[Simbelmynë]

Memory timings play a huge role with the Ryzen architecture.


I am currently trying to figure out what timings will be optimal for my 3700X.


Base system: Ryzen 7 3700X, ASUS Prime X570-pro, Linux: Solus 4.0, OpenFOAM 7
2x8 GB KFA2 Hall of Fame 4000 MHz (Samsung b-die) single rank. SMT On.


No overclocking on the CPU other than what AMD applies automatically.


The memory runs fine up to 4400 MT/s, but the secondary and tertiary timings applied by the BIOS (AUTO) will be really poor. Also, since the Infinity Fabric stops at about 1800 MHz (x2) it will have to run at 2:1 ratio (so that is 1100 MHz (x2) when the memory is at 4400 MHz). This has a negative impact on the results, but I do not know how much. Currently I suspect the extremely poor secondary and tertiary timings to be the main culprit.


Anyways, the best result so far is running the memory 3200 MT/s @ 14-15-15-35 with many of the secondary and tertiary timings locked to the default 2133 MHz profile.

It finishes the benchmark, using 6 cores, in 222 seconds.

This is quite impressive for single rank memory considering that the 8700k finished the benchmark in 247 seconds using 3200 MT/s dual rank memory. *)

The sweet-spot is 3733 MT/s according to AMD (if the infinity fabric can manage 1866) so there might be substantial gains still to be had.

I am impressed with what AMD has managed with this release. It is likely that Ryzen 5 3600 can obtain the same results to a very very competitive price.

Anyone else that can share some results with the new Ryzen chip?


*) To be fair, the 8700k result is on OpenFOAM 6 and also on a potentially slower Linux. But then again, the results were before Meltdown etc. so there may be some penalty for that as well.

[flotus1]

Thanks for the results, I was almost ready to order a system myself to do some testing, before upgrading to Epyc Rome. Seems to be looking good so far.

Also, since the Infinity Fabric stops at about 1800 MHz (x2) it will have to run at 2:1 ratio (so that is 1100 MHz (x2) when the memory is at 4400 MHz). This has a negative impact on the results, but I do not know how much.

This is one of the things that I am most curious about. Maybe you could run the same memory speed and timings (e.g. the ones you found @DDR4-3200) and compare 1:1 Infinity fabric against 1:2

[Simbelmynë]

Originally Posted by flotus1

This is one of the things that I am most curious about. Maybe you could run the same memory speed and timings (e.g. the ones you found @DDR4-3200) and compare 1:1 Infinity fabric against 1:2

I will try!


Right now I am trying to go below 200 seconds in the benchmark


3400 MT/s @ 14-15-15-35 ---> 211 seconds


3533 MT/s @ 14-15-15-35 ---> 204 seconds


3600 MT/s @ 14-15-15-35 ---> not stable



This chip is a beast though. I got the hardware last night and there are still many issues with the Bios and with Linux compatibility so this will only improve I reckon.


(Also, July 29 will be interesting)

[Simbelmynë]

So with the Ryzen DRAM calculator v 1.6.0 released today I got some fantastic results.

Running 3600 MT/s @ 14-14-14-28 (and all other sub-timings really tight) I managed 181 seconds on the benchmark (re-ran 3 times as usual)! This means that the 3700X can almost rival a Threadripper 1950 or an Intel 7940X running 3200 MHz memory in dual rank. I imagine that the same results can be had with a Ryzen 5 3600 on a an old B450 board (with BIOS update). What a fantastic value for light workload CFD!


I did also run a loaded latency check on the settings with the following results

Originally Posted by flotus1

This is one of the things that I am most curious about. Maybe you could run the same memory speed and timings (e.g. the ones you found @DDR4-3200) and compare 1:1 Infinity fabric against 1:2

I have not managed to run the IF at a lower ratio when the memory is <3600 due to stability. However, running IF @ 1866 MHz when memory is @ 4200 MHz (cl 16) gave nearly identical results (182 seconds) as when IF is @ 1800 MHz and memory is @ 3600 MHz (181 seconds). The results are reasonable if we disregard IF. This gives an indication that the IF speed may not be detrimental when running with an offset. With Rome I guess this is a no-issue since it will most likely be 1:1 always.



