Sony Pixel Power calrec Sony

What's the ROI? Getting the Most Out of LLM Inference

09/10/2024

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
LINK: https://blogs.nvidia.com/blog/llm-inference-roi/...
See more stories from nvidia

Most recent headlines

05/01/2027

Worlds first 802.15.4ab-UWB chip verified by Calterah and Rohde & Schwarz to be demoed at CES 2026

Worlds first 802.15.4ab-UWB chip verified by Calterah and Rohde & Schwarz to be ...

17/11/2026

Sinclair's Bill Fanshawe Sets Retirement

Share {alert(Link copied)}).catch(err => {console.error(Failed to copy: , err)}) data-platform=copy_link data-action= class=font-ui-heading py-2 text-sm tex...

17/11/2026

WCVB Boston Names Michael Murgia Director of Engineering

Share {alert(Link copied)}).catch(err => {console.error(Failed to copy: , err)}) data-platform=copy_link data-action= class=font-ui-heading py-2 text-sm tex...

17/11/2026

Two Scripps Research faculty named inaugural recipients of the Schultz Innovative Investigator Award

LA JOLLA, CA-Two Scripps Research faculty, Assistant Professor Christian Diercks...

16/11/2026

ESPN Edge Partners Chart a Future Built on Data, AI and a New Generation of Fans

The ESPN Edge Innovation Conference once again gave key ESPN technology partners as a chance to take to the stage and discuss the future of sports fan experienc...

16/11/2026

IMGs Ashley Shaner: How Micro-Dramas Can Build Sports Media Audiences and Why Content Creators Matter

As rights holders rethink who tells their stories and where, IMG is preparing th...

16/11/2026

Heard' at SVG Live!: Audio Needs To Make More Noise

From the sidelines, sonic savants advocate for greater emphasis on sound as sports and entertainment increasingly merge SVG Live!, a one-day event on Nov. 10 i...

16/11/2026

Yamaha To Acquire Disguise, Expanding Into Live Sports, Venue Production, and Immersive Entertainment

Deal combines Yamaha's audio technology with Disguise's visual-productio...

16/11/2026

New plug-in suite from Berry Effects

Creative effects designed to encourage experimentation Berry Effects is the brainchild of Abe and Ralph, a pair of multi-instrumentalists and audio engineer...

16/11/2026

Anukari 1.1 now available

Physics-based synth gets first major update Anukari have just released a new version of their innovative physics-based software synth, marking the instrumen...

16/11/2026

Hit'n'Mix RipX DAW 9 arrives

Both versions kitted out with new tools & improvements Hit'n'Mix have just launched the latest major version of their innovative DAW software, with ...

16/11/2026

Eventide to launch SplitEQ 2

Dynamic bands, unmasking features & new stereo tools Launched in 2021, Eventide's SplitEQ plug-in introduced an interesting new approach to EQ, with the...

16/11/2026

Rohde & Schwarz Japan demonstrates multi-vendor ASA-ML communication with highperformance oscilloscope

Rohde & Schwarz Japan demonstrates multi-vendor ASA-ML communication with high p...

16/11/2026

CNAO and Hitachi launch advanced rotating proton therapy system in Italy

November 16, 2026 Hitachi system enables proton beams to be delivered from 360 degrees around the patient, expanding treatment capacity and supporting more pre...

16/11/2026

NBC Celebrates Centennial

Share {alert(Link copied)}).catch(err => {console.error(Failed to copy: , err)}) data-platform=copy_link data-action= class=font-ui-heading py-2 text-sm tex...

16/11/2026

Apple TV Adding Podcasts

Share {alert(Link copied)}).catch(err => {console.error(Failed to copy: , err)}) data-platform=copy_link data-action= class=font-ui-heading py-2 text-sm tex...

16/11/2026

Fabric Insights Partners With Fonn Group on Enhancing Media Workflows

Share {alert(Link copied)}).catch(err => {console.error(Failed to copy: , err)}) data-platform=copy_link data-action= class=font-ui-heading py-2 text-sm tex...

