[HH] Capped double precision performance in consumer video cards

[Tom Metro]

A colleague (who previously worked on HPC at UNH) shared this rumor with
me. Might be of interest to the few of you working on high performance
clusters.
 -Tom


-------- Original Message --------
Something I found interesting recently is how the GPU vendors
differentiate their products for different market segments.

ATI/AMD and Nvidia use the same silicon in both their consumer and their
workstation markets.

The workstation cards are more expensive, so how do they differentiate?

They used to do things like put more connectors on the workstation cards
so that more monitors could be driven.

Lately, Nvidia (and now AMD apparently) has been artificially capping
the double precision performance of the consumer cards.

(Not sure if it's blown fuses or what on the silicon. Apparently Intel
fuses off broken portions of its chips to sell as low-end products, but
not sure what the GPU manufacturers do.)

Nvidia capped Kepler GPU double precision performance to 1/24 of single
precision, and Maxwell GPUs use a ratio of 1/32.

AMD was apparently slower to start capping DP performance which may
explain the popularity of AMD GPU for bitcoin mining prior to the use of
ASICs.

See the posts here: GPU computing with AMD Hardware?
<http://www.researchgate.net/post/GPU_computing_with_AMD_Hardware>

It's not an issue if you only care about single precision performance,
such as this 24 node cluster <http://blog.cr.yp.to/20140602-saber.html>
described by Dan Bernstein for crypto research. It uses consumer grade
cards.

These artificial limits on double precision performance are not widely
known or reported in reviews of consumer-grade cards.