OTOY Forums

As some of you might have read elsewhere on these forums, Redshift isn’t currently running well on GTX970/980. The good news is that we know why that is and are currently testing a fix.

The reason for the slowdown appears to be driver-related (with a possibility that the Windows driver model is responsible) but thankfully we know how to bypass it. We’ll be contacting NVidia with our findings but, given that we might not receive a helpful answer from them, we’ll most likely go ahead with the fix - if we detect these videocards. If you’re curious, the issue has to do with the driver-side memory management.

So how does the GTX970 perform after the fix?

Well, testing with “ray reserved memory” left at zero (the default), reveals that the new GPUs are indeed as fast as they claim to be. When compared against an artificially-memory-limited K6000 (it normally has 12GB, we limit it to 4GB to make it ‘equal’ to the 970), the GTX970 is actually running faster! It’s not faster across the board: certain aspects of the code run faster on the K6000 due to its superior memory bandwidth and more cores but, thankfully, there is enough code in Redshift that does care about Maxwell’s architectural improvements. So the final result is indeed favorable for Maxwell. The gains we see so far range between 10% and 20% which are not earth-shattering but considering that we’re talking about a videocard that costs around $330-$380… it’s pretty impressive! smile The 980 should offer even greater benefits.

Since we only bought our GTX970 today, we obviously haven’t had the time to do exhaustive benchmarking. We’ll try to do more of that in the following days. Some of you already own these cards so I’m sure you’ll be posting your results too.