opelika native is learning how to lead shows and capture big moments
The future of live sports produ...
08/10/2026
The Television Academy and Prime Video have announced a six-year agreement making Prime Video the exclusive global home of the Emmy Awards beginning with the 20...
08/10/2026
Techex will exhibit at NAB Show New York 2026 (Booth 623, Javits Center, October...
08/10/2026
X Games has announced the appointment of Steve Rogers as Chief Operating Officer. Rogers previously served as COO of World Supercross and spent nearly a decade ...
08/10/2026
Fonn Group has appointed Trevor Spielmann as Sales Director for Germany, Austria, and Switzerland. Spielmann joins from EVS, where he held sales and business de...
08/10/2026
The flagship program will originate from major PG events with a four-camera production, live distribution across PGTV, social platforms
Perfect Game is bringin...
08/10/2026
Innovation rarely happens in a vacuum. More often, it happens at the point of collision: when worlds that have run on separate tracks begin to converge. That is...
08/10/2026
For decades, the production and broadcast services business has run on a familia...
08/10/2026
Uplynk, Zixi, BroadView Software, and Grass Valley have published The IP Converg...
08/10/2026
The NFL has announced the return of the Big Data Bowl, the league's analytics competition, now in its ninth year and again powered by AWS. Registration is o...
08/10/2026
Eutelsat has renewed its partnership with T l diffusion d'Alg rie (TDA), Algeria's national broadcasting network operator, for continued distribution of...
08/10/2026
Disney will livestream Super Bowl LXI on February 14, 2027, alongside the game's broadcast on ESPN and ABC. All Disney subscribers in the United States wi...
08/10/2026
RHStv and Red House Streaming have announced an expansion of their partnership w...
08/10/2026
Gray Media has announced a distribution agreement with TEGNA and DAZN to simulca...
08/10/2026
SVG readers can save 10% on registration and receive a complimentary Solutions H...
08/10/2026
PROTON Camera Innovations GmbH has been named a recipient of a 2026 Technology a...
08/10/2026
Industry leaders discuss real-world cloud workflows for live game production, pl...
08/10/2026
The next-gen audio technology joins Dolby Vision as part of Apple TV's immersive F1 experience...
08/10/2026
David France, Jesse Short Bull, Mato Wayuhi, N. Bird Runningwater, Jodi Archambau, Jhane Myers, Paul McGuire, Amber Morning Star Byars, Holly Macarro, Nick Tils...
08/10/2026
Dynamic panning plug-ins get an upgrade
Sound Particles have just announced major updates for two of their popular dynamic panning plug-ins. Energy Panner 2...
08/10/2026
New expandable instrument boasts full DMNO engine
UDO Audio have just announced the upcoming release of a new, expandable synth that packs the power of the ...
08/10/2026
Two new subwoofers added to monitoring line-up
ADAM Audio's monitoring range has just been expanded with two new DSP-equipped subwoofers, the Sub8 PRO a...
08/10/2026
eds3_5_jq(document).ready(function($) { $(#eds_sliderM519).chameleonSlider_2_1({ content_source:......
08/10/2026
Calrec gives broadcasters freedom of choice at NAB Show New York We're looking forward to meeting up with you all at NAB New York on 21st-22nd October.
Cal...
08/10/2026
Share
Copy link
Facebook
X
Linkedin
Bluesky
Email...
08/10/2026
Share
Copy link
Facebook
X
Linkedin
Bluesky
Email...
08/10/2026
Share
Copy link
Facebook
X
Linkedin
Bluesky
Email...
08/10/2026
Share
Copy link
Facebook
X
Linkedin
Bluesky
Email...
08/10/2026
Share
Copy link
Facebook
X
Linkedin
Bluesky
Email...
08/10/2026
Share
Copy link
Facebook
X
Linkedin
Bluesky
Email...
08/10/2026
Globecast, a leading provider of broadcast, media and entertainment managed services, will showcase its next generation of media orchestration services at NAB S...
08/10/2026
Clear-Com played a critical role in the production communications infrastructure for Rock in Rio 2026, one of the world's largest and most recognized music...
08/10/2026
Sohonet Launches ClearView Rush: A Leap Forward in Studio Review, Built for Crea...
08/10/2026
Share
Copy link
Facebook
X
Linkedin
Bluesky
Email...
08/10/2026
Share
Copy link
Facebook
X
Linkedin
Bluesky
Email...
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.
