Tim Sweeney: The end of the GPU roadmap [pdf] | Hacker News Reader

Tim Sweeney: The end of the GPU roadmap [pdf]

graphics.cs.williams.edu

64 pointssimonb18 comments

One thing that didn't seem to be addressed in the slides themselves is the reason that GPUs have fixed function texturing operations, a fixed pipeline stage configuration, etc. It saves on memory bandwidth by making the memory accesses more coherent and dramatically magnifying the usefulness of relatively small caches.

Basically this is a claim that there's a massive amount of graphical quality that will be unleashed once we're unchained from the tyranny of the fixed function GPU behavior that remains, and that this is worth whatever loss in computational power we'll have when we increase the sizes of caches to cover the less-coherent memory access.

I'm not sure if I buy that. Games look quite nice already, and I'm not sure if I could personally tell the difference between what we see in something like Crysis and something like Toy Story, which didn't have the shackles of fixed function pipelines.

An easier explanation that's long been a pet theory of mine is that Larrabee is essentially a thread parallel FP-heavy processor for the scientific market motivated by GPGPU's encroachment on that market segment, and necessarily labeled a "CPU/GPU" in order to not step on toes in the wrong places at Intel.

It's definitely going to be an interesting next couple years in graphics, though.

I think he mostly hopes for something like REYES for games, which would improve quality somewhat (motion blur and DOF look nice), but that's probably never going to become part of the fixed-function pipeline. That's why he's calling for a "back to software" approach.

WRT Larrabee, I agree with you, and I wonder if the scientific market can recover their development costs.

Pixar has their own style that is intentionally not photorealistic. Also, Toy Story is from 1995. Try instead comparing Crysis to the effects from a modern blockbuster, which have to hold up well when compared to the live action shots.

Ditching the fixed-function pipeline isn't all about realism, though. More flexibility will allow more experimentation with non-photorealistic rendering. Maybe we'll see more games with graphics that look like they were painted or drawn.

Another benefit of more flexible rendering should be easier game development. Writing a complete renderer from scratch will be hard but most games will use middleware for that. The most expensive part of a modern game is the art, and a flexible renderer could do a lot to make generating that art easier.

I do share your pet theory about Larrabee trying to avoid stepping on toes at Intel by labeling themselves a GPU. However, in order for Larrabee to be worthwhile for Intel it needs a mass market, and graphics is it. If Larrabee fails at graphics it will likely not survive.

"Load 16-wide vector register from scalars from 16 independent memory addresses, where the addresses are stored in a vector!"

Wont that cause 16 memory reads that are much costlier than what one can benefit on a 16-way SIMD ALU operation?

What I wanted when I tried to write highly optimized computer-vision code was to have multiple "cursors" in memory that i would read from and increment. I.e. that the hardware would prefetch data and my code would operate as if data was a stream.

Your processors fetches a whole cache line at once anyways, so it isn't more memory reads. The memory address needs to be aligned to 16 byte, though.

But it says the addresses in the vector are independent. What if they don't all point into the same cache line? You could have 16 cache misses.

What have I misunderstood?

yes, you could have 16 misses, but the next n (where n=cacheelinesize/size of entry) after that will all be hits...

You have to risk that miss at some point. And it does not matter whether you see the misses one-by-one or 16 in one go, it's the same penalty.

A more interesting worst case would be one where the 16 slots take turns in missing the cache, so that almost everything is read from the cache, but there is still the penalty for a cache miss between every two SIMD instructions. But if you write your code in that style you get what you deserve :-).

Even some current architectures [edit: like, for example, x86] have instructions that give hints to the MMU that the program intends to stream some part of the memory in the near future, I guess these will rise in importance when more programs have to avoid these kinds of problems.

That's a really great presentation!

One observation though, what is sold as a 'game' today is more of an interactive movie.

As to the larrabee, I can't wait until it comes out. That chip is going to give NVIDIA serious headaches.

Lots of good stuff in there, I just skimmed it will go back later to read it again.

Larrabee seems very Itanium-like to me... And something I'd never really considered, but seems like GPUs actually take a lot from the Itanium camp.

Perhaps they were on the right track, but simply too early (afaik one of the root-cause problems of Itanium was the die size).

Itanium?!

Itanium is not a many-core processor: it is a VLIW processor.

If the Larrabee can remind someone of another processor, it's the Niagara family from Sun.

I read "itanium-like" not in terms of architecture, but in market timing, delivery on performance promises, and scalability for future revisions.

I think Larrabee is a lot more promising than Itanium ever was. At the bare minimum, it's a server-like x86 processor.

I would prefer something with a more elegant ISA than an x86, but that's what we have in front of us now, thanks to the huge mass of non-portable x86 code that's so critical for success on the desktop...

That depends on what you mean by "interactive movie". I've seen some games that rely too heavily on cutscenes, and other games that keep the character pretty strictly on rails (e.g. Call of Duty and Half-Life series), and I generally detest both of those.

I like to get into high tactics, if not strategy, in the FPSes I play, which means I like games that have the scope and openness for emergent action. Typical examples would be Far Cry and Far Cry 2 (but not Crysis, being spoiled by the suit's power). These aren't really what I'd call interactive movies, because even though they have plot to keep the character moving along, the main meat of the game is in how you tackle "action bubbles" - but those action bubbles are nice and broad that there's plenty of scope for deception, diversion, distraction, improvisation with explosives in the environment, and an openness to lots of different approaches that linear games like Half-Life (particularly 2 and episodes) can never really match.

The epic gaming guys always have the most interesting presentations, they seem more up on more current software techniques like STM and functional programming, I don't hear a lot of other game companies talk about.

It's particularly interesting to read Sweeney's views on this subject, considering that he's been talking about the shift back to software rendering since 2000 or so, when GPUs started becoming programmable.

Check out this old interview: http://archive.gamespy.com/legacy/interviews/sweeney.shtm

2006-7: CPU's become so fast and powerful that 3D hardware will be only marginally benfical for rendering relative to the limits of the human visual system, therefore 3D chips will likely be deemed a waste of silicon (and more expensive bus plumbing), so the world will transition back to software-driven rendering... If this is a case, then the 3D hardware revolution sparked by 3dfx in 1997 will prove to only be a 10-year hiatus from the natural evolution of CPU-driven rendering.

His timeline was off by a few years, but I think he basically had the right idea all along.

Note that now he's not saying that GPUs are becoming margunal, it's the fixed-function pipeline he dislikes. The "software rendering" he calls for will most likely run on a GPU.

The return to CPU-based rendering might do wonders for PC gaming's popularity. It hadn't occurred to me until I read this, but it seems like the decline of PC gaming has mapped fairly closely to the increase in reliance on top-end graphics cards. There is probably a huge slice of people at the margins that have shunned PC gaming (actively or passively) because of the various burdens that requiring a graphics card adds.

Reply on news.ycombinator.com