AMD Kaveri: Open-Source Radeon Gallium3D vs. Catalyst 14.12 Omega Driver

[dimsam]

here, my results from an oc kaveri

http://openbenchmarking.org/result/1412161-LI-20141216K75

[tomtomme]

RANK has no impact on performance.
Maybe you wanted to write single/dual channel?

wrong

http://www.techpowerup.com/197145/amd-a-series-kaveri-performs-better-with-dual-rank-dimms-report.html

[bridgman]

RANK has no impact on performance.
Maybe you wanted to write single/dual channel?

AFAIK dual vs single channel has a big impact on performance, while dual vs single rank has a smaller impact on performance but still makes a difference.

[drSeehas]

wrong
http://www.techpowerup.com/197145/am...ms-report.html

Thank you for the link.
Did you read some of the comments?

techpowerup refers to

http://www.computerbase.de/2014-01/amd-kaveri-arbeitsspeicher

Again: Did you read some of the comments?

4 ? dual rank wasn't tested.
Even on page 2 you can read "sich bei reiner CPU-Beschleunigung keine Unterschiede zwischen Single-Rank- und Dual-Rank-Speicher ergeben".

Interesting link:

https://en.wikipedia.org/wiki/Memory_rank

I correct myself: RANK has VERY little impact on performance.

[bridgman]

I correct myself: RANK has VERY little impact on performance.

Yeah, IIRC having >1 rank allows you to interleave accesses on a single channel. If we didn't have burst transfers that would make a huge difference (maybe 50% higher throughput) but since memory controllers do long bursts into cache wherever possible the benefit of dual (or quad) rank is reduced as the average burst length goes up.

I'm pretty sure that using dual rank parts also allows different pages to be open on the two ranks, which improves your page hit rate.

[drSeehas]

Yeah, IIRC having >1 rank allows you to interleave accesses on a single channel. ...

ComputerBase tested 4 x SR which results in more than 1 rank per channel. This was 5 % slower than 2 x DR.
Any explanation?

[bridgman]

ComputerBase tested 4 x SR which results in more than 1 rank per channel. This was 5 % slower than 2 x DR.
Any explanation?

I haven't been involved in memory controller design for a long time, but even back in the dark ages there was a fair amount of guessing (called "prediction" these days, which sounds more scientific ) about what was likely to happen next, and if you didn't guess right you could slow things down by enabling "performance" features. Best guess is that either the workload used for those tests wasn't representative of typical workloads and screwed up the predictions, or could just be that 4x rank isn't supported as well as 2x.

Another possible explanation is that multi-rank DIMMs load down the bus just like multiple DIMMS, so it's possible the memory had to run more slowly in order to stay reliable with the higher bus load. Would have to go through the article in detail and don't really have time at the moment.

[Ansla]

From personal experience RAM overclocking (2400Mhz is considered overclock for A10-7850K) rarely works with 4 DIMMs, it certainly doesn't for me. So if you want best performance definitely stick to just 2 DIMMs.

And even at the same speed my A10-7850K appears slower for both CPU and GPU intensive tasks then when it was using 2 DIMMs. But I didn't spend time benchmarking it with 2 DIMMs after I finally got it working at 2133Mhz with all 4 DIMMs. Just compared current state with some older benchmarks so it might as well have been software regressions since then.

[drSeehas]

... 4x rank isn't supported as well as 2x. ...

??? In both cases (4 x SR, 2 x DR) there are 2 ranks per channel.

Another possible explanation is that multi-rank DIMMs load down the bus just like multiple DIMMS, so it's possible the memory had to run more slowly in order to stay reliable with the higher bus load. ...

But wouldn't the results in this case (total 4 x SR or 2 x DR) be equal and not differ as they do in the ComputerBase tests?

[bridgman]

??? In both cases (4 x SR, 2 x DR) there are 2 ranks per channel.
But wouldn't the results in this case (total 4 x SR or 2 x DR) be equal and not differ as they do in the ComputerBase tests?

Yeah, I didn't recognize "4 x SR" as meaning "4 single rank DIMMs" and took it to be some shorthand for a 4-rank DIMM. Remember I hadn't actually read the article.

Don't know why 4 x SR performed midway between 2 x SR and 2 x DR unless the BIOS downclocked the memory controller a bit when it recognized the presence of 4 DIMMs (4 x SR would probably add more capacitance to the lines than 2 x DR).