Your CPU will be able to do CUDA.
We've heard rumors that Nvidia been dipping its toe into the x86 CPU market, and today the graphics company made an announcement related to that – but it's not what you think.
Nvidia CEO Jen-Hsun Huang revealed that the company will bring its CUDA programming language to "any computer, or any server in the world," with the help of the Portland Group (PGI).
Specifically, this means that systems without Nvidia GPUs will be able to process CUDA code, giving the company its answer to Microsoft's DirectCompute and the more open OpenCL.
Nvidia says that its CUDA without a GPU will run best on multicore CPUs and will be ideal for servers.
(Source: Electronista.)
Well now you can run CUDA code WITHOUT the need for a Nvidia graphics card and will also allow them to compete with OpenCL and Directcompute.
And maybe it's cheaper to use what you have already (example in this case, a Supercomputer with say.. 100 CPU's), and would be cheaper to simply use those 100 CPU's instead of spending more cash on GPU's.
it's all in the text.
Well now you can run CUDA code WITHOUT the need for a Nvidia graphics card and will also allow them to compete with OpenCL and Directcompute.
And maybe it's cheaper to use what you have already (example in this case, a Supercomputer with say.. 100 CPU's), and would be cheaper to simply use those 100 CPU's instead of spending more cash on GPU's.
it's all in the text.
Except that OpenCL and Directcompute are compatible with all GPUs. CUDA is useless without GPU acceleration.
CUDA was toted as being nvidias answer to give exceptional processing power over x86 after Jen-Hsun Huang bashed it for so long.
Now that port THERE pride and joy to the thing they bagged for so long?
4chan made something called tripper for cuda. it runds tripcodes in cuda, arguably the best way to get trip codes you want. now a single core cpu can do i believe 1-5 million trips a second, an i7 920 can do i think 22 million, a gtx285 is capable of almost 2 billion and if its not faked i have seen numbers up to 15 billion but i know for a fact that this is the BEST use of the gpu in practice as a gpgpu. without gou slow, with gpu fast.
point being that 100cpus are outdone by a quad sli, if you have the ability to get 100cpus, just get 4 cpus.
But it says CPU's with CUDA capability.... I'm not getting it.
If it's faster on the CPU, it renders the GPU pointless.
If it's slower on the CPU, why use it?
You're thinking of PhysX. Wrong pun.
exactly my point, for the cpu we can use normal code, and either Microsoft or opencl and get all gpus to help. it still makes cuda over all pointless.
OpenCL is meant to make programming parallel programs easier, whether that means a graphics card or a bunch of traditional processors. I'd imagine that nVidia is looking to compete with that model now, rather than focusing only on programming parallel programs for graphics cards.
So you are not programming 2 ways in 1 application to add the CPU's power to the raw CUDA crunching.
It's also a way to learn CUDA without needing Nvidia hardware or a way to develop apps without needing Nvidia hardware. (I would assume you would deploy the finished app to CUDA hardware though)
OpenCL is owned by apple.
Honestly I trust Nvidia more from a "Nazi control freak over our IP" perspective.
Actually, I was thinking of nVidia's practice of not getting along. Disabling features would be consistent with their past practices of doing that (e.g. with PhysX, as you point out) if certain competitors' equipment is in your system.