Chipping Away At X86 Hegemony In the Datacenter
December 14, 2020 Timothy Prickett Morgan
Here at The Four Hundred, we have a saying: Anything that makes Power Systems stronger makes IBM i last longer. And part of making IBM i stronger, oddly enough, means just getting behind the idea of diversity of compute in the datacenter and that specifically means countering the notion that the X86 processor (and specifically the Intel Xeon SP implementation of it, but not exclusively because we now have AMD Epyc processors that are viable) is necessarily the only processor in the future of the datacenter.
We have always held this opinion, as you well know, and have long espoused this view, particularly in regard to the System z and Power processors from IBM, which have their own significant architectural benefits. We are not seeing a resurgence in these processors here at the tail end of 2020, so don’t get too excited. But we are seeing other architectures picking away at the hegemony of the Xeon SP CPU, and that is interesting and helps strengthen the case for an alternative choice like a Power9 or Power10 chip by extension.
Again, it is important to not get too excited this early in the game. To its great credit, Intel has earned its spot as the general purpose processor during the General Purpose Compute Era that ran from maybe 1999 through 2019, with some fuzziness around those edges. In the early 1990s as Intel created server-specific Pentium Pro processors that had features in them that did not go into the desktop Pentium processors and that have allowed Intel to jump from the desktop to the deskside server to the datacenter and, through very aggressive pricing, relentless software ecosystem development, and impressive engineering, finally becoming the no-braining choice for compute in the datacenter – and in the data closet and at the deskside, too.
In the quarter ended in September, IDC believes that X86 servers accounted for $20.93 billion of revenues – 92.8 percent of all server revenues – and very nearly all of the 3.07 million machines that shifted from manufacturers to customers and the channel in that 13-week period. (IDC does not supply this number, but we reckon that it is around 99.4 percent.) That was only 1.6 percent revenue growth for X86 server platforms and somewhere around two-tenths of a percent shipment growth, by the way. Non-X86 iron, which includes System z mainframe and Power Systems sales at IBM as well as Power server sales at Inspur as well as any Arm-based server sales. The Arm servers are seeing some action, despite many setbacks. Arm servers based on Marvell ThunderX2 processors and now Fujitsu A64FX processors are the common ones shipping in volume these days, but Ampere Computing also sells Altra Arm chips and perhaps most importantly Amazon Web Services etches its own Graviton2 chips and has absolutely rolled them into production servers manufactured by one of the ODMs (probably Quanta but maybe Foxconn, too) and these are in the numbers, too. Given that we know IBM’s server sales were down around 15 percent in Q3, and non-X86 server revenues were up 11.1 percent to $1.64 billion by IDC’s numbers, then we know these others systems did pretty well.
Now, to be fair, this was against a very easy compare to Q3 2019, when non-X86 server revenues were only $1.47 billion. But as you can see in the chart above, the general trend in non-X86 server revenues is slightly on the rise, and if Nvidia buys Arm Holdings for $40 billion as planned and aggressively designs, manufactures, and sells Arm server chips and perhaps whole systems, that curve could point upwards. The opening up of the Power processor ISAs and the use of Power systems in China will also help the non-X86 cause, and given the advantages that Power has over Arm, particularly for high throughput workloads, we think there is a good chance that all non-X86 boats will rise. Intel’s position is already weakened by its inability to manufacture server chips in volume at any process smaller than 14 nanometers, and at the same time AMD is coming in with very aggressive Epyc designs that have even more aggressive pricing. This will mitigate, to a certain extent, the architectural gaps between Power and Arm architectures against X86 architectures because Intel will cut prices to remain competitive in terms of price/performance. But that still won’t change the throughput and scalability and reliability advantages of a chip like Power9 and Power10.
At this point, non-X86 server revenues trend somewhere between $1.5 billion and $3 billion, depending in large part on where IBM is at with its System z and Power product cycles.
We estimate that IBM represented only 47.3 percent of non-X86 server sales in Q3 2020, and that the other suppliers of non-X86 iron accounted for $860 million in sales, up 53.2 percent. A lot of that was Arm-based systems, which grew by a factor of 5.3X according to IDC, admittedly against a very small revenue base. But the fact remains that sales of these non-IBM non-X86 machines helped fill in the sales declines that IBM saw in its System z and Power Systems lines, and if trends persist, they will be running at a $1 billion rate in 2021 and possibly a $1.5 billion run rate in 2022. Couple that to IBM mainframe and Power Systems cycles that will get a boost in late 2021 and early 2022, with the System z16 and Power10 launches, and non-X86 platform sales could be humming along at somewhere between $4.5 billion to $5 billion in 2022. If this should come to pass, then that would be levels of sales that the server industry will not have seen for a decade, when they were last running at that level in 2012.
There is even an outside possibility that the memory area network that is at the heart of the Power10 processor spurs adoption of Power iron both at IBM and Inspur – and maybe even among hyperscalers like Google, who claimed to be enthusiastic about Power chips for big workloads some years ago.
We remain hopeful.