Original URL: https://www.theregister.com/2013/01/23/amd_q4_2012_server_drilldown/

AMD to get dense about servers – but in a good way

Hires two more chip hotshots to drive down the roadmaps

By Timothy Prickett Morgan

Posted in Systems, 23rd January 2013 23:59 GMT

Maybe SeaMicro, the upstart maker of low-power microservers that dragged Intel kicking and screaming into the market, should have bought AMD instead of the other way around. Or maybe after the AMD transformation is all done some years hence, it will look like that is what happened anyway so the difference will be moot.

There's no question that AMD is not having an easy time right now, as its latest financial results show. After having to renegotiate its wafer supply agreement with fab partner GlobalFoundries, the company posted a $473m net loss compared to the year-ago period's gain of $71m as revenues dropped 31.4 per cent to $1.16bn. If you ignore that renegotiation, AMD still lost $55m, and you can't really ignore it anyway.

The problems AMD faces are complex, and so are its answers to them.

A resurgent Intel in the data center has hurt AMD's Opterons since early 2009, and many server makers either abandoned AMD entirely or only give it minimal support in their servers. HP, Dell, and Super Micro are probably AMD's biggest friends, and the loss of Cray as a volume customer, which is shifting to Xeon E5 chips for its latest XC30 supers, will hurt AMD for years to come.

AMD has excellent GPUs, but seems to have lost its way competing against Nvidia in the fast-growing market for GPU accelerators for supercomputer workloads. The company's Fusion APU chips could turn out to be useful on this front, if AMD can marshal the software tools and partners to make it so. Thus far, Nvidia has stolen all the headlines and most of the business with its CUDA programming environment and Tesla GPU coprocessors.

The PC biz on which both Intel and AMD are dependent is hurting, and neither can get into tablets, smartphones, or embedded devices fast enough to fill in the gaps. AMD is going to deliver ARM-based Opteron processors for servers next year, but that didn't help 2012's financials and it won't help 2013's numbers, either.

So it was not much of a surprise that AMD's Computing Solutions division saw revenues fall by 37 per cent to $829m in the fourth quarter ended in December and by 20 per cent for the full year to just a hair over $4bn. And thanks to restructurings and the rejigged wafer contract with GlobalFoundries, Computing Solutions had a $231m operating loss for the year.

This is a tough business, and it is going to get a lot tougher as ARM processors continue to ride up the smartphone and tablet waves and make their way into PCs and servers.

AMD does not break out specific revenue and shipment figures for its server processors and chipsets, but on a conference call with Wall Street analysts on Tuesday after the market closed, newly appointed CFO Devinder Kumar, said that the server business was up sequentially, helped no doubt by the rollout of the Opteron 6300 processors in early November and the Opteron 3300 and 4300s in early December.

Server chipset revenues and server CPU shipments were down, as expected, but Opteron revenues overall were up sequentially thanks to their adoption in dense hyperscale servers. And, while Kumar did not say this, Cray's many big deals for Opteron-based XE6, XK6, and XK7 supers, with many hundreds of thousands of cores shipped in the second half of the year, certainly helped a little, too.

Server and fabric interconnect maker SeaMicro, which AMD purchased for $334m back on Leap Day (February 29) in 2012, was privately held before AMD bought it and AMD doesn't talk about its business except very generally to this day.

CEO Rory Reed say that AMD "recorded significant revenue growth" for the SeaMicro servers, which is great, but you have to remember that AMD only just started shipping an Opteron-based SM15000 machine late last year using a custom Opteron 4300 part.

SeaMicro did pretty well with its initial Atom-based machines, but a year ago when it dropped a Xeon chip in the boxes, the business really took off. So while having SeaMicro on the rise is good, it is helping sell Intel chips up until now and, perhaps more importantly, AMD is competing with server maker customers who might otherwise build machines based on Opterons to attack these hyperscale workloads.

Eventually, AMD will be able to sell SeaMicro's "Freedom" interconnect and both x86 and ARM Opteron parts to server makers as raw components. But that day is not here yet, and the question that AMD has not answered to anyone's satisfaction is how it will transition SeaMicro from a server maker to a set of components for server makers.

"We will combine our extensive 64-bit design experience, X86 processor IP, and ARM processor cores with our SeaMicro Freedom fabric to continue to drive leadership as the industry transitions to dense servers," Read said on the call, saying that Q4 was the best quarter ever for dense servers.

"We believe we already have significantly more dense server customer installations than any other competitor, thus making our SeaMicro technology the most-tried dense server solution available in the industry."

That's Read talking like a server vendor, not like a parts supplier. And that's fine, so long as AMD has concluded that it wants to fight the war in the data center instead of selling bullets. And honestly, given how badly it was treated by Sun/Oracle, Fujitsu, Acer, and IBM, which dropped Opterons like hot potatoes as soon as there was a bug in Barcelona Opterons as an excuse to not have to incur server engineering costs, you can't blame AMD if it wants to control its own systems and control its own fate.

SeaMicro would have had to come to the same conclusion if it were an independent and fast-growing company, too. It would have had to license ARM processors and maybe even Opteron processors and design its own processors to mesh with its fabric. The reason is simple: Networking and computing will not be separate in future systems.

It will be interesting to see how this all plays out, but for the moment, AMD is in transition and no one, including AMD, really knows.

But the company has added some more techie brains to help sort out all of the complex issues it is facing. In a separate announcement today.

Wayne Meretzky has rejoined AMD after a long and stories career. Meretzky is a hardware engineer with experience in early Unix and supercomputers back in the early 1980s and who was also the technical lead for MacOS at Apple in the 1990s.

He was also co-author of the x86-64 spec for the Opterons at AMD, where he was made an AMD Fellow, and the he was director of hardware and software development at Power chip upstart PA Semi (the carcass of which ended up inside Apple). Meretzky took a position as corporate vice president of software development at AMD in December, as it turns out.

Charles Matar, who worked on the PowerPC chip for the Motorola portion of the PowerPC alliance (which also included IBM and Apple back in the late 1980s and early 1990s), is also joining AMD as corporate vice president of SoC development.

SoC is short for system on chip, of course, and that means Matar is going to be helping AMD form its battle plan for both x86 and ARM SoCs for all kinds of devices. Matar worked on the HAL Computer Systems clone of Sparc processors way back when and for the past decade has been vice president of engineering at Qualcomm, working on various aspects of SoC design. Matar also did a stint at AMD working on the K8 processor, better known as Opteron in the server, Athlon in the desktop, and Sempron in the laptop.

"Charles and Wayne will serve as key members of our engineering brain trust, bringing with them years of expertise in SoC design and developing 64-bit software ecosystems, respectively," said AMD CTO Mark Papermaster in a statement.

"The fact that these computing experts have returned to the company underscores AMD’s unique position and opportunity, based on differentiated IP, to take a leadership position in low-power clients and dense cloud servers." ®