F@H Computational Horsepower Quadruples due to COVID-19

[nivedita]

Nice blog.

Could you make a post about any of Nvidia Quadro GPUs? I’ve heard that Quadro is almost twice more efficient in double-precision ( FP64 ) calculations than non-Quadro equivalent. FP64 is what you need for Folding. It’s very interesting.

Most of F@H is still single-precision, and note that in any case most recent Quadro GPUs do not have high-performance double-precision. Only the Quadro GP100 and GV100 do among the current generations after Kepler, and they are incredibly expensive. None of the other Maxwell, Pascal or Turing Quadros have anything beyond single-precision.

which means an Nvidia Quadro GP100
is more effective than a GTX1060 ?

if so, the question is, can i put it in the same slot of a midi tower or is the Quadro GP100 too big for that?

i ask because i don't want to spend money senselessly, because otherwise i save the money i invest the money in a GTX1660 Ti with 6 gb ?

I wasn't kidding when I said they are extremely expensive. A quadro gp100 will set you back around $7,000. Yes, it is still a standard-sized dual-slot card so it will fit, but you'd have to be nuts to drop $7k on a card just for F@H. It might fold slightly faster than a gtx1080ti but the difference will be in the noise.

[ipkh]

There is a need for both SP and DP floating point in fah tasks. GPUs without FP64 aren't usable in fah anymore.
But as a cost/benefit calculation Quadro cards don't make sense. You can get a multitude if RTX cards for the same price and collectively outperform it.

ECC ram isn't neccessary for most things and the normal expected error rate in RAM is extremely low. And the folding core will catch those errors and redirect the calculation to correct it.

[nivedita]

There is a need for both SP and DP floating point in fah tasks. GPUs without FP64 aren't usable in fah anymore.
But as a cost/benefit calculation Quadro cards don't make sense. You can get a multitude if RTX cards for the same price and collectively outperform it.

ECC ram isn't neccessary for most things and the normal expected error rate in RAM is extremely low. And the folding core will catch those errors and redirect the calculation to correct it.

Are there GPUs without FP64 at all? I thought most were just neutered with 1/16th or 1/32nd of the compute resources for FP64.

[ipkh]

There have been some discussions on some AMD cards not supporting enough do to be used for folding. All nvidia cards do to some extent but as you said it is very limited for consumer variants.
I think current Quadro cards simply match the Rtx cards in DP performance.
I checked and the rtx5000 quadro card matches the 2080 in DP performance so there is no advantage to using quadro unless you are running specific software like AutoCAD and Maya.

[Joe_H]

There are some nVidia GPUs, especially among the mobile ones that do not support DP. But it is more commonly available across the various lines of their consumer grade cards than within the AMD offerings. Most of the cards based on Fermi or newer chips from nVidia have DP support. Some of the oldest don't support OpenCL 1.2 which is generally required.

In the AMD lineups, the high end cards have had DP support for years, as you get closer to the present most have DP support on the desktop. The oldest cards that can be used are the HD 5800 and HD 6900 series, the lower level cards can't. As you move up to the later series, support for DP moves down into the midrange cards. Ffor example the HD 7700 and above cards support DP.