Intel has a number of announcements at ISC 2023. We want to cover some of the key highlights. Among them, we have a range of updates, but perhaps the two most impactful are that Aurora did not submit a HPL Top500 run but we have more specs and its Max GPU will be shipping. Falcon Shores appears as though it is going to bring Habana on board.
Intel ISC 2023 Not Yet For Aurora and an Upcoming Falcon Shores Habana Merge
Intel went over a lot that it disclosed during the Intel DCAI 2023 Update a few weeks ago. We are going to focus on the incremental disclosures. before we get to those, there were a few ones that were very interesting. Perhaps the most interesting was that, although dual CPU-only systems still make up the majority of the Top500, the Top100 has been trending towards having different CPU and accelerator ratios.
Intel had a bit about how that ratio has evolved, but perhaps the most interesting is what it put as its converged HPC and AI CPU to GPU and AI ratios. While Intel is not trying to convey exact numbers, it seems to suggest that CPU only will shrink, 1-4 GPU will be smaller in terms of overall acceleration, and that >4 GPUs per node will be common.
That is still a future-looking guess, but it is telling that Intel is focused on the GPU and AI accelerator being the most common components.
At the last ISC in 2022, folks were eagerly awaiting Aurora numbers. A year later, Intel said there has been progress, but we do not have a Top500 entry. Still, Intel gave us new specs on the machines including the fabric, memory and storage specs.
Intel says it has now delivered over 10,000 blades and that 75% of the blades have been upgraded to Xeon Max CPUs that were originally shipped with non-HBM parts.
The other Ponte Vecchio/ Intel Data Center GPU Max series announcement is around a x8 UBB. Intel tagged this as a new disclosure. We knew that Ponte Vecchio was OAM so I had assumed this was already part of the plan. We first looked at Inspur 8x OAM UBB almost four years ago. Now Inspur and Supermicro we are expecting will be the first two with the 8x GPU Max systems.
The vision of UBB is that it can fit different accelerators from different vendors. The AMD MI250X uses UBB as does the Intel Habana Gaudi 2 AI accelerator.
Here is the expected availability of the new GPU Max series.
It looks like we will start to see it more outside of Aurora deliveries soon.
Intel Xeon / Intel Xeon Max Updates
On the Xeon Max series, there were a lot of Genoa and Milan-X comparisons. The challenge Intel has is that with Genoa-X AMD is going to get 50% more cores, the microarchitecture/ memory update, and then the cache update, and AMD has said that is coming soon. Still, there will be many applications where the capacity of onboard HBM memory will be a major driver of performance.
We are going to leave that one until later this summer to discuss it from a comparative perspective. We are hoping to get to Xeon Max soon though.
Intel also discussed its upcoming Xeons, and again discussed MCR DIMMs. We covered that in the March 2023 update linked above.
While that is the current Max series, Falcon Shores is coming.
Intel Falcon Shores Update
Intel discussed that it thinks it is too early to integrate CPU and GPU like NVIDIA and AMD are doing. Instead, that will be a product past Falcon Shores. We covered that Intel announcement previously. Intel at ISC was talking about how the GPU will use HBM3, have Ethernet connectivity, and CXL.
It also made a minor roadmap tweak, and this did not get marked as a new disclosure because it was part of the recent earnings package. Intel shared that past Habana Gaudi 3, it will have AI accelerator integration options with Falcon Shores.
Intel knows it needs to continue pushing on its AI capabilities as that market is growing rapidly while traditional HPC is not on as big of a growth curve.
Overall, we have some good updates on a number of HPC and AI fronts from ISC 2023. We will update this article during the keynote if there are more disclosures made. The key messages were clear in the pre-brief. Intel is serious about expanding its AI portfolio in the data center and the Intel Max series will be a big focus this year.
Will there be an Emerald rapids for HEDT, like the W-24xx? at what price, similar to Sapphire Rapids? When?
P.S: thank for your incredible work.
I saw a lot of slides about Falcon Shores without “Light Bender” on them. You guys need to get Jeff McVeigh drunk and give him a napkin and find out what’s up.
To be fair, I did hand Jeff my second beer when we were in the bar line at an event in Hamburg and the bartender gave me two.