AMD EPYC 8005 Sorano Completely Changes AMD SP6

Key Lessons Learned

The AMD SP6 socket is one that we were going to deploy as part of our storage and backup infrastructure with the EPYC 8004 series. While the SP5 socket is great, SP6 is smaller, and for many classes of servers, like storage servers, it offers a solid mix of I/O and CPU performance, but at a much lower TDP. For folks looking to get newer servers, but at lower TDP ranges, the option from AMD has been the EPYC 4005 with significantly less I/O, or the Zen 4c EPYC 8004. Now, with the EPYC 8005, we get full Zen 5. Not on the slide, but important is that really another way to think of the EPYC 8005 is that it is lower clocked, but more I/O, closer to the EPYC 4005 series, and that it can sit below the EPYC 9005 series in terms of power.

Not captured in the per CPU price data is that with fewer PCIe lanes and memory channels, motherboards for the SP6 platform can cost less than for SP5. We did not get to test these SKUs, but the EPYC 8435P and EPYC 8535P likely can consolidate well over 2:1 compared to 2nd gen Intel Xeon Scalable and at a lower power and using lower cost single-socket SP6 systems. If you need to replace dual socket nodes in a similar power envelope per node, the EPYC 8005 has a lot of advantages in this generation.

Taking a great example of the other end, the AMD EPYC 4585PX is a 170W TDP part with 128MB of L3 cache (with 3D V-Cache) and 16 cores at $709. The AMD EPYC 8125P has 16 cores, 128MB of L3 cache, a 125W TDP, and costs $799. While the EPYC 8125P has lower clock speeds, it also has three times as many memory channels, ECC RDIMM support, and 96 PCIe Gen5 lanes. When we did our AMD EPYC 4005 Grado is Great piece, folks asked for more PCIe connectivity and higher memory capacities. Given that it was in a different memory pricing environment. Still, the EPYC 8125P delivers in a similar price range.

Another big one is the Intel Xeon 6 SoC comparison. We only had up to the Xeon 6716P-B, which AMD also highlighted for comparison. The comparison is a bit challenging. Here is the way we think of it. AMD is providing lower cost per core and power consumption with more I/O. Intel is integrating its accelerators and NIC IP into the Xeon 6 SoC. Or in other words, AMD is providing better cores and enough I/O so you can build your own solution. Intel is integrating everything. Intel has been in the Xeon D space since 2015, and if you want QuickAssist and its up to 2x 100GbE networking onboard, then, being fair, it has a strong offering. The second you add an accelerator or NIC and do not use the integrated Intel solution, then AMD’s starts to look extremely attractive. An easy example is what happens when you want a 400GbE NIC instead of 2x 100GbE? What if you need a specific cryptographic accelerator that Intel does not support with QAT? Once you start adding cards that overlap with the built-in Xeon 6 SoC acceleration, then AMD does very well.

If nothing else, I think the performance was so good that the new EPYC 8005 series is not just a slightly better version of the EPYC 8004 series. Instead, I think AMD introduced a new class of computing in the same SP6 socket.

Final Words

It is pretty hard not to like the EPYC 8005 series. It offers a lot of performance, and perhaps more importantly, I/O in a lower-cost and lower-power packaged than the EPYC 9005 series. Far more than a generational core upgrade, AMD effectively re-positioned the EPYC 8005 series and SP6. While it is not as high-volume a CPU as the EPYC 9005 series, which limits the diversity of servers that can be built for this platform, the enormous performance gains in this generation likely mean it can do a lot more than folks were expecting.

This was probably one of the most surprising CPU launches we have seen, just because the changes looked so small on the slides, yet this generation turned out to be a big performance leap once we got hold of the parts.