The Cuda SDK comes with a Cuda compiler for x86. This way you can test the code on an x86 CPU for debugging purposes, or when you want to run code on a machine without Cuda hardware.
Ahh, I see. Back in 2008, I read that Nvidia was ready with an x86 CUDA compiler I didn't care to follow up but did it ever happen?
I used to say this but will repeat: Intel never really wanted Larrabee to take off. They wanted to generate just enough buzz of it to kill/slow the development of GPGPU such as CUDA and OpenCL. Only thing that's amusing to me is how it played out on gullible journalists.
Intel need to move beyond x86. Whats interesting is that ATI/NVIDIA GPU's are able to execute x86 faster than Intel's CPU's and we are not talking 30%-40%, we are talking many magnitudes faster.
Aren't all GPUs single precision only?
I think you missed Ben's point. All of that is a product of the architecture, not the instruction set. If someone was crazy enough they could build a similar chip that handled CUDA instructions just as well.x86 are very good at branching, like most other CPUs.
GPUs do NOT branch, that's the whole point.
When a branch occurs in code, the GPU will simply run both sides of the branch and disable the results of the branch that was not taken.
So it's implemented as a masking operation, rather than an actual branch.
The effect is that the paralellism goes down (it's still SIMD, so all threads have to execute the exact same instruction).
x86 is way different, as each core has its independent instruction pipeline and can branch completely independently from each other core.
Were does it say Intel has killed Larrabee.
I think you missed Ben's point. All of that is a product of the architecture, not the instruction set. If someone was crazy enough they could build a similar chip that handled CUDA instructions just as well.
Aren't all GPUs single precision only?
How are we defining DP? Because Cypress can certainly do DP.all GPUs EXCEPT fermi... fermi has been designed with one goal in mind, HPC.
Thus it is double precision.
yea i have a cypress based milkyway@home farm here which begs to differ. you have to buy 58xx chips if you wanna do milkyway on the gpu because ATI only supports DP on their highest end architecture. fermi is the only product out (that i know of) which can do GPU based IEEE DP, but that doesnt mean there arent other cards out that can do DP FP ops of some type. G80, G92, GT200 and RV770/740 based products all also have some form of DP
Aren't all GPUs single precision only?
yea i have a cypress based milkyway@home farm here which begs to differ. you have to buy 58xx chips if you wanna do milkyway on the gpu because ATI only supports DP on their highest end architecture. fermi is the only product out (that i know of) which can do GPU based IEEE DP, but that doesnt mean there arent other cards out that can do DP FP ops of some type. G80, G92, GT200 and RV770/740 based products all also have some form of DP
the ATI client is relatively recent from what i understand. the way i heard it was that some ATI engineers were doing some side work on the project, so they coded a GPU app for it with the guys at RPI directly in stream (ATI's own GPGPU API) to allow RV670 770/740 and 870 to all run the work units on a GPU. ATI is actually faster than NV for milkyway@home to boot. http://milkyway.cs.rpi.edu/milkyway/team_members.php?teamid=77&offset=0&sort_by=expavg_credit look at the team standings since i joined, that's 3 5870s going more or less 24/7 (with some hiccups) since i joined, and my recent average credit is still going up rather quickly. getting another 5870 next week for my birthday as wellI'm confused, doing Milkway@Home on a 4850 and have at times received no WUs because they were for Nvidia Vidcards only.
the ATI client is relatively recent from what i understand. the way i heard it was that some ATI engineers were doing some side work on the project, so they coded a GPU app for it with the guys at RPI directly in stream (ATI's own GPGPU API) to allow RV670 770/740 and 870 to all run the work units on a GPU. ATI is actually faster than NV for milkyway@home to boot. http://milkyway.cs.rpi.edu/milkyway/team_members.php?teamid=77&offset=0&sort_by=expavg_credit look at the team standings since i joined, that's 3 5870s going more or less 24/7 (with some hiccups) since i joined, and my recent average credit is still going up rather quickly. getting another 5870 next week for my birthday as well
ed: i should note i have no evidence which specifically says who was behind the ATI GPU client for milkyway, but given it's performance i wouldnt be surprised if some of ATI's engineers were helping. you should be able to run all the WUs on your 4850 now, about one every four minutes (vs my 5870's 1:25 per card). in comparison, it takes my 4.2ghz cpu cores about 4-5 hours a WU per core, sometimes longer. i think when they first started using the ATI client they were still reworking their work distribution software so that they could effectively use the ATI gpu influx. i have never been down for any reason thus far with my cluster due to lack of WUs for the GPUs
Interesting article on the subject:
http://www.brightsideofnews.com/new...lly-stumbled-developer-analysis.aspx?pageid=0
apparently you can't read either, Nemesis
this thread is about Larrabee the GPU dying, not Larrabee the HPC chip
There's simply no way intel can compete with AMD or nVidia's specialized architecture with what they were trying to do with Larrabee. Maybe they'll have something once we get a revolution in the graphics API department, whether it be something like ray tracing or unlimited point cloud, but until then there's just no way.