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 ...
07/10/2026
Dalet, a leading technology and service provider for media-rich organizations, today announced the latest Long-Term Supported (LTS) release of Dalet Flex. Build...
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...
23/07/2026
Share
Copy link
Facebook
X
Linkedin
Bluesky
Email...
23/07/2026
Share
Copy link
Facebook
X
Linkedin
Bluesky
Email...
23/07/2026
David Dvir, 2D founder (left), Sundeep Reddy (center), and Jean-Marc Bouchut (right), both of Zeiss...
23/07/2026
Screen Australia announces new national documentary accelerator program 23 July 2026
Screen Australia has today announced the Above the Line Documentary Accel...
22/07/2026
Among other tweaks to coverage, the broadcaster added announce positions in the ...
22/07/2026
The Australian Sports Technologies Network (ASTN) has reported a sharp acceleration in Australia's sportstech sector, with revenue growth rising from 10% in...
22/07/2026
From a retrofitted broadcast bus to packed Match Day Live! events, Bennett detai...
22/07/2026
TMRW Sports has named Brendan Donohue as President, Flag Football, where he will lead the company's professional flag football business in partnership with ...
22/07/2026
Stephen Arnold Music (SAM) designed and produced soundtracks for nine projection...
22/07/2026
Harmonic has announced a $25,000 donation to Global Empowerment Mission (GEM) fo...
22/07/2026
Telemundo has announced it has acquired from UC3 exclusive U.S. Spanish-language...
22/07/2026
The Rest Is Football, hosted by Gary Lineker, Alan Shearer, and Micah Richards, ...
22/07/2026
Sports Studio has announced a partnership with TheSoul Group to distribute Free ...
22/07/2026
Daktronics has concluded its 2026 High School Video Summit, a two-day hybrid event combining virtual sessions with hands-on host school visits across the countr...
22/07/2026
FOR-A America will exhibit at the MMCA Show in Pensacola, Florida (July 21-24), marking the company's first appearance at the house of worship AV event. FOR...
22/07/2026
The New York Yankees have selected Lumen Technologies (NYSE: LUMN) to connect Ya...
22/07/2026
Calrec will exhibit at IBC 2026 (Stand 8.C47) with demonstrations of its IP-native audio ecosystem built around the Argo console range, and will announce two ne...
22/07/2026
Solid State Logic has announced the appointment of Algam EKO as its distributor for SSL large format consoles in Italy, covering the Live, Broadcast, and Studio...
22/07/2026
FloSports has announced the launch of the FloSports Channel, a free ad-supported streaming channel premiering on Prime Video in the U.S. on July 21. The channel...
22/07/2026
Banana Ball TV (BTV), the broadcast operation of the Savannah Bananas, has insta...
22/07/2026
SiriusXM and Audacy have announced an agreement bringing 42 Audacy sports, news, and talk stations from 29 U.S. markets to SiriusXM. Twenty-three Audacy sports ...
22/07/2026
The collaboration produced more than 500 short-form videos across four cities th...
22/07/2026
At 2,600 sq. ft., the new home of digital content houses a main studio, podcast space, and an auxiliary control room
NFL franchises are accelerating their prod...
22/07/2026
Library now available in Engine Player
Another of Eduardo Tarilonte's popular sample libraries has just been moved over to Engine Audio's Engine Pla...
22/07/2026
All-in-one mic & DSP interface overhauled
Cloudvocal have just refreshed the design of their all-in-one mic, audio interface and DSP processor, introducing ...
22/07/2026
The final boss of the lost dual-op-amp circuit
Electro-Harmonix have just introduced the Deluxe Big Muff Pi 2, which they say offers their latest take on t...
22/07/2026
eds3_5_jq(document).ready(function($) { $(#eds_sliderM519).chameleonSlider_2_1({ content_source:......
22/07/2026
Share
Copy link
Facebook
X
Linkedin
Bluesky
Email...
22/07/2026
Share
Copy link
Facebook
X
Linkedin
Bluesky
Email...
22/07/2026
Proven managed IP solution helps broadcasters confidently navigate the future of broadcast distribution
LTN announces it has completed successful satellite-to-...
22/07/2026
Today, Chaos announced V-Ray 7, update 4, eliminating the need for asset conversions and tool switching, which slow down media and entertainment pipelines. Arti...
22/07/2026
An AI playground for the Project Indigo camera app
Marc Levoy July 22, 2026
0 Comments
At left is the user interface of our Playground, running on an ...
22/07/2026
Custom Consoles announces the completion of a large technical furniture project for Radio T l vision Suisse (RTS) at the networks recently inaugurated productio...
22/07/2026
Share
Copy link
Facebook
X
Linkedin
Bluesky
Email...
22/07/2026
Share
Copy link
Facebook
X
Linkedin
Bluesky
Email...
22/07/2026
Share
Copy link
Facebook
X
Linkedin
Bluesky
Email...
22/07/2026
Culver City, CA, July 21, 2026 3D streaming infrastructure provider Miris today announced glTF intake, including the binary .glb form, to its spatial streamin...
22/07/2026
Documentary Lo Invisible Captured with Blackmagic PYXIS 6K
Brie Clayton July 21, 2026
0 Comments
Camera's versatile design allowed filmmakers to c...
22/07/2026
AAF Bridge converts AAF and XML Timelines to Ableton Live - and back
Jerome Gardier July 21, 2026
0 Comments
A two-way converter for macOS and Windows...
22/07/2026
From Class to Concert: Explore a Student-Led Production of David Bowies Work Thr...
22/07/2026
NVIDIA founder and CEO Jensen Huang today visited the Naval Postgraduate School in Monterey, California, to commission an NVIDIA DGX GB300 system - bringing one...
22/07/2026
A poisoned spy, a secret war, and a small English city contaminated by a deadly chemical weapon. Available on Sky Documentaries from 4 AugustWednesday 22 July 2...
22/07/2026
Audio at Parity: Comscore Launches Transcript-Level Targeting and Measurement wi...
22/07/2026
Arvato Systems and Lobster Enter into a Partnership and Combine the Integration Platform and Managed Services
Flexible integration solutions for cloud, hybrid...
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.
