Where HPE Adds Value to NVIDIA CPUs
Transcript
One of the biggest challenges I've encountered in deploying enterprise AI systems are organizations that try and make LLMs, which are by nature non-deterministic, solve deterministic problems. What did I see at HPE Discover that helped me better understand the relationship between NVIDIA's AI factory and HPE's capability? Let's get started. >> [music] >> All right, so we consistently know you cannot simply implement guardrails via prompts to make LLMs more deterministic. It does help with results in getting better prompt returns, but the same deterministic goal or prompt does not equate to getting a result that matches your organization's workflow.
So, I've developed this concept. We I call it deterministic code in the loop, where LLMs suggest but deterministic code decides. Basically, a LLM is good at odd-shaped problems, but code or tools, when we see it in agentic terms, decide on when some a process is finished or if the requirements are met or if the guardrails are done. What does this have to do with NVIDIA and HPE? Well, NVIDIA has been pushing their arm-based platforms, uh the future Vera Rubin and the current Grace Hopper platform as a combined platform as part of the their AI factory strategy to increase the performance and capabilities of tools.
Without a doubt, I think their stats and I'll leave you links to the speeds and feeds and the relationship between HPE and Nvidia their products, but more specifically Nvidia is leaning towards this problem that we're seeing in agentic AI. Independent research is showing that more and more we're not relying on the GPU as the bottleneck for agentic workflows versus in just pure inferencing. We're seeing the CPU as the bottleneck. So, as we try to confine LLMs to their role of advising, we need more capable pipeline or or or bandwidth and capabilities out of our CPUs that's leveraging the data.
In the next video, we'll talk about how this is impacted directly with networking and storage, but in this video, I wanted to focus on how does HPE differentiate in the crowded market of Nvidia AI factories. This relationship is special because HPE's history around HPC. They've been solving these deterministic problems with supercompute for 50 years. And these same experts are bringing that expertise to AI. We not only just see it in the Grace Hopper and Vera Rubin specific DL configurations, we're seeing it when they're bringing it with GPUs and their current existing supercompute platforms that these consultants are helping organizations figure out how to exactly do what I'm saying, which is let LLMs suggest, and but let the deterministic scientific-based systems decide.
If you want to learn more about the HPE and video relationship, follow the link below. If you want to learn more about deterministic code in the loop, I have a link to a migration document that I've created. Until then, talk to you next episode of the CTO Advisor or CTO Docs.
