allanmac how did you come to the conclusion that FP16x2 is not in the GTX 1080? Is it a certainty or is it just unlikely?
I think it’s super unlikely because there has been no mention of “16.4 TFLOPS of fp16” in any of the reviews.
At least one review says it’s not there.
If it were possible then I can’t imagine a stellar number like that wouldn’t have been mentioned.
I would love to be wrong.
I asked Anandtech and they had no idea how to test it. So thats probably more the reason reviews dont mention it.
What about FP16 HGEMM? Does the 1080 support that? No information I can see really.
I could write a HGEMM routine and send it to them to test it! Why are we having to research this ourselves for these big companies?
if it was here, it will be mentioned in 1080 ad papers, like http://international.download.nvidia.com/geforce-com/international/pdfs/GeForce_GTX_1080_Whitepaper_FINAL.pdf
this paper is the source of most cuda-related info in all those reviews
http://www.nvidia.com/download/driverResults.aspx/103610/en-us
http://www.nvidia.com/download/driverResults.aspx/103608/en-us
Nvidia GeForce 368.25 WHQL driver for GTX 1080.
There is no fast fp16 in GP104.
Is this also true for GP106 and Parker, the Pascal Tegra SoC?
I can’t comment on future products.
Thank you for your work; I am looking forward to the overhaul.
If AMD is serious about this they really should have released something like CuDNN but with OpenCL. Dense linear algebra libraries are not enough.
Well crap. Time to start saving for that DGX-1
I think you’ll have to wait until Volta for GP100’s fast FP16 to be standard features across the entire GPU family line, just like GK110’s Dynamic Parallelism & Hyper-Q became standard features when Maxwell GM107 was released and then later when GM20x family was released.
i thought the GTX 1080 was going to be P100 but with GDDR5X. I guess its closer to the 980ti than the P100. Not really blowing my skirt up.
Anyone know when the final version of CUDA 8 will be available?
Since the GTX 1080 goes on sale Friday I would hope that CUDA 8 would move beyond the release candidate state.
“CUDA 8 is the most feature-packed and powerful release of CUDA yet. CUDA 8 will be available in August 2016 and there will be a release candidate available around June.”
Darn, there’s a lot of gamers that really want a 1080… managed to get one after being on the nowinstock tracker page and watching the comments update… NVIDIA GTX 1080 Pre-order & In Stock Tracker - NowInStock.net
I suspect the initial stock is sold out everywhere by now, but figured I’d leave the link above for reference.
Don’t blame the gamers; without them we’d all be buying $4k M40s instead of $1k Titan Xes
i believe that without gamers, we had buying intel xeons, even not phi
Tru dat.
Can someone run the updated CUDA 8.0 devicequery on Pascal GTX 1080/1070? Thanks to allanmac for the news.
Yes, can someone please do this ASAP !!! :-D
so… i compiled the fp16 example from cuda samples for the sm_61 architecture (GP104 should be sm_61 AFAIK) and then decompiled the fatbin.
the FP16 instructions are not software-emulated, the binary code is the same as for sm_60.
https://gist.github.com/RoBiK75/ac1282c14146bc685052bd6100f66f4e
we can see the HFMA2 and HADD2 instructions in the disassembled SASS