Hewlett Packard Enterprise Debuts End-to-End AI-Native Portfolio for Generative AI
Enterprise-class co-designed software and hardware solutions from HPE and NVIDIA accelerate development and deployment of GenAI applications
- Availability of two HPE and NVIDIA co-engineered full-stack GenAI solutions.
-
A preview of
HPE Machine Learning Inference Software . - An enterprise retrieval-augmented generation (RAG) reference architecture.
- Support to develop future products based on the new NVIDIA Blackwell platform.
“To deliver on the promise of GenAI and effectively address the full AI lifecycle, solutions must be hybrid by design,” said
“Generative AI can turn data from connected devices, data centers and clouds into insights that can drive breakthroughs across industries," said
Supercomputing-powered GenAI training and tuning
Announced at SC23, HPE’s supercomputing solution for generative AI is now available to order for organizations seeking a preconfigured and pretested full-stack solution for the development and training of large AI models. Purpose-built to help customers accelerate GenAI and deep learning projects, the turnkey solution is powered by NVIDIA and can support up to 168 NVIDIA GH200 Grace Hopper Superchips.
The solution enables large enterprises, research institutions, and government entities to streamline the model development process with an AI/ML software stack that helps customers accelerate GenAI and deep learning projects, including LLMs, recommender systems, and vector databases. Delivered with services for installation and set-up, this turnkey solution is designed for use in AI research centers and large enterprises to realize improved time-to-value and speed up training by 2-3X. For more information or to order it today, visit HPE's supercomputing solution for generative AI.
Enterprise-class GenAI tuning and inference
Previewed at Discover
Featuring a high-performance AI compute cluster and software from HPE and NVIDIA, the solution is ideal for lightweight fine-tuning of models, RAG, and scale-out inference. The fine-tuning time for a 70 billion parameter Llama 2 model running this solution decreases linearly with node count, taking six minutes on a 16-node system1. The speed and performance enable customers to realize faster time-to-value by improving business productivity with AI applications like virtual assistants, intelligent chatbots, and enterprise search.
Powered by HPE ProLiant DL380a Gen11 servers, the solution is pre-configured with NVIDIA GPUs, the NVIDIA Spectrum-X Ethernet networking platform, and NVIDIA BlueField-3 DPUs. The solution is enhanced by HPE’s machine learning platform and analytics software, NVIDIA AI Enterprise 5.0 software with new NVIDIA NIM microservice for optimized inference of generative AI models, as well as NVIDIA NeMo Retriever and other data science and AI libraries.
To address the AI skills gap, HPE Services experts will help enterprises design, deploy, and manage the solution, which includes applying appropriate model tuning techniques. For more information or to order it today, visit HPE’s enterprise computing solution for generative AI.
From prototype to productivity
HPE and NVIDIA are collaborating on software solutions that will help enterprises take the next step by turning AI and ML proofs-of-concept into production applications. Available to HPE customers as a technology preview,
To assist enterprises that need to rapidly build and deploy GenAI applications that feature private data, HPE developed a reference architecture for enterprise RAG, available today, that is based on the NVIDIA NeMo Retriever microservice architecture. The offering consists of a comprehensive data foundation from
To aid in data preparation, AI training, and inferencing, the solution merges the full spectrum of open-source tools and solutions from
Next-gen solutions built on NVIDIA Blackwell platform
HPE will develop future products based on the newly announced NVIDIA Blackwell platform, which incorporates a second-generation Transformer Engine to accelerate GenAI workloads. Additional details and availability for forthcoming HPE products featuring the NVIDIA GB200 Grace Blackwell Superchip, the HGX B200, and the HGXB100 will be announced in the future.
About
1 Based on initial internal benchmarks of llama-recipes finetuning.py that tracked the average epoch time to fine-tune eight nodes at 594 seconds and 16 nodes at 369 seconds with flash attention and parameter efficient fine-tuning.
View source version on businesswire.com: https://www.businesswire.com/news/home/20240318761801/en/
Media:
cristina.thai@hpe.com
Source: