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 ...

18/11/2026

William Kamau: Young Journalist of the Year finalist 2026

William Kamau is an investigative journalist from Kenya. He says witnessing the pain, displacement and killings' that followed the country's 2007 elect...

18/11/2026

Humberto Basilio: Young Journalist of the Year finalist 2026

Humberto Basilio is a freelance science, health and environmental journalist from Mexico. He says that working in one of the most dangerous countries in the wo...

18/11/2026

Aaron Atsahyel: Young Journalist of the Year finalist 2026

Aaron Atsahyel is a multimedia investigative journalist from Nigeria who draws on his own experiences as someone displaced by the Boko Haram insurgency to tell...

18/11/2026

SVG TranSPORT 2026: All Sessions Now Available To Watch on SVG PLAY

Held in New York City, the event explored the latest technologies and trends shaping the future of live-sports transmission and distribution...

18/11/2026

Samsung Takes the Gold at Your Olympic Games Dakar; Galaxy Phones Play Role in Coverage

The Youth Olympic Games Dakar 2026 opened in Senegal on October 31, giving young...

18/11/2026

Eventide 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...

18/11/2026

IK Multimedia announce Tonex Parametric

Parametric Tone Models emulate control responses & interactions IK Multimedia have just announced the launch of Parametric Tone Models, a new technology tha...

18/11/2026

Shure launch the MV360 Spatial Audio Microphone

Captures immersive audio on phone, tablet or computer via USB-C Shure have just announced the launch of the MV360 Spatial Audio Microphone, a compact USB-C ...

18/11/2026

SBS announces key appointments across programming and commissioning

SBS announces key appointments across programming and commissioning 18 November, 2026 Media releases SBS has announced a series of appointments and role up...

18/11/2026

The National Film and Video Foundation (NFVF) Call for Festival and Market Hosting Funding Applications is Open

The National Film and Video Foundation (NFVF), an agency of the Department of Sp...

18/11/2026

DIRECTV teams with Gracenote to enhance the sports programming experiences that fans want

Strategic renewal builds on existing engagement, supports DIRECTV's objectiv...

18/11/2026

SMPTE Publishes Engineering Report on Content Authenticity

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...

18/11/2026

Senate Commerce Committee Approves Thumann Severs for FCC Commissioner

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...

18/11/2026

wTVision Qualifies NEOiD PTZ Cameras for Virtual Sets, AR

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...

18/11/2026

Spanish Media Group Deploys Lawo IP Audio Infrastructure

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...

18/11/2026

Library of American Broadcasting Foundation Honors 2026 Giants of Broadcasting & Electronic Arts

Share {alert(Link copied)}).catch(err => {console.error(Failed to copy: , er...

18/11/2026

Monthly Subscription Streaming Service Prices Up Just 8% Since 2022

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...

18/11/2026

Spectrum News Streamlines Workflow With Amagi

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...

18/11/2026

Media Links Joins Panasonic Kairos Partner Program

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...

18/11/2026

Signiant Acquires Burnlab's Ceivo AI-Driven Media Platform

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...

18/11/2026

Democrats Urge Delay of Thumann Severs FCC Confirmation

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...

18/11/2026

Hisong Unveils AirStudio Muse Audio Capture 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...

18/11/2026

Spectrum Reach Taps Debbie Presser to Lead Sales at AdLink New York

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...

18/11/2026

Nationwide EAS Test Runs Without Apparent Hiccups

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...

18/11/2026

SMPTE Identifies Key Requirements for Content Provenance...

SMPTE, the home for media professionals, technologists and engineers, today announced the publication of an engineering report with the summarized results of an...

18/11/2026

Live from London, its Saturday night! Saturday Night Live UK returns 16 January on Sky and NOW

Wednesday 18 November 2026 Live from London, it's Saturday night! Saturday ...

18/11/2026

Sky Kids Investigates: Being a Boy

BAFTA award-winning series sees 16-year-old reporter explore topics including toxic digital influencers and disrespectful attitudes to girls in a groundbreaking...

18/11/2026

Choosing a display for TMxCore

TMxCore has no built-in display. If you use TMxCore with a network display, you may not need a dedicated display at all. If you prefer a directly connected scre...

18/11/2026

TeleData Accelerates XGS-PON Service Rollout in Germany and Unifies its Network with Harmonic's cOS Broadband Platform

Key Takeaways TeleData is building a unified DOCSIS and fiber network with Ha...

