While my blogging has been quite sporatic, I did have the opportunity to blog on the NVIDIA corporate blog "Inner Geek". In my post "The Collectors Edition", I showed off a part of my collection of PC graphics cards dating back to the mid-90s.

Intel just announced that Larrabee is officially, really, NOT going to be released as a discrete GPU any time soon. Or for the next 5 years. Or more.

Why, you may ask? Well, because it made a really weak GPU that would probably cost too much to manufacture and have returned bad margins. How could the mighty and wise Intel have spent over three years working on this project and not know it would suck as a GPU? Well, it could be because they were not reading my blog post of December 5, 2007 where I said:

"Oh, and when did we all decide that x86 was the most perfect instruction set for graphics? Hmmmm, it's not. The example where the x86 is the preferred instruction set for either graphics or high performance computing doesn't exist. The x86 instruction is an burden on the chip, not an advantage, for graphics. In fact, graphics instruction sets are hidden by the OpenGL and DirectX API, so the only people who care are the driver programmers and we are free to design whatever instruction set is most efficient for the task, while Intel has handcuffed themselves with a fixed instruction set. Not the smartest move for graphics. Intel's plans may work better for HPC though. Still, there are many alternatives to x86 that are actually better (shocking isn't it)!"

"Larrabee is based on a simplified x86 core with an extended SSE-like SIMD processing unit. In order to be efficient, the extended-SSE unit needs to be packed efficiently. How is Intel going to do that? The magic is all left to the software compiler, which is going have a tough time finding that much parallelism efficiently."

And that was what happened. In addition, with many coherent cores, the coherency traffic and stalls eat up a lot of bandwidth and limited scaling across cores.

My prediction back then gave Intel too much credit:
"But you heard it hear first: Larrabee will not reach the potential performance capability (by a long shot) and will not displace NVIDIA as GPU leader."
Because Larrabee never even got to compete.

Oh and I still stand behind this quote:
"Let me also ask this question: when has Intel EVER produced a high performance graphic chip? (Answer: never)"

A hard lesson learned by the i guys. But I could have saved them all those many millions of dollars. If only they'd have listened. Sigh.

As I've thought more about Larrabee, I have some additional observations:
I'd like to clarify my point about the x86 instruction set. It's not innately bad, but it doesn't progress the architecture. The x86 ISA has too few registers, an irregular instruction length, and highly irregular instruction decoding - none of which help with architectural efficiency. It's just all the work and R&D put into software for the x86 architecture that compensates for its shortcomings. But if given a clean slate, few would chose the x86 ISA for maximum performance or efficiency. My point is that x86 is largely irrelevant for a GPU. The x86 will help Intel in the HPC market.

With regard to the wide SIMD (SSE) structure - we'll have to see how effective Intel is in packing useful operations. It seems that the width was chosen in order to hit certain theoretical performance metrics, not for realistic workloads. The same ambition infected the Cell processor and we're still waiting for software to approach those highly promoted peak numbers.

I remain convinced that Intel doesn't have the formula to produce a world class GPU. I see no sign they have the specialized talents to make it succeed. At best, it will be a reasonably sized die with mediocre performance, but it will have Intel's brand name behind it. That will sell some units, but I hope Intel is ready for the realities of GPU ASPs in the mainstream of the market. At least it should be better than the existing Extreme (-ly bad) Graphics they have today.

I also realized that most PC buyers have no idea how bad Intel's integrated graphics really is. The gamer sites never test Intel graphics because they know how bad it is and it's not even worth the effort. But the mainstream consumers don't know how bad it is because NO ONE TELLS THEM! Both ATI and NVIDIA are guilty here because we still have to work with the Intel chipset group and neither company has been willing to take Intel on directly. Shame on us. It's also not a topic that consumer publications make a big point about, as many in the industry consider it common knowledge - its not new news to them. So Intel gets a free ride. I was waiting for AMD to push it's platform strategy and call out Intel graphics, but I haven't seem it yet. Maybe when its new DX10 chipset ships in Q1.

With regard to Larrabee's supposed process advantage, I find it difficult to image Intel will use the leading edge process for Larrabee. I expect the chip and its successors will be 6 months to a year behind the leading edge of the process. If that's the case, then TSMC will be very competitive. In addition, our design process is great at turning around designs quickly, Intel's process is geared more toward die optimizations with its custom circuit design. Intel will likely make Larrabee coherent so that multi-chip scaling will allow Intel to make the one die and still create multiple graphics card solutions. We tend toward multiple design solutions with one die optimized for a specific price point. AMD may also be moving to the same idea as Intel. We'll have to follow that development closely.


So, if anyone else would like to comment, please add your voice. I do moderate the comments, but I promise to publish any reasonable response.