How many Shaders/CUDA Cores are required to be useful?

[[WHGT]Cyberman]

As far as I understand it, memory and frequency aren't as important as the number of shaders/CUDA cores. Because the actual calculations are mainly a LOT and not necessarily complicated. (From point of view of a processor at least.)

So my understanding would be that a good indicator to look at when shopping for a cheap GPU to fold would be to check how many shaders there actually are. The chips are often the same, but the environment differs radically. But if card A has 512 shaders, and card B has 512 shaders BUT also comes with 2GB DDR4 ram (or whatever), then card A should be about as useful as card B (for folding).
(Assuming that clock speed is about the same.)

So, what would currently be a good number of shaders for a GPU to have to be useful for some time?


Usual disclaimers because this is the internet and someone will misunderstand me:
I'm not entirely sure if this is the right forum - if not, sorry.
I did look before posting, I also searched but didn't find anything conclusive.
I feel as if this has been asked before, but maybe I merely thought about asking it and never did...
I'm not looking for a card that does folding + gaming or something. I have a card for that. But occassionally I look around for cheap cards and wonder if they'd be able to do useful work.

[Neil-B]

viewtopic.php?f=16&t=34553&p=327776&hil ... rs#p327776 seemed to indicate that in order to meet deadlines something like 200/400 shaders minimum?

I have a K420 with 192 cores that is borderline - can sometimes complete within Expiration Deadline but other times fails ... tbh I rarely fold with it ... Both my M1000M 512 cores and GTX750Ti 640 cores tend to make Timeout most of the time ... so my current guess would be 250 to make expiration deadline and 500 to make timeout?

[Foliant]

so my current guess would be [...] 500 to make timeout?

Can confirm, my GTX 560 Ti, 448 Cores usually completes around Timeout.

[MeeLee]

Depends on the core frequency.
Generally speaking, a 384 core, running at or around 1Ghz will make the deadline, though barely (GT 730).
That same GPU running at 2GHz is enough to finish a WU in a day (GT 1030).
If Nvidia would have made a 192 core 2Ghz GPU, it'd be roughly as fast as a GT730.

[bruce]

The advertisements for GPUs report a value of GFLOPS which happens to be the product of the number of shaders and the frequency. plus a spare factor of 2. That means if the software is able to keep all of the shaders busy all of the time (which is actually impossible in useful code) the GFLOPS number is a reasonable way to rate GPU productivity.

The OpenMM code massages the protein into a series of blocks of work which happen to be mostly FP32 instructions plus some essential FP64 instructions and some essential serial code plus the data needs to be moved between main RAM and the GPU. How much of the calculation time can be measured by the GFLOPS value? The critical word is "mostly" (i.e.- take a guess)

In addition, you might be assigned a protein that's able to keep the shaders busy for a high percentage of the time or you might be assigned a protein that only keeps them busy a smaller percentage of the time. That's what keeps us from giving you an accurate number of shaders needed.

Personally, I'd recommend something bigger (wider) that a GT x30. Go for a GTX x50 or above. Personally, I also avoid the top-of-the-line until the price comes down.

[v00d00]

An onboard Carrizo R7 doesnt cut it. That much I found out recently.

A 1050 ti is worth about 220k/day and uses about 50-60w of juice.

Nowadays a 1650 Super or 1660 seems to be the minimum if you want a good return.