Discussion That 128MB cache on the RX 6000, seems like not enough for 4k

EliteRetard

Diamond Member
Mar 6, 2006
6,490
1,021
136
After lots of reviews from multiple sources, I get the impression that these RX 6000 cards drop off significantly at 4k vs the Nvidia RTX 3000...despite the large difference in VRAM. Performance between the two is much different at lower resolutions like 2560x1440. My thinking is that it's a limit of the 128MB cache, and the slower memory speed when it runs out.

I'm curious what other people think. There doesn't seem to be much discussion about this interesting portion of the new AMD GPUs.
The question for 4k gaming so far has been "is 10GB VRAM enough?" but I'm curious "is 128MB cache enough?"
 

Hitman928

Diamond Member
Apr 15, 2012
5,180
7,631
136
The Ampere architecture scales unusually well with increasing resolution. Alternatively it doesn't scale as well as you lower resolution. The RX 6000 series cards scale fine, compare them to the 2080 Ti and you'll see they don't have a problem scaling up to 4K. It's just that Ampere's architecture is a bit compute focused and unbalanced for gaming so it's not really able to 'flex its muscles' until you get to higher resolutions.
 
  • Like
Reactions: lightmanek

EliteRetard

Diamond Member
Mar 6, 2006
6,490
1,021
136
The Ampere architecture scales unusually well with increasing resolution. Alternatively it doesn't scale as well as you lower resolution. The RX 6000 series cards scale fine, compare them to the 2080 Ti and you'll see they don't have a problem scaling up to 4K. It's just that Ampere's architecture is a bit compute focused and unbalanced for gaming so it's not really able to 'flex its muscles' until you get to higher resolutions.

:hmm:

I'm gonna go check the charts real quick...
 

Elfear

Diamond Member
May 30, 2004
7,096
640
126
The Ampere architecture scales unusually well with increasing resolution. Alternatively it doesn't scale as well as you lower resolution. The RX 6000 series cards scale fine, compare them to the 2080 Ti and you'll see they don't have a problem scaling up to 4K. It's just that Ampere's architecture is a bit compute focused and unbalanced for gaming so it's not really able to 'flex its muscles' until you get to higher resolutions.

What he said. The doubled up FP32 of Ampere really starts to shine at 4k. It's more a function of Ampere doing unusually good at 4k rather than RDNA2 doing unusually bad.
 
  • Like
Reactions: lightmanek

Dribble

Platinum Member
Aug 9, 2005
2,076
611
136
The more memory you use the worse the cache performs. I'm guessing at 4k games are using about 8-10GB and you can see the performance is already starting to tank. So what's the point of putting 16GB on the card as you can't use it without killing performance. They would have been better off just giving the cards 8GB and knocking $50 off the price.
 

Saylick

Diamond Member
Sep 10, 2012
3,084
6,184
136
Posted before but it seems relevant for this thread:
2020-12-03-image-4-p.webp
 
  • Like
Reactions: krumme

Hitman928

Diamond Member
Apr 15, 2012
5,180
7,631
136
The more memory you use the worse the cache performs. I'm guessing at 4k games are using about 8-10GB and you can see the performance is already starting to tank. So what's the point of putting 16GB on the card as you can't use it without killing performance. They would have been better off just giving the cards 8GB and knocking $50 off the price.

From what I understand, it has far more to do with the frame buffer (i.e. resolution) than how much VRAM is being used. Again, compare RDNA2 to Turing scaling and it's fine. Ampere just has the least penalty from increased resolution than any other architecture due to its more compute focused design. Look at HWUB and their 18 game average at 1440p and 4K. At 1440p the 6800XT is 25.6% faster than a 2080Ti. At 4K the 6800xt is 25.7% faster than a 2080Ti. RDNA2 scaling is fine and scales more in line with other, less compute focused, architectures.
 

EliteRetard

Diamond Member
Mar 6, 2006
6,490
1,021
136
Posted before but it seems relevant for this thread:
2020-12-03-image-4-p.webp

I'm curious about this chart, how are you supposed to read it?
Why are there multiple points along the lines, what do they represent?

The way I look at it, it does seem to suggest that 4k has a lot more room to scale with extra cache.
At 1440 and 1080 they do taper off around that 128MB level, but 4k looks like it's still near linear.
Feels like the 5800/5900 could have benefited from at least another 64MB as 4k focused GPUs.
Leave the 128MB for the lower tier 1440 cards, and maybe 96MB for the lowest 1080 class.
 

BFG10K

Lifer
Aug 14, 2000
22,672
2,817
126
I'm guessing at 4k games are using about 8-10GB and you can see the performance is already starting to tank.
What...what? For months we've been told 8GB is just fine on nVidia parts, but now suddenly 16GB AMD parts have a problem?

So what's the point of putting 16GB on the card as you can't use it without killing performance.
Do you feel 16GB + 128MB cache will be "killed" more than 8GB with no cache?
 
  • Like
Reactions: lightmanek

EliteRetard

Diamond Member
Mar 6, 2006
6,490
1,021
136
I do wonder if AMD would have been better off with an "odd" bus width like Nvidia.

Keeping the same 128MB cache, if you could choose:

384bit 12GB @ 768GB/s
or
256bit 16GB @ 512GB/s

That could also allow a Nvidia like move, to further differentiate 6800XT and 6900XT by VRAM (12 vs 24 GB)

I wonder if AMD knew Nvidia was gonna stick with "just" 8 - 10GB for top of the line GPUs.
 

coercitiv

Diamond Member
Jan 24, 2014
6,151
11,686
136
The way I look at it, it does seem to suggest that 4k has a lot more room to scale with extra cache.
At 1440 and 1080 they do taper off around that 128MB level, but 4k looks like it's still near linear.
Feels like the 5800/5900 could have benefited from at least another 64MB as 4k focused GPUs.
The cache is there as as an enhancement within a fixed die area budget. There will be a threshold where further increasing cache size will hurt performance as diminishing returns coming from increased effective bandwidth will be compounded with lower CU count.

Whether that threshold is at 128MB or futher up is not easy to ascertain just by looking at cache hit rate info. I would argue for example that the 6800 and 6800XT would have lower performance at 1440p with 186MB IC.