Your boss tells you to build some generative AI. Dell and Nvidia are already knocking

Who wouldn't want to feed their enterprise into Project Helix for training and inference? Anyone?

Dell World Dell has hooked up with Nvidia to pitch enterprises on tools to build generative AI models trained on their own corporate data, rather than publicly available information such as that used by general-purpose large-language models (LLMs) like OpenAI's GPT.

The key to the pitch is data security. Nvidia's Manuvir Das, vice president of enterprise computing, told journalists that an enterprise building its own generative AI trained on its own domain-specific data "doesn't have to worry about their proprietary data being mixed with proprietary data of some other company during the training."

Project Helix, a scheme launched by Nvidia and Dell on Tuesday at Dell Technologies World 2023, includes the PowerEdge XE9680 and R760xa rack servers that are optimized for AI training and inferencing workloads. The XE9680, while it runs two of Intel's 4th Generation Xeon Scalable Processors, also holds eight of Nvidia's latest H100 Tensor Core GPUs connected via Nvidia's NVLink networking.

Nvidia also plans to exploit its AI enterprise software, frameworks, and developer tools – including NeMo and pretrained foundation models NeMo Guardrails – to build secure generative AI chatbots. Dell's PowerScale and ECS Enterprise Object Storage systems for unstructured data can be used with the PowerEdge rack servers, it said.

"All of this allows us to put together really a complete solution for generative AI that can be run on-prem, that is fully validated with the hardware and software, that is secure [and] private," according to Das.

Livin' on the edge

Running the training and inference workloads within a company's own datacenter is key to keeping critical corporate data from ending up in the public domain and possibly violating privacy and security regulations, according to Huang. In the case of generative AI, on-prem will increasingly will mean the edge.

"They have to do it on-prem because that's where their data is, and they have to do it close to the edge because that's closest to the speed of light," Huang said. "You want it to respond instantaneously. You also want it to be at the edge, because in the future, you want to have information from multiple modalities.

"The more contextual information we get, the better … inference that we can make. The ability to make those decisions as close to the edge as possible, where the action is, where all the data is, and where the responsiveness can be as high as possible, is really essential."

For Nvidia, which a decade or so ago put a bet on AI being a future growth engine, Project Helix further helps cement its position as a key enabler of machine learning for corporations and HPC organizations.

At a time when LLMs train on massive general-purpose datasets – in the case of GPT and the ChatGPT bot built on it, the internet – organizations want to train smaller models on their own data to address their own specific needs, according to Jeffrey Clarke, vice chairman and co-COO at Dell.

"That is the trend we see with customers," Clarke said. "How do they take their business context, their data, and help them make better business decisions? You don't need a GPT large-language model to do that. … Companies aren't going to deploy ChatGPT in a factory to make a factory work better. That will be a localized model by company X, Y, or Z with their data."

Giving more control

The push to enable enterprises to customize training models with their proprietary information and in their own datacenters is gaining momentum. Earlier this month, ServiceNow and Nvidia unveiled a partnership similar to Project Helix. The idea isn't new, but it's been supercharged with the recent acceleration in the development of generative AI and LLMs.

At GTC in September 2022, Nvidia launched the NeMo LLM service with that in mind, giving enterprises a way to adapt a range of pre-trained foundation models to create customized models trained on their own data.

General-purpose models like OpenAI's GPT-4 will work for some jobs, Das said, "but there are also a large number of enterprise companies who need to have their own customized large-language models for their own domain, for their own proprietary data, to make sure that the models are doing exactly what they need done in the context of their company."

"NeMo is a platform from Nvidia for those customers who need to build and maintain their own models."

Nvidia CEO Jensen Huang, who appeared in a video discussion with Clark during the keynote, said that "every company is at its core about intelligence."

"Project Helix … will help every company be an AI factory and be able to produce their intelligence, their domain-specific intelligence, their expertise, and then do it at light speed and do it at scale," Huang said.

Rapid innovation around generative AI also will give enterprises more options, Dell's Clarke claimed. Dell Validated Designs based on Project Helix will be available beginning in July. ®

More about

TIP US OFF

Send us news


Other stories you might like