Civilization VII Performance Benchmark

[x4it3n]

W1zzard didn't test MSAA here, only GameGPU did that

MSAA or not the 4090 should be faster than the 5080 lol

 

[Noreng]

MSAA or not the 4090 should be faster than the 5080 lol

The problem with these GPUs is that there are more SMs than the front-end can feasibly feed at times. If the 4090 scaled decently with it's SM count relative to the 4080, it should have been almost 60% faster, instead it's more like 35% faster on a good day. And it's not just the 4090 that's "underperforming" either, the 4080 "should" be almost 230% faster than a 4060, but in practice it's closer to 150% faster.

Civilization 7 is most using a very simple rendering pipeline, considering it runs on a Nintendo Switch with 2 Maxwell SMs at ~750 MHz. The explanation for why Blackwell seems to run the game "better" might be as simple as some select instructions executing significantly faster on Blackwell than on Ada and Ampere. Intel's B580 is also performing unusually well here, while RDNA2 seems to run especially poorly.

 

[odellus]

bizarre performance scaling

 

[x4it3n]

The problem with these GPUs is that there are more SMs than the front-end can feasibly feed at times. If the 4090 scaled decently with it's SM count relative to the 4080, it should have been almost 60% faster, instead it's more like 35% faster on a good day. And it's not just the 4090 that's "underperforming" either, the 4080 "should" be almost 230% faster than a 4060, but in practice it's closer to 150% faster.

Civilization 7 is most using a very simple rendering pipeline, considering it runs on a Nintendo Switch with 2 Maxwell SMs at ~750 MHz. The explanation for why Blackwell seems to run the game "better" might be as simple as some select instructions executing significantly faster on Blackwell than on Ada and Ampere. Intel's B580 is also performing unusually well here, while RDNA2 seems to run especially poorly.

The 4090 scaling is poor because it is Memory Bandwidth starved! It only has ~36% more Bandwidth than the 4080 SUPER and even the L2 Cache is limited to 72MB (out of 96MB for a full AD102) ! The 4080/SUPER have 64MB so the 4090 only has 12.5% more L2 Cache instead of 50% if it had the full L2 Cache of AD102...

The same way the 5090 has 96MB L2 Cache but a full GB202 has 128MB L2 Cache ! Nvidia are gimping all x90 GPUs too.