Within one month, both Intel and AMD announced massive, long overdue refreshes of their server processor families, with EPYC and Scalable Xeon (yes, the credit card-style precious metal ranked) CPUs out. Both of these required brand new sockets, boards, chipset, even the memory and I/O configurations. As their arrival also significantly changes the competitive performance standing between the two X86 rivals, which way should the Taiwanese vendors go towards?
Firstly, after both launches, the obvious impact is that the AMD platform has, in both benchmarks and actual application performance, come close enough to - and in quite a few cases, matching and exceeding - Intel, that it deserves far more space than before. Add to this 33% higher memory bandwidth and capacity per socket, plus equally higher PCIe scaling in the most prevalent 2-socket configurations, and there's meat to grind here - even including AMD's usual well proven ability to shoot itself in the foot when it comes to winning the market battle. Intel may have brought in AVX512 floating point into the picture, but it requires software recompilation and, having it enabled at full throughput only on some of the new Xeons (the pricey ones), it may not gain much traction until the next refresh.
Secondly, both platforms have massively increased board size and power requirements per socket, making the dual-CPU ATX-sized server or workstation mainboard, a mark of integration excellence we last saw from Gigabyte and Asus in the Broadwell-EP Xeon generation, a virtually impossible thing to do. AMD may still have a theoretical chance here, as its CPUs don't necessitate a chipset (the minimum USB 3 ports can run from few of the CPU's PCIe lanes) and, using 8 pcs of ECC SO-DIMMs per socket (plenty of them shown by Taiwan firms at Computex), squeeze quarter of terabyte of memory and two sockets into the ATX constraint. Otherwise, it's big and bigger boards only, with correspondingly beefier power supplies, liquid cooling and so on.
Third, the above increased actual board and system costs (to the makers too looking at the BOM) are coupled with the increased average and peak CPU prices themselves. As expected, Intel has broken into five-digit US$ per socket stickers with few of its "Platinum" Xeon SKUs, getting itself into the pricing range of the still architecturally superior IBM POWER9 realm which, after all, can claim true enterprise mainframe legacy, rather than PC descendance. Knowing the general customers more or less stagnant per-unit server budgets, this could put further margin squeeze on the Taiwan vendor share of the pie, nudging them to go there where lower CPU price allows for more margin space. Or, aim towards the very top end, where, again, the antique X86 architecture isn't seen as a plus – watch POWER9 & Volta likely claim the next TOP1 system with USA "Summit" this November, setting the bar for Taiwanese vendors, like Wistron, to sell aggressively priced baby 2U server versions of this globally.
In integer data-dominated commercial apps, a 2-socket 64-core AMD EPYC solution performance would roughly equal that of 2-socket 56-core Intel Xeon - but at 3x lower CPU price yet 33% more memory and I/O. That extra memory and I/O can justify the board vendor charge somewhat more for the board itself, with the related margin benefit to boot. AMD wins here.
Floating point-bound jobs, like HPC and to some extent AI, would give Intel a bit of an advantage if using AVX512 extensions, but again that is limited to certain Intel SKUs only - Intel's selective disabling of major features across different SKUs of same product has its drawbacks too. It doesn't help that these extensions are just slightly different from those used on the last generation of Xeon Phi, which Intel is expected to retire after the last Phi, the Knights Hill, appears next year - another recompile needed? So, the HPC users - an increasing customer base for Taiwan board makers - could decide to wait for Intel's next refresh where AVX512 will be more prevalent, to jump on board the new platform.
In both cases, the new AMD and Intel platforms are kind of a pilot run for their next year's immediate plug-in successors in denser manufacturing processes, which should enable now-missing features like PCIe V4, faster DDR4 and full AVX512 on both vendors CPUs. AMD definitely offers better bang-for-the-buck with their new line; however it has to prove its manufacturing rollout will match the pent-up industry demand for a true high-end X86 competitor to Intel.
I feel the best bet for Taiwan vendors X86 server strategy right now to give equal attention to both vendors in the product line spread, and watch how AMD actually performs over the course of this calendar year. Intel has exhibited obvious long time quasi-monopoly fatigue with its high-end products, with need for a radical new CPU core way overdue after over a decade since its Israel office created what would become the "Core" - and saved Intel's then-CPU business from what was then-superior AMD architecture. We come back full circle, don't we?
