Echelon is a system built on an architecture. Kepler is a continuation of Fermi on 28HP. So this means that Echelon is Maxwell.
FenderBender
Newcomer
No, this is beyond Maxwell. The slides show Maxwell as being about 15 GF/watt, so call it 4TF per GPU. The exaflop GPU design goal is another order of magnitude beyond that at 40TF/GPU, as seen by the slides.
Also the time scale on the slides show a projected 2016 production. NVidia's track record between major family updates is about 18-24 months. (G80 was early 2007, GT200 was mid 2008, GF100 was mid 2010).
Megadrive1988
Veteran
Ok so we have:
G80 (NV50), late 2006 (it was 2006 not 2007)
GT200 (NV55), mid 2008
Fermi (NV60), mid 2010
Kepler (NV65), late 2011
Maxwell (NV70), late 2013
unnamed Nvidia GPU for Echelon (NV80) - 2016
Except that it's the overhead that makes things very costly, energy-wise (as you're well aware). And that getting the data where it's needed, costs much more energy (I think order(s) of magnitude here) compared to the calculation itself.
Yes, my point was that this is a pretty good indication of the importance of architectural innovation
If the ALU power was a much higher percentage, then it wouldn't be quite as critical. Clearspeed certainly had an extremely impressive architecture in terms of perf/w for HPC, but then again it was fairly restrictive.The hierarchy of registers and customizable SRAM memory match Dally's earlier statements about minimizing data movement.
I'm curious what overheads are involved in making the memory so flexible as to allow it to subdivide itself into multiple levels of cache.
I'm curious what overheads are involved in making the memory so flexible as to allow it to subdivide itself into multiple levels of cache.
A1xLLcqAgt0qc2RyMz0y
Veteran
More details on Echelon
Link
http://www.xbitlabs.com/news/video/...mance_Needed_for_ExaFLOPS_Supercomputers.html
Link
http://www.xbitlabs.com/news/video/...mance_Needed_for_ExaFLOPS_Supercomputers.html
Megadrive1988
Veteran
Are we certain that Kepler will be a refresh/improvement on the Fermi architecture or a whole new architecture altogether?
Well, the timing is most definitely right for the next major architecture from nVidia to be a refresh of Fermi. They haven't broken their new architecture->upgrade cycle yet.
My guess is that Maxwell will be the next new arch while Kepler would be more of a refresh.
So which physicist(s) cover G80 and GT200? And if there is only one name in NV DX10 generation what is the so called "middle life kicker"? What came before Tesla (which "physicist" and what chip was the last part of that family) and what after Tesla (this is an obvious one: Fermi and GF100).
Why stop at G80 and GT200? Given that some folks round here have the really bizarre habit of wishing to map current and future NVIDIA architectures to the long-disused NVxx naming scheme (that was dropped long before NV likely even conceived of these new generations), why not do the reverse and assign physicist names to NVIDIA architectures all the way back to NV1?!
G80 obviously didn't have something like this. G stands still for Geforce and there's nothing left besides the number. GT200 was Tesla (Geforce Tesla), the mid-life kicker supposedly GTX285, but there I am not sure, that everything went according to the original plans.
GF100 is Geforce Fermi, GKx00 (maybe they'll start with 100 again - they definitely should) will be Geforce Kepler, GMx00 Geforce Maxwell.
That at least is my interpretation.
Unless my aging memory fails they've had scientists in there for a long time. It goes something like:
- NV10 - Celsius
- NV20 - Kelvin
- NV30 - Rankine
- NV40 - Curie
- G80 (NV50) - Tesla
OT: Arguably, this is pretty good trolling material if one stops to think about it - NVIDIA cards are clearly designed for serious scientific stuff, just look at their codenames...unlike those ATI bums who use plants and islands, clearly gaming focused hippies they are! I'm quite amazed it hasn't been used yet, TBH.
