Jack Poller on AI, Edge Computing, & Security by Design from the HPE ProLiant Gen12 Deep Dive Event
At the HPE ProLiant Deep Dive event, Jack Poller emphasized the growing importance of servers in AI and edge computing, noting HPE’s focus on security by design with features like the ISO chip and custom silicon. He discussed how AI workloads require specific performance characteristics like low latency and accuracy for inference, rather than just CPU power. Poller also praised HPE’s approach to efficient server design, like using 2U servers for better airflow and energy efficiency.
Transcript
I think the first and most important thing I took from the ProLiant day is that servers matter where they didn't use to. What we're seeing is a big explosion of new use cases that is being driven by, um, AI and the need to do things at the edge. And that means that we have different performance requirements, different physical requirements, different security requirements.
I think the second thing I took away is that for a lot of organizations, um, security first and security by design are just buzzwords and a lot of the organizations banding that about, and I think HPE really takes it to heart and really does it. Uh, the integration of the ILO chip into the ProLiant servers and with custom silicon is sort of the start of building a complete root of trust and building security first into the server architecture that goes all the way through the entire hardware and software stack. And so I was really impressed by that.
The ILO concept have been around for a very long time. Um, the need to do custom silicon rather than purchasing off the shelf hardware for that, a base band, generic base band controller, I think is the right way to go. Um, again, it's what's guaranteeing you your complete, um, trust in the entire, um, uh, supply chain, right?
And that that's, that's the way to go. Um, you know, there's a lot of features in it or all the right features. I, you know, doesn't the features sort in one sense don't matter as much more as the fact that you have that there.
Yeah. Um, that they're adding a secure enclave in. And again, that's getting more and more secure in the IO.
Um, it also acts as a TPM for itself and as a TPM for things like, uh, your certificates and keys for self encrypting drives and some of the other things. So that I think is the right way to go. And again, it's all about that security and, and guaranteeing it from the very, from day zero all the way through to the end of life.
I think, and, and I've had this discussion a couple times, you know, we had an argument of whether AI is just a workload or not. And I think, I look at it as AI is a workload that has a specific set of requirements, and we get, we often get cornered into thinking about AI as something very unique, uh, and something very special that, um, makes it different than anything else. And it, to me, in my mind, it isn't.
It's just it has a specific set of performance require, or a specific set of characteristics and requirements. We often think about it in the training mode because that's what gets all the press and training is all about CPU cycles and getting the absolute best CPU performance out of it in a lot of use cases. On the inference side, latency or accuracy of results or, uh, power density is much more important performance characteristic that we're worried about.
Um, and so I think it's important to distinguish that. And HPE is looking at things in those, with those other performance characteristics. So for instance, if you don't need CPU density, then why go to a, like a one U server to get the density?
Because that just increases your expense. And more importantly, it increases your power utilization. If you can go to a two U server where you get better airflow, you can have less power consumed by fans and cooling, and that's a better solution overall.
