Advanced Computing in the Age of AI | Friday, March 29, 2024

AMD To Unify X86, ARM Systems With SkyBridge 

Chip maker AMD wants the world to know that it is not just dabbling in ARM processors, but rather intends to make ARM chips peers to its X86 products. At an event in San Francisco today, the top brass of the company revealed the full extent of its “ambidextrous computing” strategy, which includes an effort code-named “SkyBridge” that will see future X86 and ARM processors have the same pin-outs.

AMD also announced that it has become a full ARMv8 architecture licensee of ARM Holdings, which will allow it to create its own custom ARM core designs. On the server front, thus far only Applied Micro has a similar full ARMv8 license. With decades of experience designing both low-power and high-performance processors, this is a logical step for AMD to take – and one that will allow the company to further differentiate itself from others in the ARM collective who want to peddle chips for systems, storage, and networking gear.

“Today is about innovation and how we can shape a new future for AMD,” said CEO Rory Read in opening up the event. “When I joined AMD three years ago, 95 percent of our business was centered over PCs, a potentially declining market in an unhealthy duopoly. We embarked on a strategy three years ago to create an ambidextrous architecture, to unlock our leadership IP – our IP – that can truly make a difference in the next generation of the cloud era.”

AMD CEO Rory Read

AMD CEO Rory Read

Read said by next year, 50 percent of AMD’s revenues will, by contrast, come from five growth markets: professional graphics, dense servers, embedded systems, semi-custom chips, and ultra-low-power clients.

So AMD is ramping up ARM designs and fleshing out future X86 designs in a two-pronged approach to capturing market share, explained Lisa Su, general manager of AMD’s Global Business Units. Here is the market as AMD sees it for processors:

amd-arm-x86-tam

The chart above shows the total addressable market for processors across embedded, mobile, PC, console, and server segments by architecture. The gray bars in the middle of the chart lump together Power, Sparc, MIPS, various mainframe, and other proprietary chips. Interestingly, you can see that the percentage of the market supported by X86 processors has fallen in recent years, but is expected to stabilize by AMD’s estimates, while ARM chips will see their share of the chip pie grow.

“The punchline here is that if you add up all of the X86 and ARM silicon over the next X years, that is where all of the silicon is going to be built,” said Su. “These are the dominant architectures.” And, she added, “AMD is the only company that can bridge the ARM and X86 ecosystems.”

That is precisely the plan. AMD is not just going to produce X86 and ARM chips in isolation, but is going the next and logical step in designing chips that will share a common pin-out. The SkyBridge effort to unify the processor packages – there could be common sockets for high-end systems, AMD did not say – is underway right now, and the first products that offer this plug compatibility will come to market in 2015.

AMD has just rolled out its “Beema” and “Mullins” Accelerated Processing Units, or APUs, which combine CPUs and graphics chips on the same package, based on the new “Puma” X86 cores. On the X86 side, the SkyBridge parts will be based on the next-generation “Puma+” cores and will support the AMD’s Heterogeneous System Architecture (HSA) extensions that allow for the CPU and GPU to share a unified memory space, among other hybrid computing features. The future X86 chips will come in both APU and system-on-chip (SoC) variants, and so will the ARM chips, according to Su. Both will be etched in 20 nanometer processes, she said. AMD has not identified who the foundry will be, but it uses GlobalFoundries and Taiwan Semiconductor Manufacturing Corp to make its processors and graphics chips.

amd-ambidextrous-roadmap

AMD is sampling its “Seattle” Opteron A1150 now, which is made using a 28 nanometer process by TSMC, and plans to ship it in volume in the fourth quarter. That is on target even though it is cutting it close. (No other 64-bit ARM server chip vendor is ramping up any faster. Calxeda went bust and Applied Micro is only sampling its X-Gene 1 chip to selected partners.) AMD is using the 64-bit Cortex-A57 core from ARM Holdings in the Seattle chip, and will create a low-power version of the Cortex-A57 for the first SkyBridge parts next year. This is made possible because AMD has come up with the core design methodologies and verification techniques that are required to tune an ARM chip design, based in part of tools from ARM Holdings and in part on its own experience and tooling for X86 chips.

To do such a tuning requires an architectural license from ARM Holdings, and that is precisely what AMD has already acquired behind the scenes. (The company did not say when it became a licensee, but it had to be some time ago for the company to be on its way to a modified A57 core by next year.) Further out in 2016, AMD is going to create its own ARM core, code-named K12, and implement it in future SkyBridge products.

If people have been impatient for AMD to make bold moves, it is understandable. The market needs that, as much as it needed for the Opteron to be launched in 2003. But AMD can’t afford to hurry, either. “We wanted to go about this in a thoughtful way,” explained Mark Papermaster, CTO at AMD, at the event. And while AMD is obviously very excited about doing a “from scratch, optimized design” for an ARM chip, Papermaster made it clear that AMD was “not backing off on X86.”

amd-arm-roadmap

The key point with the SkyBridge effort is not just that customers who design for a future AMD X86 processor can swap it out for a future AMD ARM processor. The unification effort also means that the network fabrics, accelerator interconnects, and other features external to the cores are implemented as reusable IP blocks and are common across the X86 and ARM designs. (IBM has taken this approach for years with elements of its Power and System z processors, by the way.) This way, AMD can leverage the efforts of its hundreds of chip engineers and its large patent portfolio on both architectures.

Jim Keller, chief cores architect at AMD who worked in Digital’s Alpha as well as AMD’s Opteron processors and recently came back to AMD after a stint at Apple, is in charge of the SkyBridge effort.

“I love processor design and I love complicated system design, and AMD was looking at taking a big swing on the next generation,” Keller said, explaining his return to the company. He also explained why AMD was not doubling up its processor teams – and therefore its expenses – as it moved into ARM chips.

“As it turns out, high performance computing is mostly about high performance feature verification, methodology, CAD, and these kinds of things,” Keller said. “The instruction set is actually important. The ARM architecture is very new, and it has some inherent efficiencies in it which we think is very cool. But in terms of the overall effort, we get so much leverage that it is not a double effort. Interestingly enough, going through the process of designing ARM has given us a bunch of new ideas which I think actually drives better core designs.”

EnterpriseAI