18/11/2026

Beeble launches SwitchX 3.0 with dynamic environments and greater creative control

Beeble recently launched SwitchX 3.0, a major update to its AI-powered video-to-...

18/11/2026

Film London to lead new industry PhD program in partnership with Brunel University

Film London will lead and manage a long-term industry research program in partne...

18/11/2026

Hexagon launches MANTIS to capture 3D scene data as cameras roll

Hexagon, the global leader in measurement technologies, recently launched MANTIS by Leica Geosystems, a camera-mounted 3D capture and tracking system to give fi...

18/11/2026

Successfully achieved BSI C5 attestation

Arvato Systems Expands BSI-C5 Testing Greater Confidence in Cloud Security and Regulatory Resilience Successful BSI-C5 Retest Certification in accordance wi...

18/11/2026

RT to premiere new three-part documentary series Able

Brendan O'Connor joins six disabled job seekers as they navigate the barriers of finding work and challenge perceptions of disability in the workplace Begi...

18/11/2026

When writing becomes a weapon : New true-crime podcast series to reveal the impact of anonymous poison-pen letter case

Poison Ink begins Friday 20 November 2026 on RT Listen and RT Radio 1 Listen ...

18/11/2026

RT Choice Music Prize 2027 Key Dates Announcement

RT Choice Music Prize In association with IMRO and IRMA 2027 KEY DATES Irish Album of the Year 2026 Shortlist 18th January Irish Artist of the Year 2026 ...

18/11/2026

NVIDIA RTX Spark Compact Desktops Out Now

New NVIDIA RTX Spark compact desktops from ASUS, Dell and MSI are available today, bringing local AI agents, image and video generation, and model fine-tuning t...

17/11/2026

Fabric Insights and Fonn Group Partner to Connect Newsroom Planning and Media Operations

Fabric Insights and Fonn Group have announced a partnership to connect editorial...

17/11/2026

NABJ Online Seminar Highlights Black Sports Production Legends on Nov. 25

The National Association of Black Journalists is offering up a Thanksgiving Eve treat on Wednesday, Nov. 25, in the form of a live streaming event featuring for...

17/11/2026

Bridge Technologies Adds Apple ProRes Support to VB440 Production Probe

Bridge Technologies has announced that its VB440 production probe now supports both monitoring and signal generation for Apple ProRes video streams. The additio...

17/11/2026

Canon U.S.A. Hosts Jericho High School Journalism Students at Melville Headquarters

Canon U.S.A., Inc. recently welcomed 21 Jericho High School students to its Melv...

17/11/2026

Amagi to Power Live Newscasts for Spectrum News

Spectrum News has deployed the Amagi platform to support its live news operations. Amagi says the platform lets Spectrum News streamline newsroom workflows, red...

17/11/2026

Microsoft Expands Haivision Video Contribution Deployment, Adds Makito ONE for Global Live Events

Microsoft has expanded its use of Haivision's live video contribution ecosys...

17/11/2026

Setanta Sports Deploys Ateme TITAN Live for OTT Sports Streaming

Setanta Sports has deployed Ateme's TITAN Live platform to support its OTT video delivery operations. Ateme's transcoding technology is integrated into ...

17/11/2026

Media Links Joins Panasonic's KAIROS Alliance Partner Program

Media Links has joined the KAIROS Alliance Partner program for Panasonic's KAIROS IT/IP live video production platform. The company completed interoperabili...

17/11/2026

Behind The Mic: Dan Hicks, Iconic NBC Sports Broadcaster, Passes Away at 64

Behind The Mic provides a roundup of recent news regarding on-air talent, including new deals, departures, and assignments compiled from press releases and repo...

17/11/2026

2026 Sports Broadcasting Hall of Fame: Jim Dove, an Innovator Offering Second Looks at the Greatest Plays in Sports

In live sports and in life, things go by in the blink of an eye. Savoring the mo...

17/11/2026

SVG Sit-Down: Eutelsat's Gonzalez, Carroccia on C-Band Clearing, LEO Contribution, the Future of Satellite in Sports

Satellite operators are adding a new generation of low-Earth-orbit (LEO) network...

17/11/2026

Clear-Com Delivers Portable Wireless Systems for Aircraft Communications Security Testing

Clear-Com has delivered five customized test systems for electromagnetic securit...