While this isn't new it's appalling behaviour.
Tesla K80 - codename Stella: http://forums.laptopvideo2go.com/topic/31065-inf-v5011/
Some googling unveils this presentation: http://scent.gist.ac.kr/downloads/2014hpcss_14.pdf
-> slide 7: 2014 GK210-Duo, ~2TFLOPs DP, 500GB/s
I wonder if they stay in 225W, like their previous HPC cards. But 2014 DP-perf-crown seems to stay at AMD.
Some googling unveils this presentation: http://scent.gist.ac.kr/downloads/2014hpcss_14.pdf
-> slide 7: 2014 GK210-Duo, ~2TFLOPs DP, 500GB/s
I wonder if they stay in 225W, like their previous HPC cards. But 2014 DP-perf-crown seems to stay at AMD.
Nice find..
(( other than that, this slides have surely give some headache when presented, wow, what a mixture of old graphics ( against Nehalem ) and new ones, with in addition in the middle with future Pascal architecture... i dont know if they have present it in this order but wow. )
Money is very important actually: if you pay an employee $150k+/year to use it, the cost of his equipment is less important than the efficiency by which he can operate it. Perf/$ often only plays a small part in the whole decision process.
Last week, I finally got around to watching the Pixar keynote at Nvidia's GTC last March. What's interesting about it in this context is that they still have a large installed base of Fermi class GPUs and that a large part of their in-house tools are written around the Nvidia OptiX engine. What do you think is most important to them? The fact that AMD will offer a bit more Perf/$ or that all their existing tools can seamlessly be used across their current and future hardware without needing a rewrite?
It's possible that they have some layer around it to migrate from one vendor to another should the opportunity arise, but I don't think that AMD has anything at all that resembles OptiX.
It is a dual-GPU Tesla card, see Titan Z, so TDP should be similar at ~375w.
Titan Z is 2.66TFLOPs DP and no-GK210. This card should be below 300W TDP.
No way with these specs (should be around 520Mhz core clock). The memory might also use only 1.35V (has lower clock too). The K40 though was 235W, not 225W, so it might be the same. Otherwise, not more than 250W imho.
OK. So it's rather another chip to get more DP FLOPs for N TDP via mGPU.
Sorry, but the name GK 210-Duo was not do any mystery, it is a dual gpu.. basically a dual GK110 .. like was the first Kepler GK104 based Tesla .
2Tfllops DP for 6Tflops SP .. Its a TitanZ a little bit castrated on clock ( maybe with one SM too castrated ). I suspect the card to be really close to 250W maximum but a lot castrated on clock speed.
As for NvLink, not before Pascal or not before IBM release it .. ( IBM use a different name ) .. To understand that, Nvlink is the name given by Nvidia, but it exist similar connection used by other server brands, like SeaMicro in preparation etc.. Its just Nvidia point it today when other, have not yet speak about it. ( When you release this type of roadmap, you can put what is coming in 2017 ).
Well the most strange, is maybe to release a HPC gpu like that when we was allreaedy speak about GM210.. ( GK210 seems to be aimed against the W9100 of AMD, a little bit more SP Flops ( not much, we speak about 5.8Tflops vs 6Tflops ), less DP, s similar amount of memory shared on 2cores / memory controller )
2Tfllops DP for 6Tflops SP .. Its a TitanZ a little bit castrated on clock ( maybe with one SM too castrated ). I suspect the card to be really close to 250W maximum but a lot castrated on clock speed.
As for NvLink, not before Pascal or not before IBM release it .. ( IBM use a different name ) .. To understand that, Nvlink is the name given by Nvidia, but it exist similar connection used by other server brands, like SeaMicro in preparation etc.. Its just Nvidia point it today when other, have not yet speak about it. ( When you release this type of roadmap, you can put what is coming in 2017 ).
Well the most strange, is maybe to release a HPC gpu like that when we was allreaedy speak about GM210.. ( GK210 seems to be aimed against the W9100 of AMD, a little bit more SP Flops ( not much, we speak about 5.8Tflops vs 6Tflops ), less DP, s similar amount of memory shared on 2cores / memory controller )
Last edited by a moderator:
