Yesterday’s announcement from AMD about its plans to build ARM-compatible Opteron processors is very big news, but it shouldn’t come as a shock, since AMD execs strongly hinted at this move with talk of an "ambidextrous" ISA strategy during the firm’s analyst day earlier this year.
One of the questions that lingered after yesterday afternoon’s press conference was the exact nature of the CPU cores to be used. ARM offers licensees two basic options: rights to use a pre-fab CPU core and build it into a chip, or rights to build your own CPU core that’s compatible with the ARM ISA. Most familiar ARM-based devices these days use the first option, incorporating something like the Cortex A9 into a larger SoC. A few use the latter option, licensing the ISA and building their own CPU core, including Krait, Nvidia’s secretive Project Denver, and apparently Apple’s A6.
AMD’s talk of bringing its expertise with 64-bit CPUs to the ARM ecosystem might have led you to think that it had taken the second path, licensing the ISA, but that’s not the case. Instead, AMD is licensing a 64-bit CPU core from ARM and building it into a chip—which AMD calls an SoC, or system-on-a-chip—that’s compatible with the server-oriented Freedom Fabric interconnect AMD acquired when it purchased SeaMicro. This fabric interconnect will most likely be incorporated into the silicon along with the ARM cores.
We’d expect to see the ARM-based Opterons riding on a modular server card like the one AMD showed off for its x86 Opterons at Hot Chips. These cards plug into high-density servers following the SeaMicro template, and such cards will likely be the basic unit of computing for AMD’s cloud and enterprise solutions going forward. Thus, the, uh, Opter-ARMs may not have any need for compatibility with x86 Opteron sockets or interconnect standards like HyperTransport.
These ARM cores are unlikely to match today’s Xeons and Opterons in raw performance, but they’ll probably consume relatively little power, allowing server makers to pack lots of chips into a single cabinet. The decision to target cloud computing installations, where lots of relatively low-performance cores can serve effectively, seems to fit with that profile.
Since AMD plans to bring its first ARM-based Opteron to production in 2014, the decision to license an ARM core makes sense. Going forward, AMD could potentially decide to build its own ARM ISA-compatible CPU core, but doing so would take time. Also, licensing a core saves on engineering resources, a major constraint at a cash-starved firm. AMD can pursue this avenue while still dedicating resources to its own x86-compatible Opterons, which are not going away.
AMD is unlikely to bring ARM server chips to market unopposed, given the sheer number of ARM licensees in the world. In fact, former Opteron chief Pat Patla defected to Samsung earlier this year, along with some other former AMD employees, amid rumors that Samsung might be building ARM server chips.
Meanwhile, Nvidia was briefly very open about its plans for Project Denver, which include the development of a "fully custom" ARM-compatible core targeted at desktop PCs and servers. We expect this CPU to have relatively high performance targets and to include 64-bit support. If Nvidia has done its job well, Project Denver silicon ought to outperform ARM’s licensed CPU cores. This chip will also have an Nvidia GPU on the same die, possibly making it a candidate to serve in GPU-driven supercomputing clusters, a business where Nvidia and AMD already both play, often together.
Of course, AMD has some natural advantages over such potential competitors in the ARM server space, including its own formidable GCN architecture for GPU computing, an existing Opteron business with all of the attendant relationships with enterprise OEMs, and the SeaMicro infrastructure for dense servers. What it does not have is a head start, given that Nvidia announced Project Denver nearly two years ago and has been developing its Tesla business diligently since then.
The larger question here that’s really intriguing is what the creeping competitiveness of processors based on the ARM ISA means for Intel, the x86 stalwart whose high per-chip prices and margins are being eroded on various fronts by commodity SoCs costing $25 or less. The relative openness of the ARM ISA, combined with the constantly plunging costs predicted by Moore’s Law, may be a bigger threat to Intel’s business than any challenge that AMD ever mounted with K7 and K8. Why? Because this time, everybody from Apple to Qualcomm to Nvidia is ganging up with AMD to take on the giant.