Processor Whispers: Mine, Yours and Ours

by Andreas Stiller

After a vague announcement at the SC09 supercomputing conference, Intel has now let the cat out of Schrödinger’s bag: the planned graphics processor Larrabee is dead and alive at the same time.

Nick Knupffer, the PR Manager responsible for high performance computing (HPC), officially confirmed that the Larrabee is not – as was originally planned – going to be released as a standalone product as a GPU on a PCI express card, like Radeon or GeForce, in the first quarter of 2010. It will presumably appear sometime later, but at first only for the HPC market, not for the consumer market.

Still, the already fully developed cards are not going to be pulped, but used as development platform for the subsequent Larrabee generations.

Data Bottleneck

The memory concept for the Larrabee 2 is called M-Y-O: mine, yours, ours. With a direct connection to the CPU and common memory areas, it is supposed to circumvent the problematic data transfer bottleneck. During his speech at the SC09, chief technology officer Justin Rattner not only made a Larrabee prototype peak at over 1 teraflops (SGEMM, if only with small 40x40 matrixes) but also pointed out the weak spot of current GPGPU designs: the data transport between CPU and GPU memory. No matter how fast they run, under actual operating conditions this bottleneck will mostly frustrate all performance efforts. However, according to Rattner, Intel has something far better up its sleeve, a common virtual memory system “M-Y-O” (Mine-Yours-Ours) that covers separate as well as shared address spaces. Both can then access the same data structures, which greatly simplifies the programming and widely circumvents the bottleneck issue. But it seems that Rattner had in mind the Larrabee generation for HPC after the next one, which will presumably be connected directly via QuickPath Interface (QPI), because with the initially planned first-generation Larrabee graphics cards the PCI express bus acts as a roadblock regardless of any Mine-Yours-Ours systems.

Another slow-down for Intel’s ambitions within the graphics sector may very likely have been that it’s harder than expected to convince the game developers to throw their habitual programming models overboard and adjust to the new Larrabee model. Apparently, Intel hasn’t been able to bring even one of the big gaming console manufacturers on board either. Something fuzzy came along the grapevine about Sony and the PS4 a while ago, but now Sony is much more interested in the Power7 as a potential successor for the Cell.

Maybe now, after the settlement with AMD – the 1.25 billion dollars of conciliation money have just been transferred – and the related access to all ATI patents, Intel has new options for improved graphics concepts, which it can now integrate unhurriedly. On the other hand, AMD has time to recover; after all, until now, it’s the only manufacturer to deliver Direct X11 enabled graphics cards.

Like Intel, AMD is eagerly working on the integration of CPU and GPU on a single chip (Fusion) and – in the context of the “Accelerated Computing Initiative” – on a connection to powerful discrete GPUs that’s as efficient as possible. The next generation of GPUs “Northern Islands” is supposed to partner up with the Zambesi processor and roll out in 2011, manufactured like the CPUs, in GlobalFoundries’s 32-nm process technology.

Lacking CPUs

Nvidia got the short end of the stick as it doesn’t have a proprietary x86 CPU for a direct connection. So, it seems that Nvidia will somehow have to come to an arrangement with its only two-miles-distant neighbour in Santa Clara. Unfortunately, there’s still some quarreling going on, first of all concerning the QuickPath license for connected chipsets. In spite of the dispute with Intel, Nvidia wants to continue delivering the chipset “nForce MCP89” for Nehalem CPUs, but it doesn’t dare to dabble with the Direct Media Interface that will be a feature of upcoming desktop and notebook processors like the Arrandale and the Clarkdale. Also Nvidia had complained about unfair competitive conduct by its competitor and was recently interviewed by the Federal Trade Commission (FTC) to this end; even after the settlement with AMD, the commission continues its investigations against Intel.

After the news about Intel’s new Larrabee course, rumours ran wild that Intel now plans to acquire Nvidia. Fittingly, Nvidia mocks its competitor on its website www.intelsinsides.com, created exclusively to make fun of Intel, with a comic, in which the sad Intel child reveals its most desired wish to Santa Claus: a GPU ...

At least, according to Charly Demerijan of semiaccurate.com, Nvidia is now close to completing the A3 step with wafers of the greatly delayed next Nvidia GPU generation “Fermi”. If everything goes well, he says, Nvidia could present the first A3 prototypes at CES in January. However, GeForce cards won’t hit the market until some weeks later.

Spring also promises a variety of other interesting processors; this was revealed by Intel’s unofficial roadmap that somehow made its way to channelweb.com and PConline.com. Later on, in March, Intel plans to present the new 32-nm Westmere processors with six cores, the Gulftown for desktop PCs – which will now be labeled Core i7-980X instead of Core i9 – and the Westmere-EP (Xeon X/L/E56xx) for servers and workstations, which is supposed to roll out with up to 3.46GHz on the 16^th of March. But maybe Apple will have been shown consideration again and come up with the new Xeons in the Mac Pro just two weeks earlier, right on time for the start of the CeBIT fair, like last year with the Nehalems. So the trade fair company would be well advised to lure Apple back to CeBIT with very attractive conditions; after all, partner T-Mobile and the OS X Business Park in the “anti-Microsoft/open source software” hall 2 will be presenting Apple products. To round things up for the iTuners, there’ll also be the new music and IT fair called “CeBIT Sounds”.

Possibly, Intel will present the Xeon MP 76xx (Nehalem-EX) based on the 45-nm Nehalem architecture with 6 and 8 cores, which is intended for big servers, at some time during CeBIT. However, at the moment, nothing certain is known. In any case, competitor AMD plans to enter the race with its 12-cylinder Formula One module Magny-Cours just a few days after the Westmere-EP rolls out, approximately at the start of the next Formula One season – where the Ferraris probably won’t have any AMD advertising anymore.

(djwm)