GTC—NVIDIA today launched four inference platforms optimized for a diverse set of rapidly emerging generative AI applications — helping developers quickly build specialized, AI-powered applications that can deliver new services and insights.
The platforms combine NVIDIA’s full stack of inference software with the latest NVIDIA Ada, NVIDIA Hopper™ and NVIDIA Grace Hopper™ processors — including the NVIDIA L4 Tensor Core GPU and the NVIDIA H100 NVL GPU, both launched today. Each platform is optimized for in-demand workloads, including AI video, image generation, large language model deployment and recommender inference.
“The rise of generative AI is requiring more powerful inference computing platforms,” said Jensen Huang, founder and CEO of NVIDIA. “The number of applications for generative AI is infinite, limited only by human imagination. Arming developers with the most powerful and flexible inference computing platform will accelerate the creation of new services that will improve our lives in ways not yet imaginable.”
Accelerating Generative AI’s Diverse Set of Inference Workloads
Each of the platforms contains an NVIDIA GPU optimized for specific generative AI inference workloads as well as specialized software:
- NVIDIA L4 for AI Video can deliver 120x more AI-powered video performance than CPUs, combined with 99% better energy efficiency. Serving as a universal GPU for virtually any workload, it offers enhanced video decoding and transcoding capabilities, video streaming, augmented reality, generative AI video and more.
- NVIDIA L40 for Image Generation is optimized for graphics and AI-enabled 2D, video and 3D image generation. The L40 platform serves as the engine of NVIDIA Omniverse™, a platform for building and operating metaverse applications in the data center, delivering 7x the inference performance for Stable Diffusion and 12x Omniverse performance over the previous generation.
- NVIDIA H100 NVL for Large Language Model Deployment is ideal for deploying massive LLMs like ChatGPT at scale. The new H100 NVL with 94GB of memory with Transformer Engine acceleration delivers up to 12x faster inference performance at GPT-3 compared to the prior generation A100 at data center scale.
- NVIDIA Grace Hopper for Recommendation Models is ideal for graph recommendation models, vector databases and graph neural networks. With the 900 GB/s NVLink®-C2C connection between CPU and GPU, Grace Hopper can deliver 7x faster data transfers and queries compared to PCIe Gen 5.
The platforms’ software layer features the NVIDIA AI Enterprise software suite, which includes NVIDIA TensorRT™, a software development kit for high-performance deep learning inference, and NVIDIA Triton Inference Server™, an open-source inference-serving software that helps standardize model deployment.
Early Adoption and Support
Google Cloud is a key cloud partner and an early customer of NVIDIA’s inference platforms. It is integrating the L4 platform into its machine learning platform, Vertex AI, and is the first cloud service provider to offer L4 instances, with private preview of its G2 virtual machines launching today.
Two of the first organizations to have early access to L4 on Google Cloud include: Descript, which uses generative AI to help creators produce videos and podcasts, and WOMBO, which offers an AI-powered text to digital art app called Dream.
Another early adopter, Kuaishou provides a content community and social platform that leverages GPUs to decode incoming live streaming video, capture key frames, optimize audio and video. It then uses a transformer-based large-scale model to understand multimodal content and improve click-through rates for hundreds of millions of users globally.
“Kuaishou recommendation system serves a community having over 360 million daily users who contribute millions of UGC videos every day,” said Yue Yu, senior vice president at Kuaishou. “Compared to CPUs under the same total cost of ownership, NVIDIA GPUs have been increasing the system end-to-end throughputs by 11x and reducing latency by 20%.”
D-ID, a leading generative AI technology platform, elevates video content for professionals by using NVIDIA L40 GPUs to generate photorealistic digital humans from text — giving a face to any content while reducing the cost and hassle of video production at scale.
“L40 performance was simply amazing. With it, we were able to double our inference speed,” said Or Gorodissky, vice president of research and development at D-ID. “D-ID is excited to use this new hardware as part of our offering that enables real-time streaming of AI humans at unprecedented performance and resolution while simultaneously reducing our compute costs.”
Seyhan Lee, a leading AI production studio, uses generative AI to develop immersive experiences and captivating creative content for the film, broadcast and entertainment industries.
“The L40 GPU delivers an incredible boost in performance for our generative AI applications,” said Pinar Demirdag, co-founder of Seyhan Lee. “With the inferencing capability and memory size of the L40, we can deploy state-of-the-art models and deliver innovative services to our customers with incredible speed and accuracy.”
Cohere, a leading pioneer in language AI, runs a platform that empowers developers to build natural language models while keeping data private and secure.
“NVIDIA’s new high-performance H100 inference platform can enable us to provide better and more efficient services to our customers with our state-of-the-art generative models, powering a variety of NLP applications such as conversational AI, multilingual enterprise search and information extraction,” said Aidan Gomez, CEO at Cohere.
The NVIDIA L4 GPU is available in private preview on Google Cloud Platform and also available from a global network of more than 30 computer makers, including Advantech, ASUS, Atos, Cisco, Dell Technologies, Fujitsu, GIGABYTE, Hewlett Packard Enterprise, Lenovo, QCT and Supermicro.
The NVIDIA L40 GPU is currently available from leading system builders, including ASUS, Dell Technologies, GIGABYTE, Hewlett Packard Enterprise, Lenovo and Supermicro with the number of partner platforms set to expand throughout the year.
The Grace Hopper Superchip is sampling now, with full production expected in the second half of the year. The H100 NVL GPU also is expected in the second half of the year.
NVIDIA AI Enterprise is now available on major cloud marketplaces and from dozens of system providers and partners. With NVIDIA AI Enterprise, customers receive NVIDIA Enterprise Support, regular security reviews and API stability for NVIDIA Triton Inference Server, TensorRT and more than 50 pretrained models and frameworks.
Hands-on labs for trying the NVIDIA inference platform for generative AI are available immediately at no cost on NVIDIA LaunchPad. Sample labs include training and deploying a support chatbot, deploying an end-to-end AI workload, tuning and deploying a language model on H100 and deploying a fraud detection model with NVIDIA Triton™.