Gloomy
Golden Member
- Oct 12, 2010
- 1,469
- 21
- 81
It's going to be interesting to see how AMD's brute force HBM stacks up against big Maxwell's bandwidth efficiency improvements.. The enhanced color compression of Maxwell adds about 25% to the actual effective bandwidth, so GM200's bandwidth should be close to 400 GB/s. Also if NVidia doubles the L2 cache from 2MB to 4MB, then bandwidth efficiency will be even greater. I don't expect big Maxwell to have issues with 4K resolution considering that GM204 already competes with Hawaii in that aspect very well, even though Hawaii has significantly higher raw bandwidth.
But if R9 390x has nearly twice as much bandwidth as GM200, then it should conceivably handle 4K plus additional eye candy effects better than it's competitor.
Tonga already showcased Maxwell-level memory compression. AMD actually claims 40% and this is evidenced in tests:
With AMD maintaining their same texture units and same tex:FP32 ratio for Tonga, the end result is that R9 285s texturing performance is virtually identical to R9 280s. Ignoring any possibility of caching or bandwidth bottlenecking for the moment, R9 285 can push texels just as well as R9 280 could. [with 30% less bandwith]
70% is the best case for it. 40% is an optimistic expectation. But it is typical to see the 285 match the 280-- so in most cases it is at least a 30% efficiency improvement.Unlike our texel test, AMDs delta color compression technology introduced on GCN 1.2 has an incredible impact on R9 285s pixel throughput. This pixel test is normally memory bandwidth bound, providing something that approaches a best case scenario for AMDs compression technology. As a result despite possessing nearly 30% less memory bandwidth than the R9 280, the R9 285 tops our charts at 19.9 GPix/sec, blowing past the R9 280 by 68%. Even the R9 290 with its 512-bit memory bus and doubled ROP count still falls short here by over 3GPix/sec, or 16%.
This benchmark in a nutshell is why AMD can deliver the average performance of the Tahiti based R9 280 without Tahitis memory bandwidth. By improving their color compression to this point AMD can significantly reduce their memory bandwidth requirements on GCN 1.2, allowing them to do more with less. In real games the result wont be anywhere near this remarkable since this is a pure pixel fillrate test, but it goes to show that AMD has been able to neutralize their memory bandwidth deficit in graphics workloads.