At AMD Financial Analyst Day 2022, the company gave an update on its data center business and specifically the AMD data center roadmaps. We already covered the Zen roadmap in AMD Technology Roadmap from AMD Financial Analyst Day 2022 and also the new AMD EPYC chip updates. Forrest Norrod is going to talk about the AMD Instinct MI300 so get excited.
This is being done very fast live, so please excuse typos.
AMD Data Center Solutions Update at AMD Financial Analyst Day 2022
AMD has been building a larger data center portfolio. Adding Xilinx and Pensando recently.
That is leading to a larger TAM.
Today, the MI250X is the top GPU for HPC right now. This is the chip that was deployed for the #1 supercomputer, Frontier. It also has taken a lot of spots in the latest Top500 list, really shifting momentum on the new systems added to the list.
AMD says the MI250X is also being used for AI.
The next generation of AMD data center GPUs are not just GPUs, they are APUs. Moving to the AMD Instinct MI300, AMD is adopting a trend we have seen others go down, unifying the CPU and GPU with shared memory between the two. Apple does this with the M1 and recently M2. Intel will be doing this with Falcon Shores.
AMD says that the new MI300 coming out in 2023 will offer 8x the AI training performance versus the MI250X.
Pulling this from the CDNA 3 presentation, this will package the CPU, GPU, caches, and HBM together in 3D chiplet packaging to make a big APU.
AMD will have a unified memory architecture to reduce redundant copies. This is actually an important nuance. NVIDIA Grace Hopper will have LPDDR5X for the Arm cores and HBM for the Hopper. AMD is going unified more like the Apple M1/ M2 but with HBM and caches.
AMD is going beyond just CPUs and GPUs. With Xilinx, it now has Solarflare NICs that are latency optimized, Alveo for SmartNICs, and Pensando for DPUs.
AMD discussed its Alveo FPGA NIC. For those wondering, this is a bit different than the Intel IPU we just got hands-on with.
Next up is the Pensando Elba DPU we covered years ago. We are excited to hopefully see Pensando DPUs at some point since we have already looked at solutions from other vendors. Pensando has yet to be able to get us a DPU, but hopefully, that changes with AMD.
Pensando has a software stack that can be used in whole or in part. For AMD, adding this stack is a big deal since frankly, AMD did not have all of this before the Pensando acquisition.
AMD is talking about how Aruba has a switch with Pensando and NetApp uses Pensando on some systems. Just a note, NetApp often uses technology backed by certain VC firms.
This is all part of AMD’s securing the data center solution. AMD was far ahead on technologies like confidential computing and secure memory encryption and now it is extending a vision beyond the CPU and out to other devices.
A lot of great stuff in this presentation.
The Instinct MI300 is super exciting and we hope to see more on Pensando soon. This move to the more APU-like data center GPU is important because it greatly reduces data movement. Power has become a huge limiting factor in the data center and moving data adds both power and performance penalties that the new APU architecture should help with.
On the Pensando side, hopefully, we will get to show STH readers that one of these days like we have solutions recently from NVIDIA, Marvell, and Intel. We also have Broadcom’s Stingray now that we will show off in a few weeks.