16/11/2026

SDVI Names Marcy Lefkovitz as VP of Strategic Accounts

Share {alert(Link copied)}).catch(err => {console.error(Failed to copy: , err)}) data-platform=copy_link data-action= class=font-ui-heading py-2 text-sm tex...

16/11/2026

Monthly Streaming Costs Up, Internet Fees Down for 2026

Share {alert(Link copied)}).catch(err => {console.error(Failed to copy: , err)}) data-platform=copy_link data-action= class=font-ui-heading py-2 text-sm tex...

16/11/2026

Hexagon Introduces Camera-Mounted 3D Capture and Tracking System

Share {alert(Link copied)}).catch(err => {console.error(Failed to copy: , err)}) data-platform=copy_link data-action= class=font-ui-heading py-2 text-sm tex...

16/11/2026

Bridge Technologies Adds Apple ProRes Support to VB440 Probe

Share {alert(Link copied)}).catch(err => {console.error(Failed to copy: , err)}) data-platform=copy_link data-action= class=font-ui-heading py-2 text-sm tex...

16/11/2026

Rebuilding Broadcast on Air: Inside the 2026 SMPTE Media Technology Summit

Share {alert(Link copied)}).catch(err => {console.error(Failed to copy: , err)}) data-platform=copy_link data-action= class=font-ui-heading py-2 text-sm tex...

16/11/2026

Hiltron HMAM Chosen by Major Satcom Service Provider

Hiltron Communications has received a repeat order from one of Europes longest established communications satellite operators. Central to this project is a Hilt...

16/11/2026

Fabric Insights and Fonn Group Partner to Unlock Media Wo...

Fabric Insights and Fonn Group today announce a partnership to connect editorial planning and media operations across their complementary product portfolios. ...

16/11/2026

Density 1.0 brings a complete color grading workflow to Final Cut Pro

Density 1.0 brings a complete color grading workflow to Final Cut Pro Brie Clayton November 16, 2026 0 Comments Filmmaker Pascal Basel's plugin pr...

16/11/2026

Berklee Students to Perform at Boston's CITGO Sign Relighting

Berklee Students to Perform at Boston's CITGO Sign Relighting Kuper Walker, Ollie Mann, and dreamcorp. will perform at the free public celebration in Kenm...

16/11/2026

The Hollywood Reporter Names Berklee Among World's Best Music Schools

The Hollywood Reporter Names Berklee Among World's Best Music Schools The annual list spotlights Berklee's screen scoring programs and new master'...

16/11/2026

Taurus Rooms - A Five-Year Vision Proven In Meeting Rooms

Five years ago, Lightware launched the original Taurus UCX with a clear vision: USB-C would become the central connection in meeting rooms. The goal was simple:...

16/11/2026

Sonniss: bundles over 80% below their libraries' prices, plus up to 60% off over 2,700 sound effects libraries

Sonniss: bundles over 80% below their libraries' prices, plus up to 60% off ...

16/11/2026

Singaporean Comedy Luck My Life Shot with Blackmagic URSA Cine 17K 65

Singaporean Comedy Luck My Life Shot with Blackmagic URSA Cine 17K 65 Brie Clayton November 16, 2026 0 Comments Blackmagic PYXIS 12K and DaVinci Resol...

16/11/2026

NVIDIA DGX Spark Developer Kit Turns a Desk Into a Supercomputing Lab

At SC26, an annual supercomputing conference taking place this week in Chicago, NVIDIA is introducing the DGX Spark HPC Developer Kit for researchers, academics...

16/11/2026

U sees new original drama Fleet Street from the creators of Miss Scarlet

Images available HERE UKTV has announced Fleet Street (6x60'), a new period crime drama from Miss Scarlet creators Rachael New and Ben Edwards, coming to f...

16/11/2026

Tribeca Festival Lisboa Unveils 2026 Film Lineup And Talent Roster For Third Annual Edition, December 9-13

November 16th, 2026 TRIBECA FESTIVAL LISBOA UNVEILS 2026 FILM LINEUP AND TALE...