Imagine if RDIMMS would work in 3600 MT/s (cl14) on ROME.

[aparangement]

Originally Posted by Simbelmynë

Memory timings play a huge role with the Ryzen architecture.





Anyways, the best result so far is running the memory 3200 MT/s @ 14-15-15-35 with many of the secondary and tertiary timings locked to the default 2133 MHz profile.

It finishes the benchmark, using 6 cores, in 222 seconds.

Would you mind post the detailed list of memory timing?

I tested the ryzen3600 on ddr3600c14-15-15-30 with tRC48, however finishing the benchmark with 6 threads (HT-OFF) would take 247s, not a very impressive advange over 2700x, which could finish in about 265s with 3200c14. I am wondering if other timings have major impact.

Thanks a lot!

[Simbelmynë]

Originally Posted by aparangement

Would you mind post the detailed list of memory timing?

I tested the ryzen3600 on ddr3600c14-15-15-30 with tRC48, however finishing the benchmark with 6 threads (HT-OFF) would take 247s, not a very impressive advange over 2700x, which could finish in about 265s with 3200c14. I am wondering if other timings have major impact.

Thanks a lot!

Those are old results.


Since Ryzen DRAM Calculator v1.6 became available I use that ("fast" setting). You will get a huge improvement when all sub-timings are set properly (I did at least!). Anyways, I manage less than 180 seconds on the benchmark.


If you do not wish to use the DRAM Calculator then I suggest setting the memory to the default 2133 profile and reboot. Then you can lock the timings (i.e. remove "Auto") and gradually increase the frequency (and voltage to the memory).

[aparangement]

Originally Posted by Simbelmynë

Those are old results.


Since Ryzen DRAM Calculator v1.6 became available I use that ("fast" setting). You will get a huge improvement when all sub-timings are set properly (I did at least!). Anyways, I manage less than 180 seconds on the benchmark.


If you do not wish to use the DRAM Calculator then I suggest setting the memory to the default 2133 profile and reboot. Then you can lock the timings (i.e. remove "Auto") and gradually increase the frequency (and voltage to the memory).

Thanks for the suggestion.

Actually I was using the DRAM Calculator, but previously only control the first 6 timings (until tRC).

I just finish one run with all timings set accoring to the "FAST" suggestion, it's 212.34s @ddr3600. That's very sensible improvement compared with that only leading timings being controled.

To me this is not very good news since playing with timings takes time, and capabilities of DRAMs from different vendors vary a lot.

By the way, even the "FAST" suggestion form DRAM calculator does not go as low as 14-14-14-28, so your settings is basically from AMP at ddr2133?

[Simbelmynë]

Here are the timings according to v1.6.0.3. I think I have v1.6.0.1 at home though and there might be some difference. Also, don't forget FCLK and other settings

[aparangement]

Originally Posted by Simbelmynë

Here are the timings according to v1.6.0.3. I think I have v1.6.0.1 at home though and there might be some difference. Also, don't forget FCLK and other settings

I have beeing using 1.6.3, but I think the difference comes from "R-XMP"...

I use the calculator on my laptop running Winsows 10, rather than the computer running the benchmark. But I am not sure about the latency inputs like "tCL(CAS) ns, tRCDWR ns" etc, So I click the "R-XMP"... the calculator does read something but I guess that's totally irrelevant ...

So you run the calculator on the computer you tested to get the latency settings?

[Simbelmynë]

Originally Posted by aparangement

So you run the calculator on the computer you tested to get the latency settings?

No, I have the calculator on a different machine.


If you wish, then I may be able to check my settings tonight. Most of it (if not all) correlate with what I just posted though as far as I remember.

[aparangement]

Originally Posted by Simbelmynë

No, I have the calculator on a different machine.


If you wish, then I may be able to check my settings tonight. Most of it (if not all) correlate with what I just posted though as far as I remember.

Thanks a lot if it is not too much trouble.

For now I could just use the settings you just post for testing.

[Simbelmynë]

Here are the v1.6.0.0 results.