At high performance systems:
Wikipedia says (#1):
As of 2010, the fastest six-core PC processor reaches 109 gigaFLOPS
(Intel Core i7 980 XE)[16] in double precision calculations. GPUs are
considerably more powerful. For example, Nvidia Tesla C2050 GPU
computing processors perform around 515 gigaFLOPS[17] in double
precision calculations, and the AMD FireStream 9270 peaks at 240 gigaFLOPS.
i7-980X @ 32nm = 239 mm sq. (#2) 109 GFlops/d... 130W max
C2050 @ 40nm = 526 mm sq. (#3) 515 GFlops/d... 247W max
The power consumption scales with chip, the performance at GPU is
however the double of the CPU's (that might be due to limited, but
massively parallel architecture).
I bet today you can get multiple of that performance in the GPU - the
limit is the way the heat is removed, if somebody makes a cooler capable
of dissipating 1kW from 2000 mm.sq area, I bet there would be such chips :)
By taking into account that nvidia is doing its best to design a power
efficient core (their primary target is high-performance) and they are
ahead of Intel in terms of efficiency, it leaves only a little chance to
you.
The question is how can you do better then them?
Maybe just make a small GPU with low core count - there are already to
be found in devices targetting the mobile market (powervr, mali, etc) so
in reality what you can do is just to make yours easier to use and well
documented. There is a big chance that none of the embedded cores will
have public documentation in near future.
From my point of view (as a board designer), I would really welcome a
discrete PCIe graphics, which has its ddr memory dies integrated in the
package. That saves a lot of design time, and you can guarantee the
performance/behaviour of the combination.
And there are then interfaces like C2C (found in Omap5 from TI) that let
you share the processors ram for the sub-processor cores (like the
baseband processor) so there is no need for a separate memory die. But I
would not go this way, an integrated ddr3/gddr die would be nicer.
Daniel
[1] http://en.wikipedia.org/wiki/FLOPS
[2] http://ark.intel.com/products/47932/Intel-Core-i7-980X-Processor-
Extreme-Edition-12M-Cache-3_33-GHz-6_40-GTs-Intel-QPI
[3] http://www.techpowerup.com/gpuz/7rgnu/
[4] http://www.siliconmechanics.com/files/C2050Benchmarks.pdf
On 12/10/2012 10:50 PM, Timothy Normand Miller wrote:
To motivate the energy efficiency problem with GPUs, it would be good to
have some statistics, and I know that some people on this list have
their fingers on this information. I'm trying to finish my CCF proposal
due Monday, and this could REALLY help.
Given state of the art CPUs and GPUs with the same process tech, what
are their transistor counts, per die? How about total area (mm^2)?
What is the relative power draw per die? (Keep in mind that some GPUs
are multi-chip)
The more clearly and specifically I can show that GPUs have urgent power
draw problems, the more reviewers are likely to fund the work.
Thanks.
--
Timothy Normand Miller, PhD
Assistant Professor of Computer Science, Binghamton University
http://www.cs.binghamton.edu/~millerti/
<http://www.cse.ohio-state.edu/~millerti>
Open Graphics Project
_______________________________________________
Open-graphics mailing list
[email protected]
http://lists.duskglow.com/mailman/listinfo/open-graphics
List service provided by Duskglow Consulting, LLC (www.duskglow.com)
_______________________________________________
Open-graphics mailing list
[email protected]
http://lists.duskglow.com/mailman/listinfo/open-graphics
List service provided by Duskglow Consulting, LLC (www.duskglow.com)
