engadget.com— Reg says this technology has the potential to kick up performance by a factor of 10 and possibly as high as 30 in some computing applications
Sep 21, 2006View in Crawl 4
I'll wait for the full announcement, but I'm hoping this will (at the least) have something to do with offloading performance for Directx10 and OpenGL processes on PCI-Express connected devices.
It's probably more along the lines of video encoding, ray tracing, simulations, rc5-72, F@H, seti.In particular, anything with a tight loop, needing minimal external data, and that fits in the cache should be (stupid fast)*(number of pipelines)
Actually, stream computing is meant for easy programming. It's essentially SIMD with an emphasis on caching. That is, the vectorsized codes are executed on the GPU, but in a manner that the data from memory is collected into a "stream" (buffer really) where data can be operated on in parallel. Look-up PeakStream and RapidMind for an idea of how this stuff is programmed; anyone with an understanding of MATLAB should be able to do it.
... but only if it needs little more than 32bit-precision floating-point math.Which, considering how badly the x86 preforms at FLOPing, would be a welcome boon indeed.
cquinndSep 21, 2006
I'll wait for the full announcement, but I'm hoping this will (at the least) have something to do with offloading performance for Directx10 and OpenGL processes on PCI-Express connected devices.
lordskywalkerSep 22, 2006
This means nothing when even now programs can't utilize the performance power of high-end processors.
Closed AccountSep 22, 2006
More like benchmarks.
vuke69Sep 22, 2006
It's probably more along the lines of video encoding, ray tracing, simulations, rc5-72, F@H, seti.In particular, anything with a tight loop, needing minimal external data, and that fits in the cache should be (stupid fast)*(number of pipelines)
Closed AccountSep 22, 2006
Actually, stream computing is meant for easy programming. It's essentially SIMD with an emphasis on caching. That is, the vectorsized codes are executed on the GPU, but in a manner that the data from memory is collected into a "stream" (buffer really) where data can be operated on in parallel. Look-up PeakStream and RapidMind for an idea of how this stuff is programmed; anyone with an understanding of MATLAB should be able to do it.
fgsfdsSep 22, 2006
... but only if it needs little more than 32bit-precision floating-point math.Which, considering how badly the x86 preforms at FLOPing, would be a welcome boon indeed.
Closed AccountSep 22, 2006
Not on Linux it won't. I've already got my shiny new Nvidia card on order.
kcoraxSep 22, 2006
Until this is a free gain from GCC and Java or .Net JIT compilers I can't see it gaining momentum.
nanostuffSep 22, 2006
It's dual channel all over again!