A closer look at AMD’s server assault in 2014

Last month, AMD announced its new Kyoto server processors – low powered, Kabini-based APUs designed to compete with Intel’s S1200 family. Today, the company is updating its roadmaps with information on 2014, including data on the launch of new ARM processors, server APUs, and the first server chips based on the upcoming Steamroller core. These products are a major assault on a server market that has slipped mostly back to Intel, which currently controls some 96 per cent of server shipments.

Berlin: APUs go mid-market

Kyoto is AMD’s first server chip to combine CPU and GPU, but it’s also a SoC that’s principally aimed at emerging markets. Berlin will ship into an established market segment where server CPUs have traditionally placed very little emphasis on the GPU. AMD is clearly hoping to change this – the Berlin family will incorporate a GCN GPU, four Steamroller cores, and support dual-channel DDR3-1866.

The fact that AMD is listing Berlin as a replacement for the Opteron 3300 series says good things about the potential efficiency of the new hardware. The Current Opteron 3300 parts are a mixture of four and eight-core parts clocked at 1900MHz – 2800MHz. That should put the Berlin family ahead of the quad-core Opteron 3300s, but likely behind the eight-core Opteron 3380. While we expect Steamroller to be significantly faster than Piledriver, it’s unlikely to deliver a 2x performance improvement.

Unlike previous Opterons, which had fairly large L3 caches, Berlin drops the L3 altogether but keeps the shared L2 at 2MB. This could hurt the chip in certain server workloads, but it’s noteworthy that AMD is confident enough about the benefits of its heterogeneous system architecture (HSA) that it is pushing graphics into mainstream servers.

We’ve yet to see any benchmarks showing how HSA improves performance compared to non-HSA systems, but that should change in the months ahead.

AMD is going to position Berlin for web/enterprise services, possibly with some cloud computing workloads, but those are likely to be the province of its first ARM architecture, codenamed Seattle.

Seattle: AMD’s ARM architecture ascendant

AMD’s decision to break with decades of x86 expertise and build an ARM core was the ultimate mixed message. On the one hand, it was an admission that the company couldn’t afford to build its own ultra-low-power x86 core and attempt to chase the smartphone market against Intel, Qualcomm, and Samsung. Going with ARM, however, gave AMD the ability to leverage pre-existing IP and build a unique solution of its own in vastly less time than it would take to build an ultra-low-power variant from scratch.

What’s surprising about Seattle is that AMD is claiming “2x-4x the performance of Opteron X-Series.” The X-Series is Kyoto, the Kabini-based part; the quad-core X1150 is a 2GHz CPU running at 17 Watts. Given that AMD only claims a “2x+” performance benefit for Berlin over Kyoto, this implies that Seattle will actually be much more efficient than Steamroller. For now, we’re going to take that with a grain of salt. Generally speaking, low power processors have smaller caches and lower total memory bandwidth, and this could prove to limit the CPU when compared to higher clocked APUs.

AMD is clearly serious about stealing market share with Seattle; the new SoC integrates 10GbE, features 8-16 cores, supports up to 64GB of RAM, and leverages SeaMicro’s “Freedom Fabric” for dense compute solutions. Interestingly, they don’t mention SeaMicro’s technology in conjunction with Kyoto, Berlin, or Warsaw. It’s possible that Chimpzilla is holding back on talking up the SeaMicro connection, possibly to avoid sabotaging its arrangement with Intel to continue selling the company’s processors in those systems.

Warsaw: Steamroller strikes back

When AMD updated its roadmaps last month, it showed Steamroller/Kaveri as launching in 2013, but gave no word on whether or not we’d see a Steamroller CPU core for desktop. That question hasn’t been answered yet, but we now know more about the server variant of Steamroller, codenamed “Warsaw.”

This is a little troubling. Both Berlin and Seattle got enthusiastic figures attached to them with claims that the cores are “2x-4x” better than the other low power solution, Kyoto. With Warsaw, AMD’s praise is an anaemic “better performance/Watt than the 6300 family.” That’s certainly a good thing, but it doesn’t tell us much. It’s possible that AMD is playing coy with this information to hide its expected performance from Intel, that it hasn’t settled on clock speeds yet, or that it’s emphasising the parts it believes are more critical to its roadmap (ARM, low power x86), and going light on details for big cores.

Alternatively, yes, it’s possible that AMD isn’t talking up performance because Steamroller, while an improvement, doesn’t do enough to put the company back on a competitive footing with Intel. We don’t know yet.

News that these chips are drop-in replacements for AMD’s server hardware, however, increases the chance that the CPUs will be drop-in replacements for AM3 hardware, as well. No, that’s not a given – we’re not claiming that current AM3 boards will be upgradeable to Steamroller cores, when they become available on the consumer market, but if the chip is a drop-in replacement for one set of hardware, that increases the chances it’ll be a drop-in replacement for the other. Even if Steamroller is a modest upgrade to Piledriver, giving the AMD faithful a path forward one more time would be a smart move.

Takeaways from AMD’s 2014 roadmap: ARM the torpedoes

After seeing the company’s roadmap, we expect AMD to focus its efforts on building support for its ARM-based hardware and in leveraging HSA/OpenCL acceleration. The company’s verbiage implies that the next-generation Steamroller parts will debut in 8-16 core solutions as well, which means AMD won’t be beefing up the multi-core offerings any more than it has already. That’s likely a smart move; improving single-thread performance is more important than pushing additional thread capacity.

Analysts, meanwhile, are fond of AMD’s ARM ambitions. Barrons has an enthusiastic article on AMD’s chances in the server market, with some analysts believing the stock could hit $8 (£5). AMD closed today at $4.05 (£2.50), and while that’s still way down from the $6 (£3.80) price it enjoyed a year ago, it’s far better than the sub-$2 cost we saw last year. How much you buy the hype around ARM in server markets depends largely on how willing you think the market is going to be to embrace a non-x86 solution.

There’s a pro and con to that argument. On the one hand, the 1990s server companies sneered at Intel as a company that made “personal” computers – toys, really – but that couldn’t compete against the big iron from Sun, Alpha, SGI, HP, and IBM. The parallels are hard to miss – ARM could be the 21st century company that made ‘smartphones’ but couldn’t handle real computing – until, of course, it can.

Even given the historical precedent, we’re leery of the “AMD + ARM = Intel crusher” argument. Twenty years ago, Intel went after big iron vendors that refused to admit their own weakness. ARM’s server initiatives will deploy into the teeth of 22nm, out-of-order Atom solutions. When we’ve talked to Intel about the price question, the company has quietly assured us that it won’t be out-priced in a market where it chooses to compete.

Ultimately, we agree with Barrons that ARM server parts could gain significant market share and that AMD is making a smart move by investing in them. The question is, how much of that market will belong to AMD? If 2013 has been AMD’s year to make a major push in tablets and notebooks, 2014 is the year of the server. Warsaw, Berlin, and Seattle will collectively hit the market in the first half of the year with the goal of clawing back market share in multiple scenarios.