I'm interested to get some comparison points from the community on their poclmembench scores (GitHub - kruzer/poclmembench: calculates your gpu memory speed ).
My R9 270 incorporates 2GB of Epilda W2032BBBG 6A-F GDDR5 and achieves:
148.8 GB/s @ 925/1375 MHz (game stable) 777133200000000029BD472A60550C0F231D87030044840022AA1C085C0B14204A8900A000000120110D20234A1D2411
170.3 GB/s @ 1050/1575 MHz (game stable)
180.4 GB/s @ 1200/1625 MHz (unstable)
I'd like to see some scores from other Pitcairn, Tonga and Polaris users (7870, R9 270/370/285/380, RX 470/480/570/580), noting memory type, whether the memory timings are stock or modified, and what clocks are used. I've seen scores for Polaris but not other Pitcairn and Tonga based GPUs.
Happy to take comments and suggestions re: Epilda GDDR5 custom timings.
Yes that issue is discussed here: RX 580 8G Last GB of VRAM is slow..? Reminds me of GTX 970 last 0.5GB | guru3D Forums
Seems to be the consequence of memory swap issue.
Also in that thread i note that a GTX 970 (7GHz GDDR5) bench comes in at 145.8 GB/s.
Polaris poclmembench tests i've seen so far... 159.1, 184, 191, 192.7, 200, 205.6 GB/s
The low numbers seen for some memory chunks on some GPUs is likely not representative of actual memory performance. However, lets not throw the baby out with the bathwater. Putting these outliers aside, Poclmembench indicates about 90% achievement of theoretical peak bandwidth is achievable on Pitcairn given a high enough GPU core clock to GDDR5 frequency (75%+). Poclmembench is a repeatable quick to run benchmark that is very sensitive to GPU clocks and timings, which makes it rather useful to those people tuning their timings and as a yardstick to assess & compare different memory systems.
Updated original post.