How do the "games-oriented" NVIDIA-GPU-based cards behave w.r.t. CUDA computation?

Question

NVIDIA makes GPUs, which are used both for graphics and for graphics-unrelated computing. It also publishes compute-related specs of its GPUs.

Now, when it comes to cards:

• there are some cards that NVIDIA sells itself as the sole vendor, such as the Quadro and the Tesla series (with the latter being solely compute-oriented, lacking any display outputs); these have their own GPUs which you don't see on consumer-oriented cards.
• There are the consumer-oriented cards which vendors such as MSI, GigaByte and EVGA sell, which are graphics-oriented and use NVIDIA GPUs. These come in all sorts of varieties, such as "SuperClocked", "FTW", "D5", "OC", "Extreme Edition", "Green/Red/Blue/Black" etc.
• Finally, NVIDIA has started over the past couple of years to sell "Founders' Edition" cards with the same GPUs (chip-wise) used by the other consumer-oriented card vendors.

My question is: Regarding compute work, can I only trust the "Founders Edition" to deliver consistent performance in exact correspondence to the specifications? Or can I also trust the other vendors' cards to not "mess things up", and conform to the specs (modulo stated changes in core and memory clock speeds)? And if I can't fully trust the non-NVIDIA vendor cards, how would they "misbehave"?

If the scope is too general, I'm specifically interested in GTX 1080 Ti cards (GP102-350-K1-A1 chips).

Answer 1

We use off-the-shelf consumer-grade GTX 1080's for compute work. They are not "Founders Edition" and work without fuss. They fail like any normal GPUs would. No one will give you the guarantee you seek. Not NVIDIA, not MSI, GIGABYTE, nor EVGA. But they will work consistently across GPUs and over time.