Larrabee Team Is Focused On Rasterization

Vigile writes "Tom Forsyth, a well respected developer inside Intel's Larrabee project, has spoken to dispel rumors that the Larrabee architecture is ignoring rasterization, and in fact claims that the new GPU will perform very well with current DirectX and OpenGL titles. The recent debate between rasterization and ray tracing in the world of PC games has really been culminating around the pending arrival of Intel's discrete Larrabee GPU technology. Game industry luminaries like John Carmack, Tim Sweeney and Cevat Yerli have chimed in on the discussion saying that ray tracing being accepted as the primary rendering method for games is unlikely in the next five years."

Comments

*Sigh*

[dreamchaser]

Intel has been saying with each and every iteration of graphics hardware that it's created that it would be 'competetive'. None have been except at the very, very low end. I like Intel's CPU's quite a bit, but I have heard the boy who cried wolf too many times from them with regards to GPU's to take them very seriously at this point.

Re:*Sigh*

[QuantumRiff]

One would think that a company that could do a complete turn around after it got its 64bit ass handed to it (thanks AMD) would be able to dedicate just a bit of brain-power to their graphics.

Re:*Sigh*

[Anonymous Coward]

Whether you take them seriously or not, this is a serious effort to be a major player in the discrete graphics market. (a market not likely to disappear soon as some seem to think)

I happen to know a great many people that work at Intel. And I just happen to also do product testing and marketing focus groups for them. All centered around gaming.

This was a topic that intel did not take seriously 5-10 years ago. They take it deadly serious now.

I spoke with paul otellini on one occasion on the topic of intel gaming. It went more or less like this.

Paul- Which Intel chip do you have in your machine at home?
Me- It's an AMD actually.
Paul- You work for Intel, your family works here and you buy an AMD?
Me- I run what gives me the highest performance in what I do. It also happened to be cheaper, but thats secondary.
Paul- They only beat us in gaming! Our chips are better at EVERYTHING else.
Me- Gaming leads the market.
Paul- No it doesn't.
Me- No one upgrades twice a year to keep up with MS office. We upgrade to keep up with Carmack.
Paul- If I offered to give you a couple of our next gen processors, would you use them?
Me- I'd try them out, but if they can't beat my current machine I won't use them. Even if they are free. Neither will anyone I know. We literally spend a couple thousand dollars a year keeping our machines state of the art so we can squeeze an extra frame per second out of our systems. We aren't going to use anything that isn't the best.

You want me and my market segment to take you seriously? Take us seriously. We make up a small segment, but we are fanatical.

___
A couple years later, I got an email from him.
It was actually sent as a response to several key divisions in intel, because several people had asked why we (intel) care about gamers, they make up less than 5% of the PC market (it's actually closer to 1%).
___
Paul- We care about gamers because gamers grow up. They grow up to work mainly in IT fields. The gamers from 5-10 years ago are now the IT professionals we most want to be on our side. They are the ones making purchasing decisions and recommendations and they do so based on what they know. They know AMD better than us because we ignored them for so long.

Why do we care about games? We don't. We care about the people playing them and we want them to identify with our products.

____

So now you have some insight as to where intel thinks this is all going. It's not that they care about gaming or graphics, because they really don't. They care about the people behind it, and getting them hooked into a brand that "supports" them.
Then there is the really obvious reasons for Intel getting into graphics, VISTA, and other next gen OS's and GUI's are going to use a lot of hardware acceleration. Which means discrete graphics cards aren't for the desktop anymore, they are for the server and the workstation too.
Add to that using the GPU to do certain types of parallel processing at much better thru-put than you can get from a CPU.

The motivation should be obvious.

*Posted AC for my sake. I like my contacts at Intel. I'm hoping Paul doesn't remember talking to a PFY about his companies gaming culture.

Re:*Sigh*

[Anonymous Coward]

Geforce 9800 (Finished and can be bought now)
shader cores 128
clock 1.7Ghz
128x1.7 = 217.6 Gflops
70.4GB Bandwidth

Larrabee (Not released until Q1 2009)
16-24 cores
Clock 1.7 to 2.5Ghz
2.5*24 = 60 GFlops
DDR3 Memory bandwith far less (Even faster DDR3-1600 has 12.800 GB/s speed)

This shows the Larrabee is at least 3.6 times slower processing speed. Plus memory bandwidth is around 6 times slower. Plus the Geforce 9800 isn't even the fastest. The GeForce 9800 GX2 is nearly twice as fast and available now.

