- Messages
- 5,317
TL;DR - I have found through Chips and Cheese that the 3.2 and 3.8 teraflops of claimed FP64 performance on TechPowerUp and Notebookcheck for the 7900 XT and XTX respectively are both very wrong and should instead be 1.6 teraflops and 1.9 teraflops respectively. I have emailed both sites about this and the databases should be updated to reflect this soon.
This started when the 7900 XT was first launched and I saw TechPowerUp's listing for the card saying that just the 7900 XT alone had a whopping 3.2 teraflops of FP64 performance. Naturally, when I saw this with a card that also had very damn good FP16 (for AI) and real-time ray-tracing performance, it made me pretty excited. Finally, a consumer card that could do it all and wouldn't completely demolish your bank account! But then... I dunno, I guess my paranoid side was nagging me a little too much saying perhaps this number was too good be true and the number was wrong. FP64 is, nowadays, almost ALWAYS shafted on consumer GPUs in favor of cramming yet more FP32 performance into it.
But hey, maybe this AMD card would finally be a modern exception! So I dug a little further and... Nothing. I couldn't find ANYTHING that concretely stated what the FP64 performance of these AMD cards were. So I messaged TechPowerUp directly. They didn't know and emailed AMD about it. Several times. AMD didn't respond. Then I saw that Notebookcheck advertised the same numbers. I emailed them about it. They didn't quite know for sure, but they gave me a very helpful link to an article by Chips and Cheese which finally concluded after some testing that the FP64 performance on RNDA3 has actually regressed as opposed to RDNA2. FP64 is now 1/32nd of the performance of the FP32 number (a 1:32 ratio instead of the 1:16 ratio that it was on the older RNDA2 cards).
Chips and Cheese also emailed me thus, confirming this reading of the results:
This started when the 7900 XT was first launched and I saw TechPowerUp's listing for the card saying that just the 7900 XT alone had a whopping 3.2 teraflops of FP64 performance. Naturally, when I saw this with a card that also had very damn good FP16 (for AI) and real-time ray-tracing performance, it made me pretty excited. Finally, a consumer card that could do it all and wouldn't completely demolish your bank account! But then... I dunno, I guess my paranoid side was nagging me a little too much saying perhaps this number was too good be true and the number was wrong. FP64 is, nowadays, almost ALWAYS shafted on consumer GPUs in favor of cramming yet more FP32 performance into it.
But hey, maybe this AMD card would finally be a modern exception! So I dug a little further and... Nothing. I couldn't find ANYTHING that concretely stated what the FP64 performance of these AMD cards were. So I messaged TechPowerUp directly. They didn't know and emailed AMD about it. Several times. AMD didn't respond. Then I saw that Notebookcheck advertised the same numbers. I emailed them about it. They didn't quite know for sure, but they gave me a very helpful link to an article by Chips and Cheese which finally concluded after some testing that the FP64 performance on RNDA3 has actually regressed as opposed to RDNA2. FP64 is now 1/32nd of the performance of the FP32 number (a 1:32 ratio instead of the 1:16 ratio that it was on the older RNDA2 cards).
Chips and Cheese also emailed me thus, confirming this reading of the results:
Ultimately, this makes pretty much zero difference for gaming on the card and its many other uses, but even so, I think people should still probably know just in case they buy the card with the expectation of that, in reality, fake 3.2+ teraflop FP64 performance number.
