Watch for the Compute Capability

Up until now, my graphics cards have all worked fine with TensorFlow and PyTorch. And clearly they are still supported by Nvidia's cuda drivers.

However, much to my surprise, suddenly TensorFlow and PyTorch have effectively EOL-ed one of my cards : Which caused me to look at Nvidia's Compute Capability listings.

This post highlights cards that are now 'out', those on the verge of 'falling out', and those that should be Ok for a bit...

The Current Cut-Off

Now (and it seems to have been recently changed, since the card worked before), the Compute Capability required by TensorFlow and PyTorch stands at 3.5.

Run-down of cards

I made the following list pretty quickly : If in doubt, please refer to the original source. The main take-away is that :

• The level of each card within a series can be pretty arbitrary
• The spectra of obsolescence is upon cards that were pretty useable until the frameworks 'moved on'.

Pretty safe (5.2 and up)

• Nvidia Titan V/Xp/X
• GTX 10-series
• GTX 9-series (except the M-series below 965M)
• 910M

Strangely middling (5.0)

• 960M/950M/940M/930M
• 850M/840M/830M
• 750/750Ti

Now border-line (3.5)

• Nvidia Titan -/Z/Black (REALLY?)
• GTX 920M
• 780/780Ti
• 730/720

Now obsolete (below 3.5)

• 880M/870M
• 820M/800M
• 770/760
• 740
• 6-series, 5-series, 4-series, ...

Bottom Line

Compute Capability 'Expiry' is a problem. Specifically for my GTX 760.

Huge shout-out to the salesperson at Sim Lim Square (a large electronics mall in Singapore) who 'upgraded' me from a 750Ti to a 760 : And doomed my GPU-for-ML plan to premature obsolescence... /s

Cross-fingers that 5.2 stays valid for a while (and the Titan X that I won in an Nvidia poster competition remains relevant).