Plus its also important to note that Next generation of NVidia cards will be out around July.

Geforce G200 (Expected in July this year, not in over 6 months time like Larrabee)
* 512bit bus so double width of Geforce 9800
* 1Billion+ transistors so very likely to be more than 128 cores
* Clock speed undetermined but its 55nm so its going to be around at least Geforce 9800 speeds.
* Plus GDDR5 will give double speed of GDDR3

But even assuming G200 isn't clocked faster plus doesn't have more than 128 shader cores, once we have the GDDR5 version, its going to be at least 4 times faster than Geforce 9800. It could end up being nearer 5 times faster.

This means we are going to get around 16-20 times the processing speed and at least 12 times the memory speed from a Geforce G200 compared with the highest estimates of Larrabee. And this assumes against a single graphics card. A Duel card will be double this speed.

Plus ATI are going to match these speeds. Intel however are showing they are laging far behind.

So now tell me which top end gamers are going to buy. Intel, no way. Still will be low end.

Intel are still not considering gamer requirements for the future, which they have to, when designing new hardware expected to be used in years to come. Its all Intel marketing hype and talk about thinking of gamers. They are simply trying to limit the effects of GPGPU, which has shown up the X86 architecture as slow by comparision, plus bloated down by lagacy features. Larrabee hopes to clean away some of the legacy features which is a good thing, but for gamers, Larrabee is too little too late at this stage. They really need to up their product spec, if they want to be taken seriously.

Re:Stupid debate

[pclminion]

Then why did you call these graphic engine experts "old men" as if they are just set in their ways? It sounds to me like they know just what they are talking about.

It's been my experience in the software world that people are incredibly stubborn about dropping old, familiar technology when something better comes along. It's certainly not limited to these folks. But even the smartest people get blinded by the familiarity of their ways.

Re:Stupid debate

[ardor]

WHEN (and only when) the technology is fast enough for real time recursive ray tracing, it will be the end of rasterization in 3D applications. Oh yes, the brute force solution. It will be a very long while until it is fully obsolete. Expect hybrids to stay for a long time. An example: many people claim terrain culling methods to be fully obsolete nowadays. Then they try to render really large terrains...

Also, given that hybrids are a no-brainer, I bet both pure raytracers and rasterizers will be extinct in games.

Cache coherency problems can be fixed by making an enormous cache, or simply making the RAM itself so damn fast it doesn't matter anymore. Ehrm .... got any other wishes?! You do realize that this would be a revolution and might not be possible? One of the reasons why the cache is fast is that the signal propagation delay can be much lower due to the proximity of the cache to the cpu. In other words, there will always be a cache. As for an enormous cache: cache mem is very expensive, is likely to be because cache performance doesnt stand still (read: it will always use more expensive special hardware), and huge caches have issues with cache misses.

Adaptive subdivision of pixels for antialiasing is not exactly a first year student problem but not enormously difficult either. I didn't say its enormously difficult, just not nearly as trivial as with rasterization.

Honestly, I want to see the technology blow right past raytracing and go straight to radiosity 1. Radiosity is not the ultimate. Just try doing specular stuff with radiosity.
2. Algorithmic complexity will always come back to haunt you. O(nÂ) will always be worse than O(n), unless you have small scenes. So you have your geforce19000 and can render ... ONE room with realtime radiosity! Nice! Somebody else fakes it and renders an entire city. Guess what will be chosen for games.
3. You could have said path tracing or photon mapping at least.

Finally, these people don't particularly favor raytracing simply because it does not pay off for games. Games usually don't feature fully shiny scenes, games are expected to run at interactive framerates. In, say, 5 years, entirely new (and demanding) effects are en vogue; if raytracing steals too much time, it will be dropped, its results faked. This is what the "old men" do all the time in their games: fake. In the offline world, things are wildly different, so don't compare them.

Larrabee will be awesome ... for everybody!

[Joce640k]

Read the article - Larrabee is designed for general purpose programmability.

If your motherboard has Larrabee you could use it for the physics calculation while your add-in GPU does the graphics.

This makes a whole lot more sense than trying to get a single GPU to do both tasks.