AMD’s microserver confidence

The man in charge of the company’s server business says ARM is coming for Intel

21 November 2013 by Yevgeniy Sverdlik - DatacenterDynamics

AMD’s microserver confidence
AMD's SeaMicro 1500 servers deployed at a Verizon Terremark data center

For Andrew Feldman, microservers and System-on-Chip (SoC) are a lot more than just another variety of compute. The man in charge of AMD’s server business expects nothing short of a revolution, and a big reason for that revolution, he says, will be ARM Holdings, the UK company that licenses its low-power processor-core architecture to chip makers. ARM chips power most of the world’s smartphones, but many vendors in the data center IT food-chain are working around the clock to crank out servers, storage and networking devices with low-cost and low-power ARM-based SoCs instead of the expensive network silicon or powerful and even more expensive x86 processors.


Today, the server business is a tiny drop in the ocean for ARM. Lakshmi Mandyam, director of server systems and ecosystems at ARM, says out of the 8.7bn ARM-based chips that were shipped last year, only a few thousand were server-oriented products. “If you look at it from a volume perspective, it will always be relatively small,” she says. It will be a very different picture from a value perspective, however. Size of the enterprise silicon market in 2017 is expected to be about US$20bn, about $3.5bn of which will be specifically server silicon, she says. ARM, by that time, expects to have 5-10% share of that market, “depending on how everything goes.”


We sat down with Mandyam and Feldman in September to ask how and why they think the market will move in that direction.


Intel wins the first-to-market race

AMD announced its first 64-bit ARM product back in June. Codenamed Seattle, the eight- and 16-core SoCs will support 128Gb of DRAM, operate at 2GHz and higher and have integrated 10Gb Ethernet. The company is proposing the chip for Memcached, Apache Hadoop, Cassandra, video transcoding and streaming, cloud-hosted gaming, virtual desktops, hosting and other applications along these lines.


Feldman says customers will be able to sample the part in the first quarter of 2014. AMD expects to start shipping it in the second half of the year. Intel has used announcements by a number of companies planning to bring ARM chips for servers to market in 2014 to illustrate how much ahead of the competition it is. The company said at the Intel Developers Forum in September that its 64-bit Atom-based server SoCs, codenamed Avoton, were already out.


But Intel competitors are saying that’s not true. Both Feldman and Karl Freund, VP of marketing for Calxeda, another SoC vendor using the ARM architecture, say they have not been able to actually buy an Avoton SoC. “You can’t buy it,” Feldman says. “They will be available beginning of the year.” Calxeda has tried to get its hands on the SoC both in the US and in Taiwan and was unable to, Freund says. “It’s not available today.”


Still, the two concede that Avoton will be available before any competing ARM product for servers. “Avoton will be four-six months ahead,” Feldman says about Intel’s lead on Seattle. While he has not actually seen one, he does no believe Avoton will be that good of a part. “Avoton is better than the existing Atom – better than Centerton – but not a big-enough step forward at all,” he says. “Remember, I’m the only guy with any experience actually building servers with Atom parts.” SeaMicro has been – and continues to do so today as part of AMD – building servers using both AMD and Intel chips.


“And it’s part of the sort of schizophrenia that you hear out of Intel,” he says. “They’re unsure whether they want to have a part in the category, and they want to be sure [when they do decide to produce the part] that you think it’s a leadership part, but they’d actually really rather people buy off the Xeon end of the line.”


Jason Waxman, Intel’s general manager of high-density servers, told is in an interview that he expected Atom-based SoCs to replace Xeon servers only for certain workloads. These would primarily be applications like hosting, networking and cold storage (low-performance arrays that store infrequently accessed data). When they do replace Xeon chips, however, it will not be a one-to-one replacement, Waxman said. Intel expects multiple Atom servers to replace every Xeon-based x86 product.


It’s not just AMD and Intel

Intel is obviously not AMD’s only worry on the ARM data center front. Contenders include fairly young companies, such as Calxeda and Applied Micro, as well as veterans, the likes of Texas Instruments and Huawei. The same is true in the system market, where AMD’s ARM server business will be competing both with the likes of HP and Dell as well as with smaller newcomers like Marvell, Cavium and Penguin Computing.


Feldman’s answer to this portion of the competitive landscape is again SeaMicro’s – and by inheritance, AMD’s – experience building and selling servers. “One of the ironies here is that in the ARM space, there are a ton of people who have made ARM SoCs, Qualcomm and Samsung being the best, without question,” Feldman says. “Those guys are monsters at making ARM SoCs.” These are companies that manufacture ARM processors inside the bulk of the world’s smartphones. Apple designs the ARM processors inside iPhones but uses Samsung as manufacturer. None of them, however, have made an ARM server before, Feldman says.


