Conclusion

EPYC is a bold attempt by AMD to get back into the high-performance enterprise CPU market. With the diminishing Opteron brand based on the inefficient "Piledriver" architecture, only a completely new brand coupled with a radically different platform approach could put AMD back in competition with Intel and its bloating lineup of Xeon-branded SKUs. There are, as this is written, over 100 active SKUs based on the Xeon brand in the market, across half a dozen sockets and form factors and spread across three micro-architectures ("Broadwell" thru "Kaby Lake"). This gives AMD the opportunity to cut through the clutter with fewer sockets and SKUs to target very specific form factors in the enterprise space, namely datacenters. With a brand name that sounds sufficiently catchy to today's 30-something datacenter engineers making purchase decisions, AMD is out on the right foot.

With EPYC, AMD is imitating Intel's comeback chip circa 2005, the 65 nm "Conroe." This dual-core silicon made its debut with the Core 2 Duo and was so successful against AMD's lineup across client and enterprise segments that Intel didn't bother with a larger monolithic quad-core silicon till 2008. It built quad-core parts using multi-chip modules of these chips. It's relatively easier to build MCMs with a silicon you've perfected than trying to scale up your "perfect" silicon. For AMD, that silicon is the 14 nm "Summit Ridge," and it has had a nice start with the client segment.

AMD's entry to the enterprise segment with the "Zen" CPU micro-architecture it's based on is a large multi-chip module with four such dies, which makes for 32 cores, with SMT enabling 64 logical CPUs, 64 MB of L3 cache, and a vast 128-lane PCI-Express gen 3.0 root complex, which should power a great number of redundant flash-based NVMe storage devices and multiple 10 GbE or even 100 Gbps fiber NICs. There is, most importantly, a gargantuan 512-bit (eight-channel) DDR4 memory interface, with a 2 TB ceiling on how much RAM you can deploy per socket. AMD has clocked its 32-core EPYC SKUs relatively high compared to Intel chips with similar core counts, and that ends up giving AMD the competitive performance edge.

According to benchmarks put out by AMD, a machine with a single 32-core EPYC 7551P outperforms a dual-socket 2x 12-core/24-thread Intel Xeon machine, and a dual-socket machine outperforms Intel Xeon chips at all price points. Coming from AMD, a lot of cherry-picking, iffs and buts went into developing these numbers, but we're convinced that EPYC can be competitive with the very latest Xeon chips, and that IT managers should look up performance reviews from trusted sources before making purchasing decisions. Besides performance, AMD has added a cutting-edge hardware security engine which relies on a dedicated processor on each die besides the 8 CPU cores, which handles security. AMD is also deploying its advanced memory encryption engine with hardware-based key management. AMD has also deployed its latest adaptive power-management features it introduced with its client-segment chips.

In conclusion, AMD has a compelling 1P and 2P product offering, which competes with not just Intel's 1P and 2P, but also its vast 7000-series high core-count chips designed for the virtual server market. Besides good CPU resources, these chips also give you a solid platform with broad memory and PCIe interfaces that keep you from deploying multiple blades to accommodate your add-on hardware. AMD now has to deal with problems not related to the product itself.

AMD has robust industry relations with the game-console manufacturers and has monopolized semi-custom SoC manufacturing for that market segment. It also pushed the software industry for favorable yet open standards, which its client hardware is inherently good at. The company must now direct its efforts towards building similar industry relationships with the most popular server OEMs, who not only make products based on EPYC, but also allocate their own marketing budgets to promoting EPYC-powered solutions to their longstanding or prospective customers. The onus is also on AMD to ensure that EPYC chips live up to the reliability standards the industry has come to expect of competing Xeon chips, and to provide a robust after-sales support net to customers.

AMD must also nurture a vibrant ecosystem built around EPYC, which includes not just hardware and OEMs marketing them, but also the matter of pioneering software standards in the enterprise segment. This must include not just open- and open-source standards, but proprietary ones from the likes of Microsoft or Novell. Fortune 500 companies need to know that EPYC-powered machines can be relied upon to power their businesses. AMD should also build the brand in the high-performance compute (HPC) arena with turnkey solutions that combine its enviable GPGPU leadership with its homebrew EPYC processors. We're inclined to think NVIDIA isn't fooling around by investing billions of dollars into machine-learning, and AMD is in a position to combine its CPU, GPU, and software resources into uniform and reliable solutions driven by open standards. The PC and server CPU industry dominated by Intel is finally getting a much needed shakedown, and it will only benefit the industry at large.