In a story on PC Pro, Nvidia architect John Montrym (whose name was incorrectly spelled "Mottram") quoted my recent blog post on Larrabee as concluding that "the 'large' Larrabee in 2010 will have roughly the same performance as a 2006 GPU from Nvidia or ATI."
Alas, this isn't really what I said or meant.
What I actually described as equating to "the performance of a 2006-vintage...graphics chip" was a performance standard defined by Intel itself--running the game F.E.A.R. at 60 fps in 1,600 x 1,200-pixel resolution with four-sample antialiasing.
Intel used this figure for some comparisons of rendering performance. If Larrabee ran at 1GHz, for example, Intel's figures show that it would take somewhere from 7 to 25 Larrabee cores to reach that 60Hz frame rate.
Larrabee will probably run much faster than that, at least on desktop variants. Let's consider the most commonly mentioned clock-frequency target in public reports: 3GHz. Also, let's assume that Larrabee performance scales linearly with frequency and core count.
So that means we're probably looking at needing about eight Larrabee cores to render F.E.A.R. at 60Hz.
THAT would be performance comparable to a 2006 GPU, and there will probably be Larrabee variants--especially for CPU or chipset integration--with that level of performance.
But discrete Larrabee graphics chips may have more cores. Published estimates suggest 32 cores, and that would be much better than a 2006 GPU. At least on F.E.A.R, a 32-core Larrabee could be faster than Nvidia's current GTX280.
F.E.A.R. is a DX9 game, of course, and the GTX280 was not optimized for those older games. On a game like Crysis, I suspect the GTX280 would display relatively much better performance. Nevertheless, a big Larrabee could be pretty fast.
But at what cost?
A 32-core Larrabee running at 3GHz would probably make the GTX280 look inexpensive and power-efficient. If the rumors I heard at Siggraph are correct, Larrabee's made on Intel CPU fabs, one-half to one generation behind the company's leading processors. That probably means 65nm right now, maybe 45nm by the time Larrabee ships.
Good technology, sure, but Larrabee isn't built like Intel CPUs and doesn't work like Intel CPUs. Intel only gets four to six CPU cores on a 45nm chip today. I'd estimate Larrabee's cores are a third to half the size of a Core 2 Duo core, so a 32-core Larrabee could be 3 to 4 times the size of an Intel quad-core processor (not counting the caches)--call it 400 to 600 mm^2.
Also unlike a CPU, when running hard, most of Larrabee is running at the same time. Intel CPUs rarely experience this condition--they have a lot of execution units and big caches, and no real-world workloads can beat on everything at once. That's why CPU vendors have to create custom-written thermal-virus programs to test maximum power consumption.
Larrabee, on the other hand, gets most of its performance from the wide vector units, and these are relatively easy to keep busy.
So I think a 32-core Larrabee in 45nm could draw 300W or more for real-world workloads, and that's just the chip--the local memory and any other needed circuitry will draw more power. Since this kind of power consumption is probably impossible to support in a real graphics card, something will have to give. The number of cores or the operating frequency could be lower, the overall efficiency could be lower than implied by the Larrabee disclosures to date, or perhaps the 32-core variant will just have to wait for 32nm process technology--putting it yet another year behind AMD/ATI and Nvidia.
Bottom line, I feel like we need to wait for Intel to give us better performance estimates before we start making specific comparisons with competing GPUs, but I don't expect Larrabee will be fully competitive on power or performance for the price.
(This post isn't the big Larrabee followup I've been planning, which will go into many other details of the chip's design and operation based on information I've gathered at Siggraph 2008 and the Intel Developer Forum. I'm waiting on that until after Hot Chips and NVISION 08 this week, where I'm sure I'll learn even more about this forthcoming chip.)