Has AMD's Clock Run Out?

AMD recently announced that they massively missed their quarterly numbers, and that they plan to lay off 15% of staff. Of course they are blaming it on the economy, but as a long-time follower of the CPU scene (and a perpetual fan of the underdog AMD), I can say that isn't really the case. Sure, PC sales may have ticked down a bit, and even Intel's numbers lagged this quarter, but the real problem is that AMD has become irrelevant by failing to release competitive chips. They relied on the low-end market for the bulk of their sales, which was and extremely dangerous move (as I will explain later).

About ten years ago, I wrote a column about AMD's upcoming Athlon 64 processor, which was to be the first mainstream 64-bit processor. At that point, 64-bit processors were exciting because they finally allowed PCs to address more than 4GB of RAM, and also offered a number of improvements that are extremely exciting to CPU geeks. While Intel had offered 64-bit processors in a few high-end workstations, they were ridiculously expensive, and never really caught on. AMD was essentially offering an affordable processor with next-generation capabilities.

I actually bought one of the first-gen Athlon 64s soon after they launched. I loved that machine, and every now and then, I drag it out of the closet whenever I need an old Windows XP machine for some testing. The Athlon 64 had a long life as a desktop processor, and it was also popular in servers. A number of top tech companies used it in their servers because it could do more work per watt of power consumed (which is the most important thing once you reach that scale). In fact, Dell sold a diminutive computer just a couple years ago with an original Athlon 64 inside.

The problem was that the Athlon 64 was pretty much the last in AMD's string of hits. Intel quickly eclipsed them by releasing the first dual and quad-core CPUs, and managed to scale performance much more quickly. While AMD did pretty well in the server market, 64 bits didn't offer a huge advantage on the desktop, and Intel eventually released their own processors that were 64-bit compatible. By the end of the decade, AMD needed a new hit. They did manage to release an affordable 6-core processor  - I happen to own one of those, being the CPU nerd that I am. However, Intel's midrange i5 offered more performance at tasks that don't require massive numbers of cores, so they sold modestly at best.

AMD needed a new hit. They bet on an entirely new architecture, nicknamed Bulldozer, which completely rethought the way that chips were designed (I won't go into the specifics, but it attempted to compete with Intel's Hyperthreading technology). It promised up to 8 cores on a single chip, where Intel offered a maximum of 6. The problem was that Bulldozer was about a year late, and when it finally came out, performance was much lower than expected. In fact, the Bulldozer processors are usually slower than the generation that they replaced. AMD pretty much admitted defeat on that one, and announced that it was basically withdrawing from the high-end PC market (although they continue to make processors for the server).

So AMD retreated to the low-end, releasing chips that performed modestly for a reasonable price. Since acquiring ATI in 2006, they have focused on creating chips with integrated graphics capabilities. At the low-end, they offer significantly better graphics performance than Intel's integrated graphics, and a lower price. The problem is that the margins on these low-end PCs are small, and they are extremely sensitive to economic fluctuations (AMD actually had to take a write-down because they have so much unsold inventory). Although AMD managed to turn a profit in the past based on these processors, it's unlikely they will be able to continue this in the future. More importantly, ARM is taking over the low-end.

As I've pointed out in a few previous posts, ARM is quickly increasing its chip performance to the point where it overlaps with low-end PCs (for the uninitiated, ARM designs power most smartphones and tablets on the market these days). People are already replacing their laptops with tablets (one of the real reasons AMD's sales have dropped), and within a year or two, we will be seeing PCs and laptops based on the ARM platform. Google's just-announced Chromebook is only exceptional in that it features an ARM processor rather than an Intel chip. And rumor has it that Apple is considering moving its Macbook Airs to ARM, which would allow them to decrease size without compromising on battery life.

So AMD is being squeezed on both ends. At the top-end, Intel is beating them on performance, and on the bottom, ARM manufacturers are eating their lunch. They have a few reasonably competitive products in various niches, but that's about it. This is pretty much the classic story of a company undergoing disruption by competing technologies. My guess is that it struggles along in mediocrity for a few more years, and then either exits the processor business entirely or is acquired by a competitor. It's possible that AMD will pull an ace out of its sleeve with the next generation of processors (currently not even on the roadmap), but it seems like this might be the end of the rope.

1 response
So yes, Intel tried to cripple x86 and make 64 bit itanium only at a huge price and price/performance penalty.

AMD created the x86-64 instruction set, sold at at zero premium and improved their market share. Enthusiasts got behind it, and the rest of the market followed. Intel's P4 while delivering high clock speeds didn't deliver particularly good performance.

AMD was the first (for x86) to move the memory controller on board. This improved memory latency, decreased power usage, and improved performance.

Then with the first opteron they added TWO memory busses per socket (4 per dual server) and it took years for Intel to kill off the shared north bridge and bring out the nehalem based xeons. The difference was so big that in general unless you were cache friendly, CPU performance scaled on AMD opterons and did not on Intel.

Hypertransport was also huge, allowing 2-8 sockets, performance scaled well, and it was a cheap, fast, and low latency interconnect that saw some popularity even outside of x86-64 systems. Mips, some network chips, some routers, and related adopted it. Sadly despite removing the custom glue needed for quad sockets AMD never pushed their advantage with quad sockets. Instead AMD quads were priced like Intel. If they sold their quad chips at the same price as the dual I think they would have really pushed their advantage. Intel was quite hamstrung by the lack of onchip/per socket memory controllers, lack of hypertransport, and the lack of glueless SMP.

AMD's real weakness however is Intel has consistently had a year or more advantage on the CPU process. So AMDs always been behind, but has used better designs to compete. Unfortunately for AMD, Intel has improved on memory latency, now has 4 memory busses per socket (just like AMD) and brought out a hypertransport of their own (QPI).

So Intel's been shipping 22nm with similar technology: integrated pci-e, integrated memory controller, qpi, etc. AMD is trying hard to compete on 32nm, but this has meant lower IPC, lower single thread performance, but still pretty good on parallel workloads. Unfortunately desktops have a mix of serial and parallel codes, and thus AMD doesn't do particularly well there.

However even today AMD does quite well on servers. Depending on your app the 16 core/socket bulldozers are quite competitive. Sure Intel still wins single threaded benchmarks, but that's less common on servers and AMD ships quite a few more cores. 32 core AMD servers typically provide more throughput than the similarly priced 12 core/24 thread intel servers.

So AMDs got a tough battle, I would rule them out. Some possibilities:
* better OpenCL or similar vectors than intel. Intel's still lagging on both
* Seamicro (purchased by amd) has some promising technology for excellent
perf/watt while maintaining a healthy interconnect
* Increasingly power/watt, mobility, and flexibility are taking priority over
compatibility. WinTel is no longer a requirement for a large part of the market.
AMD seems more willing than Intel to abandon x86. Especially now that
Microsoft seems to be focussing more on non-x86 (Windows RT), even at
the cost of lack of (*GASP*) x86 compatibility.

AMD has announced they are going to stop bashing their heads against Intel in a battle over the mostly stagnant x86 market and pursue higher growth markets. I've not heard anything about the details, but a arm license seems like a possible solution. Add Seamicro's interconnect, hypertransport, and a cortex-a15 on a chip might just provide perf/watt than an intel x86 couldn't match.