Slashdot Mirror
With 8 Cards, Wolfenstein Ray Traced 7.7x Faster
An anonymous reader writes "As Slashdot readers know Intel's research project on ray tracing for games has recently been shown at 1080p, using eight Knights Ferry cards with Intel's 'Many Integrated Core' architecture. Now a white paper goes into more detail, documenting near-linear scaling for the cloud setup with 8 cards, and gives details on the implementation of 'smart anti-aliasing.' It will be interesting to see how many instances of Intel's next MIC iteration — dubbed Knights Corner, with 50+ cores — will be required for the same workload."
so, how does it stack up against a Riva TNT2?
(ducks)
Not to mention that those 8 cards probably don't have as much raw pixel pushing power as a pair of SLI or CF cards from the other guys, really not that impressive. Frankly Intel should have tried to buy out Nvidia years ago, the thing that gimps their chips more than anything is the crappy IGPs, instead they shoot themselves in the face by killing the Nvidia chipset business thus making Atom completely worthless. With ION it was actually a nice little unit but with Intel IGPs Atom is horrible.
That is one place you really have to give AMD credit, they saw the direction the market was going with low power mobile devices and bought ATI which makes excellent chips and ended up with Brazos, a dual core with HD graphics that sucks less than 18w under full load. I mean sure 8 cards might be nice for a render farm, but how many are gonna be willing to pay that electric and cooling bill with the prices constantly going up?
While I appreciate the effort Intel's answer to graphics has never been very good and if they need to throw 8 fricking cards at it I don't even want to know how many watts this sucker is pulling. The future is obviously mobile and while Nvidia has tegra, AMD has brazos, what's Intel got that can compete with that level of graphics in that small a wattage envelope?
ACs don't waste your time replying, your posts are never seen by me.
Welcome to the current generation of ray traced game engine demos. They're poor looking. Apart from the metal sphere floating above the water and the odd glass refraction effect.
You'd think that they could hack something impressive looking together, particularly as they are competing against whatever it is the next generation of polygon rendering will come up with.
Michael Reed, freelance tech writer.
Spot on! ION was the best thing that ever happened to the Atom platform. Really, it was the only thing that made it into a usable HTPC or ultra-low-power desktop. They really need to stop shitting down NVidia's throat because they are precisely the kind of aggressive, performance-driven company that would fit alongside Intel's model.
-Billco, Fnarg.com
Note that these are raytracing cards, not rendering. Raytracing is a very different technique which can do cool effects like refraction through glass (shown in the chandeliers and scopes), jawdropping water, and realistic lighting effects that rendering cards simply cannot do.
It's also much more demanding on hardware. One of the big drawbacks is it requires a lot of scattered reads out of memory making caching much less effective. You need tons of bandwidth to low latency memory to make it happen. We're still a very long ways out from having this possible in reasonably-priced consumer GPUs.
Rag on Intel for their integrated graphics if you want (though I consider them a good non-gaming graphics chip with very good open source support), but these cards are not related to those in any way. These are full-featured x86/x64 processors with 32 cores per die. In other words, they created a 256-core system capable of software-raytracing the whole thing at high resolutions.
That is quite an accomplishment, and rest assured, it is top-tier performance in the raytracing world. This isn't meant to be a practical gaming system; this is pretty clearly being done by Intel to show off the benefits of their many-cores processors, and it is an impressive show.
To the GP: They're using Wolfenstein because it's one of very few games that has a ray-traced variant, and it exists only because Intel created it as a testbed. More on that here: https://en.wikipedia.org/wiki/Wolfenstein:_Ray_Traced
The problem with these demos is, they use ray tracing like we did in 1980 (i.e. Whitted style). All computations are highly coherent and efficient. As soon as you want to have more natural rendering, with diffuse illumination etc. Parellization doesn't scale proportionally anymore. Rays become heavily incoherent, memory access scatters and you get cache misses etc. So the real feat would have been if tey show 7.7x speed with diffuse global illumination.
[--- PGP key and more on http://www.root42.de ---]
Apparently this is the newer wolfenstein games; I wanted to see what 8 GPUs worth of fancy effects could do to the original pre-Doom Wolfenstein :(
I mod down anyone who says "I will be modded down for this", regardless of the rest of their comment
Rent the Pixar Shorts DVD and watch what they did before Toy Story 1. Red's Dream, Tin Toy and Andre'B are all pretty crappy looking short films that demonstrate what a couple of guys in a lab could do at the time - when viewing them, you're supposed to imagine what could be done by a larger studio with funding, not bash them because a couple of guys in a lab given 3-6 months aren't producing something competitive with hundreds of people given millions of dollars and a couple of years.
I believe you're mistaken. Raytracing IS the technique where you're tracing light much the way it happens in the real world. The techniques usually used in GPUs are quite backward. It hasn't really been all that downhill, though; they've gotten pretty good at faking a lot of the effects, but when it comes to things like shadows, local lighting, radiosity, and refraction, Raytracing is where it's at.
Examples:
https://upload.wikimedia.org/wikipedia/commons/e/ec/Glasses_800_edit.png
http://hof.povray.org/images/chado_Big.jpg
http://hof.povray.org/micra1_09.html
http://hof.povray.org/images/warm_up_Big.jpg
http://hof.povray.org/images/kitchen.jpg
All of those are from POV-Ray. There are plenty more in their gallery over here:
http://hof.povray.org/
Feel free to send some counterexamples of other techniques doing it better.
Wrong. Nearly the entire raster pipeline would be ignored for ray tracing, and you don't really need a lot of shading units for the rest (don't need to multitexture in the background and whatnot). The main use for the GPUs in ray tracing would be collision detection, which could be written into shaders as long as the entire scene was loaded into each GPU's memory, so Wolfenstein is actually a good choice - a large scene would have problems because of memory constraints. Ray tracing works very well with lots of parallel CPUs, but usually is memory constrained (dependent on memory access more than anything else) in that scenario, so splitting it off onto multiple GPUs is a way to remove that constraint, but basically it still works like a lot of parallel CPUs accessing the same scene in memory.
It's also much more demanding on hardware. One of the big drawbacks is it requires a lot of scattered reads out of memory making caching much less effective. You need tons of bandwidth to low latency memory to make it happen. We're still a very long ways out from having this possible in reasonably-priced consumer GPUs.
Yes, it is exactly what Intel Mic card are awesome for. They are generic x86 core with 4-way SMT and a buttload of memory bandwidth. I worked with Knight Ferry prototypes and studied the scalability of the worst case of algorithms for scattered memory access: graph algorithms. (The paper will be published soon but the preprint is available at http://bmi.osu.edu/hpc/papers/Saule12-MTAAP.pdf .) Basically, we achieve close to optimal scalability on most of our tests.
These MIC card are designed to scale in good cases (compact memory and SIMDizable operations such as dense matrix vector multiplication, or image processing) but almost in the bad cases (lots of indirections, accessing caches lines in pathological scenarios such as sparse matrix vector multiplication, graph algorithms.)
I am excited to get a hold on the commercial card (we worked on prototypes) to make a CPU/GPU/MIC comparison.