Pipelines – different things to different folks

Posted: 09.19.05

My ears are ringing from all the pipelines that have been piped into them this week. Nvidia has one in their 6150, two in their 6200, and 24 in their 7800, while ATI has six in their 800, and 16 in their 1800 (relying purely on early leaks on the web about Ruby's new r520).

Then in the Southeast we saw the other pipelines go dry as refineries shut down before they were blown over. It's been a tough couple of weeks for the suppliers of pipelines.

In GPUland we're about to enter into the great debate about what is and what isn't a pipeline, or what is a good pipeline or a great pipeline. It will be a discussion about fragments, and memory I/O, and floating points vs. vectors. All that noise will be unleashed when ATI officially announces the r520 and explains why it's the real deal and Nvidia's award winning 7800 is just good enough.

The news coming out of Taiwan via Digitimes, combined with IDC's upbeat forecast for the Pacific region, is that Q4 is going to be great for PCs and GPUs. Well, I couldn't be happier. That, combined with the huge announcement of the Xbox 360, is going to leave most of the consumers of pixels somewhere between stunned and bedridden. And when Microsoft starts to explain how and why their unified shader is more better'er than Sony's conventional function-specific shaders, you won't know a pipeline from a pipe cleaner.

And it's not going to let up, because the next crusade will be HDR and the right and wrong way to do that. Is HDR done in the shader? Where? Is that part of the pipeline? When does a shader stop being a ÒshaderÓ and become a FPP and not a pipeline? If you saw a pipeline would you recognize it?

I don't have an answer to these riddles because they are questions that can't be answered with technical agreement. And that's because they have been created by marketing and not engineers. So marketing types will create new and exciting names for things that process pixels and vertexes and do memory management and perform software operations that polish pixels, and we'll be expected to remember them, and understand them. We don't really have to understand them, we just have to be able to repeat what ATI and Nvidia say in their white papers, so other people will think we understand them. And to help us, the words will be either all capital letters, or a combination of capital letters. As a matter of fact, I was thinking about writing one of these editorials in all caps, but Kathleen and Marilyn threatened to hide all my games if I did.

Nonetheless, we need a way to discuss the parallelism in today's GPUs and a way to relate implied performance. But we're at risk of falling into the same trap that Intel did with the MHz game—at a certain point it doesn't reveal enough about the real performance, even though more is needed. No doubt we will see GPUs with 256 "pipes" in them, and it won't be that far away. But when we cross the 32-pipe threshold (next year), the scaling of performance will shift to other factors and they will be more difficult to turn into simple sound bites.

The next stage of graphics performance will be measured in compiler efficiency. As operating systems like Apple's System 10 and Microsoft's Vista bring more of the GPU to the forefront and simultaneously steal some of the differentiation away from the GPUs like memory management, the GPU sup-pliers will have to find other parameters to distinguish their parts. Even benchmarks won't be enough now that -Nvidia has been successful in downgrading the importance of 3Dmark and we now score dozens of FPS counters in each game (and when will that meaningless game end?).

And while we're struggling to distinguish and praise the newest crop of GPUs, the little supercomputers on a chip that are a marvel to contemplate, their little brothers the IGPs are growing up and sporting pipelines, too.

This is terrific fun and what gets me out of bed and onto too many airplanes every day. It's exciting and it never stops. But we are at risk of getting too far ahead of the people with the money—those consumers. If the games can't keep up and take advantage of these new multi-dozen-pipelined GPUs, and the purveyors of them can't communicate their value in simple terms, the consumers will turn off. And its damn hard to lead a parade when you're the only one in it, no matter how many pipe organs you've got.