Slashdot Mirror
3D Raytracing Chip Shown at CeBIT
An anonymous reader submits "As noted at heise.de Saarland University is showing a prototype of a 3D Raytracing Card at CeBIT2005. The FPGA is clocked at 90 MHz and is 3-5 times faster in raytracing then a Pentium4 CPU with 30 times more MHz.
Besides game engines using raytracing there was a scene of a Boeing with 350 million polygons rendered in realtime."
So, the question is: Can these guys get ATI or nVidia to buy their chip?
They are trying (surprise, surprise).
From their site (already melted (yes, yes, mirrordot)): We are very much interested in evaluating new ways for computer games and therefore like to cooperate with the gaming industry. Thus if you are in such a position, please send us an email!
CC.
TaijiQuan (Huang, 5 loosenings)
It's doubtful that people from ATi or nVidia would have anything more than a casual interest in such a chip. More than anything with the way GPU's have been heading, this device would be more of a backwards step than a forwards one. GPU's these days are far more flexible than they used to be, and there's every indication that trend will only continue, allowing developers to do what they want with the hardware, rather than being told what they can do with it and having no real choice either way.
So to sum up, don't expect to see vastly specialised GPU's for raytraycing hitting the market, at least not for the mainstream buyer. It's more likely that we'll see GPU's become more generalised to the point where raytracing can implemented on software. Will they be as fast as a purpose built chip like this? No, more than likely they won't. Will developers be able to do a whole lot more with them? Most definitely, and though that will come at a significant performance penalty for the moment, I think it's the right trade off to make as we should see far more creative uses of hardware put into practice, such as work being done already to use GPU's for something other than Graphics Processing.
There's a standard model of a Boeing aircraft (777, IIRC) that's used as a something of a test scene in the computer graphics community. It's about 350M triangles (everything down to nuts and bolts, but modified so as not to give any trade secrets away) and over 4GB of data, so it gets used a lot for testing how the performance of an algorithm scales to large datasets.
Unfortunately the price is an order of magnitude (or two.. or three) too high for FPGAs to really be a consumer tech. The issue I think is an ASIC costs so little in volume, rather than spend all the money on an FPGA design that might be obsoleted next year anyway - a vendor is more likely to commit a design to silicon and then sell that.
:-) ) if you're interested in this - there is open source hardware out there, some really good designs at that.
There's also the speed issue - I've spent DAYS of CPU time to get a design syntheized from VHDL for a moderately complicated IC built up from available cores.
Factor in optimizing floorplans and the like, and you're talking about serious time commitments to optimize the hardware.
It works; I've been paid to do it in the past; but it's not something I can see in the consumer market for the time being.
An exciting hybrid is intersting though, putting silicon CPU cores on the same die with an FPGA. They've been around for awhile, and I haven't done any FPGA projects in ~18 months - but I haven't seen any real movement outside of areas where FPGAs are already popular.
See Open Cores (no, not sores..
..don't panic
amd only "recently" implemented sse? they've had it since 2001.
more recent amd chips have sse2, and sse3 on amd64 is just round the corner.
They have hierarchical structures to test for which group of triagles a ray may intersect and scales more like O(n log n).
I assume you're talking about kd-trees... these do indeed offer very nice performance characteristics, but they're designed for static geometry. Efficient raytracing for dynamic geometry (moving or deforming objects) is AFAIK still far from "solved".
If you add this to the fact that raytracing lets you have perfectly smooth non-polygonized objects
and take away the fact that they don't particularly like the arbitrary triangle meshes that make up the vast majority of real datasets...
Flexible and robust realistic reflection and refraction
Yes, "Flexible and robust" is the killer. And not just for refraction/reflection; there's still no fully-general, clean, robust method of shadowing for rasterizers, and it's not for want of trying. Radiosity is a joke. Attempts to get realism out of current rasterization approaches are bodges piled on kludges piled on hacks. It became clear some time ago that the technology was heading up a dead end. Of course, so much has been invested in making that dead-end fast that it's going to be hard to take the performance hit of moving to a better but less optimized approach.
I suspect we'll eventually end up with a hybrid, rather like current deferred-shading techniques. It'll be interesting to watch it all pan out.