“While others have more experience with ARM SoCs, none of them have an iota – I mean single. Shipped. Part. Ever – of the 30 IP blocks needed to build a server.


That’s the differentiation that we bring as AMD,” he says. You do not get that kind of experience by building it right once, he adds. It’s also not a one-chip race. “If you want to get in the server race, you need a new chip every 15 months for the next 10 years. And I don’t think anybody thinks Applied Micro can do that.” Applied Micro representatives did not respond to a request for comment.


Calxeda shining ARMor

Calxeda’s Freund says the company has been putting out new chips regularly and plans to continue to do so. “This is exactly what Calxeda is doing, except we might be a little closer to 12 months,” he says. “We published a road map last fall, when we began shipping our first-gen product, and we are executing to that schedule.” Calxeda expects to start shipping its second-generation SoC this fall and to sample its third-gen product in the first half of 2014. “So yeah, 12-15 months sounds like a great idea.”


Addressing the question of experience building servers, Freund says one would be hard-pressed to name a processor or a server maker Calxeda does not have an engineer from.


The Austin, Texas-based company has attracted former employees of Dell, IBM, HP, SGI, Cray, as well as AMD and Intel. “It’d be hard to find a company in the industry from whom we haven’t recruited talent,” he says.


Calxeda already has an ARM-based server SoC on the market, although its a 32-bit one. Its 64-bit product (the third generation) is expected to become available in the second half of 2014. Meanwhile, it is marketing the 32-bit part for applications like storage and networking. While he agrees that 64-bit SoCs are required for most data center applications, 32 bits is fine for now, since no operating system or software has been written for 64-bit ARM anyway. “The market is expecting 64 bit to happen really really soon, and the fact is, these things take a little bit more time,” Freund says.


From Facebook to enterprise

Both he and Feldman say growth of the segment will, to a great extent, depend on growth of the ecosystem around it – the ecosystem of hardware makers, software developers and customers. Feldman expects this ecosystem to grow very quickly because it is so much cheaper and less time consuming to build an ARM server than it is to build an x86 one. “It takes four years and $350-400m to make an x86 server,” he says. This is why there has not been a successful startup in the x86 server market in 20 years. “On the other hand, it takes $30m and 18 months to build an ARM CPU.”


This is important because it means companies can make custom chips for big internet companies, who are expected to be the first major microserver users. “What that means is we can do a custom one for Facebook; we can do a custom one for Amazon; we can do a custom one for Baidu – each individually and separately – that has their IP in it, that embed advantages [for] their particular technology,” Feldman says.


He expects ARM to win not because the end products will have slight advantages on key metrics. “They will,” he reassures us. “They win because it’s a fundamentally different approach to delivering a product.” ARM Holdings amortizes its R&D over its 250-300 licensees, while Intel amortizes its R&D over itself, and that is what makes it so much fun to watch, Feldman says. “It’s not a battle of two heavyweights [who are] just going to stand next to each other, hitting each other in the head. It’s a battle of a heavyweight versus 40 little Irish guys hitting him on the ankles.”

The reasons the big internet companies will be the first to adopt microservers, he explains, are that their businesses are extremely sensitive to cost of computing, and that they write their own software and have no allegiance to an instruction set. “They’re writing in PHP. They absolutely don’t give a shit about special instructions or optimizing. They’re writing at a layer of abstraction so high, that they don’t even know what the CPU’s doing.”


That is the opposite of the enterprise market, where one virtualized server is running 30 or so software packages on 30 different virtual machines. This is an extremely conservative market, where IT is expensive by design, and it can take as long as five years for the concept of microservers to take hold in this space. We have already seen examples of technologies developed by the likes of Google or Yahoo! – Apache Hadoop is a good example – making their way into the enterprise data center a few years later.


The big internet companies with massive data centers “want an ecosystem’s worth of alternatives,” and that ecosystem is bound to develop very quickly, Feldman says. “Right now, you’ve got fundamental changes. You’ve got demand going through the roof on one side. You’ve got customers being smart and innovative on the other. You’ve got a threat from the ecosystem against an installed giant. A lot of things happening, and that’s cool.”


This article first appeared in the 32nd edition of DatacenterDynamics FOCUSmagazine. Follow the link for a free subscription.


Sign in

Forgotten Password?

Create MyDCD account


region LATAM y España North America Europe Em Português Middle East Africa Asia Pacific

Whitepapers View All