5 ms·

Guide looks very well written, and I'm actually studying parallel programming right now...bookmarked. Side Note: Maybe this is Nvidia's way to promote sales, s

Guide looks very well written, and I'm actually studying parallel programming right now...bookmarked. Side Note: Maybe this is Nvidia's way to promote sales, since AMD cards sold like crazy cause of the mining craze.

jjoonathan 13y ago
It's not reactionary: nvidia invested heavily in getting to market first and advertising to the HPC crowd and they have been reaping the rewards of a monopoly for a few years now. Their latest high-end cards are dramatically more expensive (vs AMD at the same raw computational power) while offering completely crippled GPGPU capabilities. I recall incredulity on IRC when people found that their new 6xx cards were a step backwards from the corresponding 5xx cards for their GPGPU apps. Pro media & HPC users fork over the cash for Tesla because they can't jump to AMD due to their legacy CUDA code. Gamers don't care since they don't use double precision arithmetic. The bitcoin community was the only one nimble enough to switch to AMD. For those of us in academia, we'll be paying the piper for the foreseeable future :/
- hendzen 13y ago
  Bitcoin miners flocked to AMD because the SHA-256 algorithm utilizes a 32-bit right rotate operation that AMD cards could execute in one clock cycle, while NVidia cards took ~3 cycles to do so. At this point it doesn't really matter since GPU mining is obsolete. Also, what's wrong with NVidia's GPGPU capabilities? The last time I had to write GPU code, I found that CUDA was much more mature - and much more pleasant to write than the equivalent OpenCL. Also, the benefits of being able to run OpenCL code on multiple compute platforms were being realized by CUDA with the development of a PTX to x86 compiler.
  - hdevalence 13y ago
    It's not just about right rotation, it's also about architectural differences: the AMD cards have more ALUs to work with. https://en.bitcoin.it/wiki/Why_a_GPU_mines_faster_than_a_CPU#Why_are_AMD_GPUs_faster_than_Nvidia_GPUs.3F https://en.bitcoin.it/wiki/Why_a_GPU_mines_faster_than_a_CPU...
  - jjoonathan 13y ago
    >Bitcoin miners flocked to AMD because the SHA-256 algorithm utilizes a 32-bit right rotate operation that AMD cards could execute in one clock cycle, while NVidia cards took ~3 cycles to do so. That would have made the difference even larger, but even at 1:1 instruction timing AMD would have had a large price advantage so long as we're talking about the last 2 generations of cards or so. If the price/performance optimum lies further back than that, it may well have fallen at a point in time when AMD / NVidia were closer to price/performance parity. To clarify, by "performance" I mean "performance for my needs" which means "double precision float performance." > Also, what's wrong with NVidia's GPGPU capabilities? The price per double-precision FLOP was off by a factor of four at the consumer level last time I went shopping and the number of DFLOPS available at the consumer level was capped lower. It seemed like a move to force penny-pinching CUDA-dependent academics to upgrade to teslas. > The last time I had to write GPU code, I found that CUDA was much more mature - and much more pleasant to write than the equivalent OpenCL. It still is, but the gap has closed for many use cases (including mine). I'm not saying that CUDA isn't/wasn't a reasonable choice, especially a few years back, just that we are now paying the hidden price that comes from allowing a monopoly to develop.

jjoonathan 13y ago

It's not reactionary: nvidia invested heavily in getting to market first and advertising to the HPC crowd and they have been reaping the rewards of a monopoly for a few years now. Their latest high-end cards are dramatically more expensive (vs AMD at the same raw computational power) while offering completely crippled GPGPU capabilities. I recall incredulity on IRC when people found that their new 6xx cards were a step backwards from the corresponding 5xx cards for their GPGPU apps. Pro media & HPC users fork over the cash for Tesla because they can't jump to AMD due to their legacy CUDA code. Gamers don't care since they don't use double precision arithmetic. The bitcoin community was the only one nimble enough to switch to AMD. For those of us in academia, we'll be paying the piper for the foreseeable future :/

hendzen 13y ago
Bitcoin miners flocked to AMD because the SHA-256 algorithm utilizes a 32-bit right rotate operation that AMD cards could execute in one clock cycle, while NVidia cards took ~3 cycles to do so. At this point it doesn't really matter since GPU mining is obsolete. Also, what's wrong with NVidia's GPGPU capabilities? The last time I had to write GPU code, I found that CUDA was much more mature - and much more pleasant to write than the equivalent OpenCL. Also, the benefits of being able to run OpenCL code on multiple compute platforms were being realized by CUDA with the development of a PTX to x86 compiler.
- hdevalence 13y ago
  It's not just about right rotation, it's also about architectural differences: the AMD cards have more ALUs to work with. https://en.bitcoin.it/wiki/Why_a_GPU_mines_faster_than_a_CPU#Why_are_AMD_GPUs_faster_than_Nvidia_GPUs.3F https://en.bitcoin.it/wiki/Why_a_GPU_mines_faster_than_a_CPU...
- jjoonathan 13y ago
  >Bitcoin miners flocked to AMD because the SHA-256 algorithm utilizes a 32-bit right rotate operation that AMD cards could execute in one clock cycle, while NVidia cards took ~3 cycles to do so. That would have made the difference even larger, but even at 1:1 instruction timing AMD would have had a large price advantage so long as we're talking about the last 2 generations of cards or so. If the price/performance optimum lies further back than that, it may well have fallen at a point in time when AMD / NVidia were closer to price/performance parity. To clarify, by "performance" I mean "performance for my needs" which means "double precision float performance." > Also, what's wrong with NVidia's GPGPU capabilities? The price per double-precision FLOP was off by a factor of four at the consumer level last time I went shopping and the number of DFLOPS available at the consumer level was capped lower. It seemed like a move to force penny-pinching CUDA-dependent academics to upgrade to teslas. > The last time I had to write GPU code, I found that CUDA was much more mature - and much more pleasant to write than the equivalent OpenCL. It still is, but the gap has closed for many use cases (including mine). I'm not saying that CUDA isn't/wasn't a reasonable choice, especially a few years back, just that we are now paying the hidden price that comes from allowing a monopoly to develop.