16/11/2026

Sky and NOW bring some of the UKs biggest video podcasts to the big screen at no extra cost with Platform Media as first partner

From Graham Norton to Olivia Attwood and Joe Marler, a new collection of hit vid...

16/11/2026

Can Digital Product Passports help meet climate goals?

As countries gather to discuss climate targets at COP 31 this week, one question that is at the heart of debates is how to measure the environmental impact of t...

16/11/2026

NVIDIA Announces Rubin HPC - A New GPU to Accelerate Agents for Science, Engineering and Analytics

Announced today at SC26, NVIDIA Rubin HPC is a new GPU in the NVIDIA Vera Rubin ...

16/11/2026

CUDA at 20: How the Platform Transformed Scientific Computing

This week at SC26, NVIDIA is marking two decades of CUDA by honoring the researchers, developers and engineers who helped make GPU computing a foundation for mo...

16/11/2026

NVIDIA CUDA-X Libraries Give Agents the Tools to Accelerate Scientific Discovery

With expanded NVIDIA CUDA-X libraries, including cuPhoton, cuPotential and cuSpa, researchers can sift through vast amounts of data to accelerate research, from...

16/11/2026

NVIDIA Sweeps MLPerf Training and Endpoints as Vera Rubin Debuts With Leading Performance

In its MLPerf benchmarks debut, with results released today, the NVIDIA Vera Rub...

16/11/2026

NVIDIA Expands Slurm to Bring Quantum Computing Into Supercomputing

Quantum processor deployments at supercomputing centers have grown nearly sevenfold over the past five years, with nearly 30 supercomputing centers around the w...

16/11/2026

Lambda Boosts AI Factory Resiliency and Efficiency With NVIDIA Photonics

To train an AI model, thousands of GPUs have to work together, requiring a network that's fast, efficient and resilient. Lambda is tackling the network pie...

15/11/2026

Yamaha launch MGX, URX & CC1 updates

Mixers, interfaces & DAW controller receive new firmware Yamaha have just launched a trio of firmware updates that bring some powerful new functionality to ...

15/11/2026

Can VideoProc AI Really Reduce Noise and Enhance Blurry, Old Footage? Real-World Tests

Can VideoProc AI Really Reduce Noise and Enhance Blurry, Old Footage? Real-World...

15/11/2026

AI Day Seoul Brings Together Korea's Industry Leaders, Startups and Developer Communities

Over two days at COEX, more than 2,000 attendees converged for AI Day Seoul, bri...

14/11/2026

MLB Networks David Maybrown, 57, Passes Away

Longtime MLB Network employee David Maybrown passed away Thursday at the age of 57. David joined MLB Network in 2009 shortly after its launch, beginning as an i...

14/11/2026

Rotary Exciter & Resonator Clamps from ADDAC System

Modular innovators introduce two guitar-based products ADDAC System's recently introduced Four String Series makes it possible to integrate physical str...

14/11/2026

Marcelo Dai's Live Session Captured with URSA Cine Immersive

Marcelo Dai's Live Session Captured with URSA Cine Immersive Brie Clayton November 14, 2026 0 Comments Artist takes viewers on an immersive journe...

14/11/2026

Political Ads Spiked a Week Before Midterm Elections, Nielsen Says

Share {alert(Link copied)}).catch(err => {console.error(Failed to copy: , err)}) data-platform=copy_link data-action= class=font-ui-heading py-2 text-sm tex...

14/11/2026

FCC Sends 2027 Upper C-Band Auction Plan Into Motion

Share {alert(Link copied)}).catch(err => {console.error(Failed to copy: , err)}) data-platform=copy_link data-action= class=font-ui-heading py-2 text-sm tex...

14/11/2026

ZEISS Showcases Latest Cine Lenses at TorunCAMERIMAGE 202...

Newest Lenses, Horizon Anamorphic Seminar & Guided Lens Character Experience Highlight Creative Possibilities At the 2026 Toru CAMERIMAGE Film Festival, ZEISS...