ARM in the Cloud

I know I’m long overdue on a “Part 2”, but wanted to slip this in first. I’ve long been a skeptic on ARM becoming a mainstream processor choice for servers. But today’s announcement by Amazon Web Services of the ARM architecture Graviton2 processor and the M6g, C6g, and R6g instance families has me rethinking my position. As is often the case I’ll start with some historical perspective and then discuss today’s AWS announcement.

In the late 1980s and early 1990s it was widely believed that RISC (Reduced Instruction Set Computer) architectures would replace CISC (Complex Instruction Set Computer) architectures such as the then leading VAX, x86, IBM 360/370,Motorola 68000, etc. instruction set architectures (ISA). The reasons for this were two-fold. First, it was believed that in any given semiconductor process technology a RISC processor would have 2x the performance of a CISC processor. This was largely because with CISC you were devoting an ever increasing percentage of the available transistors to overcoming the inherent bottlenecks of the ISA, while with RISC those transistors could be devoted to increasing performance. Second, the complexity of designing CISC ISA processors had reached the point where the semiconductor technology could advance more quickly than you could design a processor for it, so you were always behind the curve of taking advantage of Moore’s Law. RISC ISA processors were easier to design, and thus would better track the semiconductor process improvement timing.

One thing to keep in mind was the original RISC concept was to create a new ISA every time you made a new processor. So you never really had to waste silicon on legacy, you did a new optimal ISA for each processor and made it a compiler problem to re-target to each processor. Of course software people quickly made it clear to the hardware folks that this was a non-starter, that the cost of supporting new processors effectively (finding and fixing ISA-specific bugs, tuning performance, issuing patches to released software, etc) would outweigh the performance improvements of not having a fixed ISA. So we moved on to fixed RISC ISAs that would survive through multiple generations of processors. By 1995 RISC was well on its way to world domination. IBM, Apple (Mac), and Motorola had moved to Power ISA. DEC moved to Alpha and HP to HP/PA. Acorn Machines was a PC manufacturer that created its own RISC processor (the Acorn RISC Machine) and operating system (RISC OS). Acorn would later shift its focus away from PCs to its RISC ISA, dropping “Acorn” in favor of “Advanced”, renaming the company ARM, and licensing its architecture and designs to various companies. Other RISC chips also appeared including the Intel i960 and the MIPS line. MIPS in particular looked that it would become “the Intel” of RISC processors, though it would eventually falter. And as we now know, ARM would be the only RISC ISA to really thrive, by riding the growth of the market for mobile devices. But at the start of 1995 it looked like we were going to have RISC everywhere.

So what happened in 1995? The Intel Pentium Pro. The Pentium Pro could hold its own on performance with that year’s RISC chips while maintaining full x86 compatibility. How did Intel do it? First off they clearly had made advances in chip design tools that let them move much faster than other companies working on CISC. And they adopted an approach of compiling CISC instructions into RISC-lie ROPS and then making the rest of the processor work like a RISC processor. But maybe more importantly, they had a generational lead on introducing semiconductor manufacturing processes. So even if the assumption that in any given semiconductor process technology RISC would be 2x CISC held, Intel being a process technology generation ahead negated that RISC vs CISC advantage.

Intel’s process technology advantage held for twenty years, allowing the x86 to extend its dominance from the desktop to the data center. With the exception of Power, which IBM continued to advance, RISC ISAs disappeared from the server and desktop world. But RISC had another advantage, its simplicity made it easy to scale down to smaller low-power microprocessors for embedded and mobile applications. Today pretty much every mobile phone and mainstream tablet uses a processor based on the ARM ISA.

A number of years ago ARM and its semiconductor partners began trying to bring RISC back to the server and PC markets where it was originally expected to dominate. On the PC front ARM has made some limited progress, particularly with ChromeOS PCs and more recently in Windows PCs such as Microsoft’s Surface Pro X. But so far that progress represents a tiny portion of the PC business. In servers we’ve seen several abortive efforts and essentially no adoption. Until now.

Last year Amazon Web Services introduced an instance family, the A1, based on an ARM ISA processor of its own design called Graviton. Side note, in most cases (Apple is the counter-example) semiconductor designers license not the ARM ISA but an actual “core” design from ARM. That is the case with AWS. This was a pretty niche offering, and (to me at least) signaled likely another failed attempt to bring ARM to the mainstream server market. For example the A1 was not something you could benchmark against their Intel-based instances and end up with a direct comparison. It was more niche targeted.

Today AWS brought their second generation ARM processor, the Graviton2, to its three most mainstream instance families. Those are the M (general purpose, or balanced), C (compute intensive), and R (memory intensive) families and we now have the M6g, C6g, and R6g families. They even did some performance comparisons of the M6g against the Intel Skylake SP-powered M5. And they were quite favorable to the M6g.  But Skylake SP is an older Intel generation, and a comparison with the Coffee Lake SP and AMD’s Rome would be more telling. These have already made their way into some models in the C5 and C5a families. Intel is also accelerating its product cycles so I expect it to regain a performance lead, though perhaps not enough to deter the growth of Graviton. Graviton is likely to retain a price/performance lead in any case.

So what happened to allow ARM to (apparently) catch up to Intel in the data center? I think there are three factors at play. First, recall the original RISC premise that in any given semiconductor process technology RISC should be 2x CISC performance and that this turned out not to matter with the x86 because Intel was a generation ahead on semiconductor process. Intel no longer has that generational advantage, and by some measures (e.g., smallest feature size) is behind semiconductor foundries such as the one AWS uses, TSMC. The second factor is we have a modern, some might say the most important modern, “system” vendor, AWS, leading the charge. Instruction Set Architectures have tended to thrive when backed by a powerful system vendor, not as pure artifacts of the semiconductor industry. x86 is the dominant PC and Server chip today because of selecting it for the IBM PC. ARM’s success came from DEC adopting it to create the StrongARM family, which was the dominant processor used in PDAs and early smartphones. Even ARM originator Acorn used StrongARM in its systems. Earlier dominant ISAs came from the system vendors themselves, particularly DEC and IBM. Now, just as DEC boosted ARM into dominance in the mobile device market, it looks like AWS will do the same for servers. Third, because AWS can optimize the ARM architecture and licensed core into its own chips for the controlled environment that is the AWS Cloud it can tune chip designs far more than someone trying to create a general purpose offering for on-premise servers.

So is it game over for the x86? After decades of watching, and working with, Intel I doubt it. Though it isn’t just AWS that Intel has to worry about. If AWS starts to show real success with Graviton than Microsoft and Google will be empowered to go full-bore with ARM in the cloud as well. And then there is the persistent rumor that Apple wants to move the Mac to ARM processors of its own design. With failed efforts to break into the mobile device market, pressure from all corners growing in the PC market, and now an apparently competitive mainstream ARM offering in the server market, Intel can’t stay on the modest improvement at high unit cost bandwagon much longer.

This entry was posted in AWS, Cloud, Computer and Internet and tagged , . Bookmark the permalink.

2 Responses to ARM in the Cloud

  1. Jim Finnerty says:

    Remarkable analysis, Hal. I really enjoyed it. Doesn’t the same argument from the optimizing compiler side still apply, though?

Comments are closed.