No article because it's not really for anyone but TR at this point due to being rough around the edges, but when it's finished I'll write it up.
Brodda Thep
Newcomer
FPGAs have more of a role on running already trained neural nets. It would be ungodly expensive and time consuming to use FPGAs for training. However, if you want extensive neural nets to run on cell phones, that may be the way to go.
However, neural net techniques are just changing too quickly. You need the resources of the community to keep up. This means CUDA. There are three main libraries, Torch, Caffe, and Theano. Then there are tons of tertiary libraries. However, they all use CUDA. nVidia is investing in this technology. AMD is not.
Neural nets are the wild west of computer science. Anyone with a shoestring budget can make large contributions and even try them in brand new areas. 12 GB gives me a ton of flexibility. I am constantly running up against memory limitations when I am training a spatially sparse network. With 12 GB this card will be worth while for years. With 6 GB this card would no longer worthwhile after the first 980Ti is released.
I am very happy that nVidia is supporting the deep learning community.
Anandtech mentioned there is another 8-pin connector.
Nvidia may be holding back custom variants to see what AMD drops.
In return, AMD may be holding back its water-cooled variant to see what Nvidia drops.
When they're done with the itty-bitty lightsabers, I'm sure we'll be the first to know.
whoaoamm whoosh bzzzt
I suppose all is relative in the hype Olympics. There could be a series of performance tiers coming up, possibly more tiers than cards to fill them, but we may be in for a bit of a silly season.
Yep, 100% culled. There are other variants of that test that use different cull ratios and different vertex and index sizes that help explore throughput a bit more, which I'll ship in the next update.
That's been on quite a lot of cards and re-used for Quadro- and Tesla-Boards which have their (sole) a-pin connector out of the back of the card. IIRC Fermi was the first to use such a L-shaped power-grid.
Oh, I missed the increase to half rate. And now even GK110 looks more like that. So yes, GM200 is underperforming, maybe GigaThread Engine was not scaled up?
According to Hardware.fr's GM204 article its 1/3 rate per SMM, reduced from 1/2 rate per SMX; the clusters are slimmer and scaled up in number relative to Kepler so I assume they wanted to reduce front-end area per SM.
I just look at those results on the TR review.
The hardware.fr article also has similar results. In this and their previous article GM107 is consistent with 1/3 rate while GM204 is somewhat overperforming a 1/3 rate on culled tri's. It is, however, closer 1/3 rate than 1/2 rate and I can only attribute the performance to GM204 hitting some rather large boost clock on what it possibly a relatively low power scenario; GM200 probably isn't hitting similar boost levels hence looks like it underperforming relative to it number of SMM's in comparison to GM204.
The hardware.fr article also has similar results. In this and their previous article GM107 is consistent with 1/3 rate while GM204 is somewhat overperforming a 1/3 rate on culled tri's. It is, however, closer 1/3 rate than 1/2 rate and I can only attribute the performance to GM204 hitting some rather large boost clock on what it possibly a relatively low power scenario; GM200 probably isn't hitting similar boost levels hence looks like it underperforming relative to it number of SMM's in comparison to GM204.
If the meausrement is exact, it would correspond to a 1.5Ghz boost clock though for the GTX 980. Seems a bit on the high side, though I'm not sure on the maximum possible boost. For the Titan X however the results would correspond to only 1 Ghz - even assuming it can't reach that high of a boost that would only be roughly base clock. Maybe there's some front end limitations somewhere?
FWIW I'm also surprised how terrible the r290x does with strips...
edit: actually these results have to be Vertices/s not Tris/s right?
Last edited:
Ok, thank you for your opinion
GM204 seems like a prime candidate.
28nm 400mm2 256bit ~180w -> 16FinFet ~220mm2 256bit ~100w
Bulk GM206 up to 10 or 12SMM and 16FinFet would put it at ~120mm2 128bit ~60w.
Titan X Black Edition?
NVIDIA is certainly holding back thus far with Maxwell. They could have easily released the Maxwell parts with 15-20% higher clocks, but I think it made the best business sense to let AMD remain competitive (selling their chips nearly at cost). Sort of like how Intel does with CPUs.
Bottom line is, NVIDIA would never put itself at a performance disadvantage in the name of power efficiency. They simply no longer have to make 300W cards to beat AMD.
