Today we have the launch of the AMD EPYC 7002 series. The last “2” digit represents that it is the second generation product, formerly codenamed “Rome.” To say this is a highly anticipated release is perhaps the understatement of the decade. As I have traveled around speaking to vendors, customers, ISVs, even teams designing hardware for future space missions, there is a common theme: giddiness over Rome.
Middle-aged folks from all over the world who have seen the industry stagnate for years have been giddy with anticipation. It is like the first time you drive in a Tesla automobile and the giant screen peers at you while launching almost silently to 60mph faster than most sports cars. We have been seeing the early ship customers express that level of excitement about the new platform. If Intel Xeon is the BMW 3-series, AMD EPYC is the Tesla Model 3 of the industry. Ask the automotive folks where momentum is there.
We are going to have a comprehensive look at the new chips in this article covering what the new CPUs can do, and why AMD EPYC is now firmly ahead of Intel Xeon in terms of bringing the future of server architectures to market. The AMD EPYC 7002 series goes well beyond just more cores, faster memory, and faster PCIe. Instead, it is the first transition into how servers will be built in the future as a paradigm shift will hit the server market. In our review, we are going to show you why this is now the hottest platform on the planet for anyone who wants to build a server product. We are also going to take a step back and look at where Intel Xeon still has a lot of value, and what the second generation AMD EPYC means to the IBM Power and Arm server ecosystems.
This is going to be a long, and very thorough piece. Our team has been working on this article, and even just keeping STH going while we have been working on it, for the past few weeks.
The AMD EPYC 7002 Elevator Pitch
I know a lot of folks are not going to be able to read this entire review on day 1. I wanted to simply share the “elevator pitch” version:
- The AMD EPYC 7002 series spans 19 public launch SKUs from 8 cores to 64 cores with up to 32 billion transistors
- AMD has made its I/O die the center of the server. This chip is the first step in disaggregating compute, memory, storage, and other I/O in servers. Other vendors will follow this paradigm shift. With AMD, the CPU core is no longer the center of the server which is a more modern systems approach.
- AMD has roughly twice the performance per socket in mainstream applications as Intel has throughout the range up to Xeon Platinum.
- AMD now supports PCIe Gen4 while Intel’s PCIe Gen3 limits them to half the I/O bandwidth per lane of AMD EPYC 7002.
- AMD EPYC 7002 has more PCIe Gen4 lanes up to 128 usable lanes in single-socket and up to 160 lanes in dual-socket servers. Effective single-socket bandwidth is comparable to quad-socket 2nd Gen Intel Xeon Scalable.
- AMD EPYC 7002 delivers the same core counts in fewer sockets and fewer NUMA nodes than Intel Xeon can in Q3 2019. The cores themselves are comparable in performance to their Xeon counterparts.
- Memory capacity has doubled, and the theoretical bandwidth has increased by 20%. Increased cache sizes help AMD deliver more performance.
- AMD now has a massive power consumption per core or performance advantage over Intel Xeon, to the tune of 2x or more in many cases.
- AMD list prices on competitive parts are ~25-35% of what the compute performance would cost to get from Intel Xeon comparables
As some perspective, the $4425 list price AMD EPYC 7702P essentially puts the capability of a quad-socket Intel Xeon Platinum 8253 ($3115 each) system into a single socket. The systems have the same memory capacity, about the same raw CPU performance, and the AMD EPYC 7702P solution has slightly more PCIe I/O bandwidth. The Intel Xeon Platinum solution there has more raw memory bandwidth but now has four times as many NUMA nodes as the AMD solution, and requires a more costly quad-socket server.
Here is a quick framing summary view:
Here are the numbers driving that view:
AMD is now engaged in an asymmetric attack against the legacy Intel Xeon architectures. Intel has a competitive response in its 2020 “Ice Lake” generation, but its 2nd Generation Intel Xeon Scalable chips released just four months ago are not competitive in many markets. This is all due to the 10nm process delays which pushed next-generation Intel Ice Lake out and therefore has the AMD EPYC 7002 competing against a legacy Xeon architecture.
To make this more salient, we have been testing systems from AMD and a number of vendors. We also have an enormous set of Intel Xeon test data. Once we started testing the AMD EPYC 7002 series we have only been buying for AMD systems except for one Intel Optane DCPMM solution for our lab servers.
In the remainder of this article, we are going to get into the details of the points above and show you many of the items that matter in this launch at a practical level, not just in marketing slideware.