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
11/12/2025
Dalet, a leading provider of cloud-native, end-to-end media workflow solutions, ...
28/11/2025
Nadia Fall attends the 2025 Sundance Film Festival premiere of Brides at the Egyptian Theatre on January 24, 2025, in Park City, Utah. (Photo by Donyale West/...
28/11/2025
It's easy to ignore those little red update available badges. But when it ...
28/11/2025
WASHINGTON Federal Communications Commission has released a tentative agenda for the December Open Commission Meeting scheduled for Thursday, December 18, 2025 ...
28/11/2025
The Professional Fighters League is looking to super-serve fans of mixed martial...
28/11/2025
Fubo has released in beta on select Roku devices a new feature that lets users display up to four simultaneous streams at once....
28/11/2025
The WNBA playoffs and Week 4 of the NFL regular season highlight the list of live sports events airing on television this weekend....
28/11/2025
The 32nd class of honorees to the B+C Hall of Fame took to the stage at New York's Ziegfeld Ballroom on September 26 for a gala induction event. Click below...
28/11/2025
We hold in our hands the very last Next Text for Next TV, the weekly back-and-fo...
28/11/2025
DirecTV said it made a deal with EchoStar to buy EchoStar's video businesses, including satellite-TV provider Dish TV and virtual MVPD Sling TV, for $1 plus...
28/11/2025
The Broadcasting+Cable Hall of Fame, the premier industry event paying tribute to the influencers, innovators and shining lights of broadcast, cable and streami...
28/11/2025
Friday 28 November 2025
Sky Sports x Slawn drop limited-edition football jersey...
28/11/2025
Rohde & Schwarz shows resilience in a challenging environment, revenue exceeds t...
28/11/2025
Unwrapped: The Toy Show Appeal - airing this Sunday on RT One and RT Player- s...
27/11/2025
LONDON Vizrt has added several AI-driven advanced features offering improved speed, intelligence and accuracy in the newest version of its media asset managemen...
27/11/2025
Prime Video has launched AI-powered video season recaps in a beta version for select English-language Prime Original series in the U.S., a move Amazon is callin...
27/11/2025
Back to All News
Netflix's Raat Akeli Hai: The Bansal Murders Marks a Grand...
27/11/2025
27 Nov 2025
GSMA brings M360 Eurasia 2026 to Samarkand in partnership with VEON...
27/11/2025
Tahar Rahim and Izuka Hoyle star in the gripping six-part Sky Original from Acad...
27/11/2025
Thursday 27 November 2025
Sky Arts Reveals the Nation's Greatest Basslines - and Queen Reign Supreme
The UK's most iconic basslines have been revealed...
27/11/2025
Back to All News
Stranger Things 5': Prepare for One Last Adventure With O...
27/11/2025
The media industry has a paradox at its core. It's an industry built on light, color and imagination, yet behind the scenes, it's powered by one of the ...
27/11/2025
Rating reflects rating progress across areas including policies, diversity & inclusion, health & safety and Net Zero leadership
Winchester, UK, 27 November 202...
27/11/2025
What are the industry standards for Retail Media? Kathryn explains that certification is based on the IAB Europe Retail Media Measurement Standards and the IAB ...
27/11/2025
World champion boxer and Irish sporting icon Katie Taylor will be in studio this...
27/11/2025
Roblox, one of the world's most popular online gaming platforms for primary ...
27/11/2025
Black Friday is leveling up. Get ready to score one of the biggest deals of the season - 50% off the first three months of a new GeForce NOW Ultimate membership...
26/11/2025
SVG Sit-Down: Prime Video EP Mike Muriano Previews Massive Black Friday Slate Fe...
26/11/2025
A cinematic snow sculpture at the 1995 Sundance Film Festival. Photo by Randall Michelson...
26/11/2025
Book podcasts are booming. On Spotify, you'll find everything from celebrity book clubs to deep dives with bestselling authors. And in markets where audiobo...
26/11/2025
Mumbai, November 24, 2025: In a first-of-its-kind initiative, JioStar, in collab...
26/11/2025
LONDON Factual content producer ITN Productions has launched a new low-latency IP gallery for news bulletins....
26/11/2025
MIAMI TelevisaUnivision said it struck a new multiyear distribution agreement with YouTube TV that includes distribution of TelevisaUnivision's U.S. network...
26/11/2025
OpenDrives, Inc., a leader in software-defined data storage and data services, today announced the launch of the Atlas Corporate Creative Solution. This new Atl...
26/11/2025
Disguise, the industry-leading company powering the world's biggest live performances, is partnering with pioneering LED wall manufacturer DVS to give atten...
26/11/2025
HighField AI, the pioneer in agentic and multimodal automation for broadcast and media production, today announced the expansion of its global channel partner n...
26/11/2025
As high-stakes Premier League fixtures approach and additional premium content launches, with MONO positioning themselves to dominate Thailand's sports stre...
26/11/2025
Hosting a wide variety of events from high-intensity NHL games to complex live music concerts and major entertainment productions, Montreal's 21,000 capacit...
26/11/2025
Vizrt, the leader in live production technology revolutionizing viewer engagement and experience, releases AI-driven advances focusing on speed, intelligence, a...
26/11/2025
ITN Productions, an award-winning factual content producer, today launched a new low-latency IP gallery for news bulletins. Responsible for delivering a leading...
26/11/2025
Ikegami reports ongoing advances throughout 2025 in developing and delivering coordinated television production solutions that maximize quality, versatility and...
26/11/2025
Following the Nov. 21 blackout of NBCUniversal channels on Fubo, the two sides have traded barbs about their inability to reach a new carriage deal....
26/11/2025
LONDON As TV sports rights become increasingly important for both broadcasters and streamers, Ampere Analysis predicts global investment in the genre will surpa...
26/11/2025
LOS ANGELES Vubiquity said it has achieved the Amazon Web Services (AWS) Media & Entertainment Competency as part of the AWS Partner Network (APN). This designa...
26/11/2025
WASHINGTON The Federal Communications Commission's Enforcement Bureau said it has entered into a consent decree with Comcast calling for the cable company t...
26/11/2025
Berklee Named to the Hollywood Reporters Top Music Schools List The publication highlights the college's screen scoring program, industry partnerships, and ...
26/11/2025
Back to All News
Animated Series Love Through a Prism' Casts New Light on ...
26/11/2025
Back to All News
NALIP Unveils Fifth Cohort of Director Incubator
Social Impact
26 November 2025
United States
Link copied to clipboard
The National Assoc...
26/11/2025
YouView Achieves Greenly Gold Certification for SustainabilityNov 26, 2025
YouView is proud to announce a Gold Certification award from Greenly for our perform...
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.
