AMD Trinity Linux Memory Performance

[Ansla]

I find it very odd that AMD has been sending sites like Anandtech etc. 1866MHz memory along with these APUs.

The highest RAM frequency officially supported is 1866MHz so it's not surprising that was sent to review sites. What's surprising is that 2133MHz is not officially supported, even though it works.

[dnebdal]

Michael (or anyone else), is there any visible reason why DDR3 1866 MHz RAM was sometimes on lower performance than the 1600 equivalent? On most tests it showed a relatively "linear" behaviour, but on others it left the queue with a jump.
I guess all of them were default and configured by SPD on chip/bar?

At a guess, the timings don't work out ideally with other clocks in the data path? If there is a 2:3 (or 5:4) ratio in there somewhere, it could make the performance a bit more weirdly sensitive to instruction timings and such.

[kukreknecmi]

What about Trinity / Piledriver with GCC using -bdver1 -bdver2 ? Will they have any positive effect?

[Pahanilmanlintu]

I find it very odd that AMD has been sending sites like Anandtech etc. 1866MHz memory along with these APUs.

[Adarion]

Michael (or anyone else), is there any visible reason why DDR3 1866 MHz RAM was sometimes on lower performance than the 1600 equivalent? On most tests it showed a relatively "linear" behaviour, but on others it left the queue with a jump.
I guess all of them were default and configured by SPD on chip/bar?

[Michael]

equally clocked 4-core bulldozer FX-4100 series.

I have no FX-4100...

[waucka]

Are you doing cpu test next? It could be interesting to see how this apu's 4-core piledriver fares vs equally clocked 4-core bulldozer FX-4100 series.

As an owner of an FX-4100 series CPU, I would be very interested in such a test.

[droidhacker]

For quite some time almost all AMD and Intel desktop chips have had dual channel memory. There was the i9x0 family that had triple channel and there is the uber-expensive SB-E with four, but most chips you're going to see have two channels. The Atom, the 'C', and 'E' series chips are all single channel.

The Opteron chips with more than two memory channels are dual-die (two chips in one package), so it's not much different than a dual socket MB with two channels/processor. It's just all in one nice (huge) package, so it makes 8P system easier to build for the hardware vendors.

We're not likely to see an 8 channel chip in the near future--at least until memory starts to get integrated onto the processor package itself. That's going to be a strange day. Of course, some of the upcoming Intel Haswell variants are expected to have *some* memory on the package in a separate die--much like the old Pentium-Pro and PII/PIII era Slot-1 processors did. Oh, and the Slot-A from AMD was that way in the beginning, too.

That said, AMD can certainly improve its memory controllers. In similar processors, Intel has had a lead in bandwidth and latency in its memory controllers at least since the Core2 family came out. DDR4 is coming in a year or so, so that will bring higher memory clock speeds and some improvements in bus efficiency.

Eh, one die, two die, red die, blue die.

What we need is more memory bandwidth. Period. The northbridge can handle it.

[tuke81]

Nice review Michael, thanks. 8GB 1600MHz kit cost something about 30? in Europe, the same as 1333 and 1060. So obviously it's slowest ram that one should buy for these apus. Faster rams cost some 10? more, so if one wants cheap performance boost, buy faster rams.

That 1866 MHz rams gives some odd results, lost couple of times even 1333MHz rams how so?

Are you doing cpu test next? It could be interesting to see how this apu's 4-core piledriver fares vs equally clocked 4-core bulldozer FX-4100 series.

[willmore]

WOW!

Ok, finally a benchmark/test out of phoronix that offers actual valuable information!!!!
But... late.

http://hexus.net/tech/reviews/cpu/46073-amd-a10-5800k-trinity-needs-faster-ram/

It seems to be a rather linear dependency that the chip has on the memory. Looks to me like AMD may have a very solid platform, but needs to focus more on memory bandwidth. These AMD chips are still dual channel aren't they? And Intel runs triple channel on some of their chips? With AMD, you need to step all the way up to an Opteron 6000 to break over 2 memory channels (Opteron 6000 are 4-channel rectangular chips, lots of pins!).

How about it AMD? 8-channel desktop chips?

For quite some time almost all AMD and Intel desktop chips have had dual channel memory. There was the i9x0 family that had triple channel and there is the uber-expensive SB-E with four, but most chips you're going to see have two channels. The Atom, the 'C', and 'E' series chips are all single channel.

The Opteron chips with more than two memory channels are dual-die (two chips in one package), so it's not much different than a dual socket MB with two channels/processor. It's just all in one nice (huge) package, so it makes 8P system easier to build for the hardware vendors.

We're not likely to see an 8 channel chip in the near future--at least until memory starts to get integrated onto the processor package itself. That's going to be a strange day. Of course, some of the upcoming Intel Haswell variants are expected to have *some* memory on the package in a separate die--much like the old Pentium-Pro and PII/PIII era Slot-1 processors did. Oh, and the Slot-A from AMD was that way in the beginning, too.

That said, AMD can certainly improve its memory controllers. In similar processors, Intel has had a lead in bandwidth and latency in its memory controllers at least since the Core2 family came out. DDR4 is coming in a year or so, so that will bring higher memory clock speeds and some improvements in bus efficiency.