Review Ryzen 3700x Memory speed scaling

[Head1985]

I just wanted know how my 3700x scale with memory speed in cpu bottleneck scenes.So i tested it
Test system:
3700x oc to 4250Mhz
2x16GB ram dual rank kingston predator 3333mhz.Subtimings same with all speeds.Also its running in 1:1:1.
1080TI 2040/12600
Resolution 1920x1080
Aida:
2133mhz

3200mhz

3600mhz

3800mhz

 

[Head1985]

 

[lopri]

3600 CL16 sweet spot! Amazing work.

 

[lopri]

Subtimings same with all speeds

Are you sure about this?

Also, it saddens me that my Gigabyte X570 board cannot do 1900 fclk in any stable manner when a B450 board can pull it off.

 

[Kenmitch]

3200MHz or better doesn't make much difference.

Wonder if 3600MHz cl14 b-die with tweaked timing would?

 

[Head1985]

Are you sure about this?

Also, it saddens me that my Gigabyte X570 board cannot do 1900 fclk in any stable manner when a B450 board can pull it off.

Yes subtimings are same.Only TRFC is different.But TRFC number is in graph

 

[DrMrLordX]

Hmm, those subtimings must be pretty loose. I used the latest 1usmus DRAM calculator for DDR4-3733, dialed clocks back to DDR4-3600 (due to weird system behavior above that speed), and got:

CPU CPU Clock Motherboard Chipset Memory CL-RCD-RP-RAS Read Speed
12x Ryzen 9 3900X HT 4375 MHz [ TRIAL VERSION ] X570 Dual DDR4-3600 14-16-14-28 CR1 56456 MB/s

CPU CPU Clock Motherboard Chipset Memory CL-RCD-RP-RAS Write Speed
12x Ryzen 9 3900X HT 4375 MHz [ TRIAL VERSION ] X570 Dual DDR4-3600 14-16-14-28 CR1 55492 MB/s

CPU CPU Clock Motherboard Chipset Memory CL-RCD-RP-RAS Copy Speed
12x Ryzen 9 3900X HT 4375 MHz [ TRIAL VERSION ] X570 Dual DDR4-3600 14-16-14-28 CR1 57870 MB/s

CPU CPU Clock Motherboard Chipset Memory CL-RCD-RP-RAS Latency
Ryzen 9 3900X 4375 MHz [ TRIAL VERSION ] X570 Dual DDR4-3600 14-16-14-28 CR1 65.9 ns

Also watch out for performance degradation at speeds above 3600. My CBR20, 15, 11.5, and 10 results all tanked hard even when I booted with stability using DDR4-3666. I think if you push IF speed too high, things get . . . weird.

 

[lopri]

It's the Gigabyte BIOS. I am pretty sure about it. I plan to get an ASUS X370 board on the cheap just to see how much time I have wasted fighting Gigabyte's BIOS.

 

[lopri]

For games, it seems to me what matters is latency. Not the CAS ratings, but the access time as shown by AIDA. It is also why Intel still has an upper hand in games, thanks to its vastly superior memory access latency.

To be sure, every game is different and some games are not as sensitive to latency. And Zen 2's massive L3 and improved prefetch do an excellent job hiding latency penalty.

This reminds me of the time when AMD CPUs had direct memory access while Intel CPUs had to deal with FSB and north bridge on the motherboard. Back then Intel combated it with larger caches as well. Yet the Athlons had upper hand in games even with smaller caches.

 

[Concillian]

FYI, I've been doing my own memory tuning, and at any speed there is a pretty decent performance improvement in CPU limited scenarios by improving subtimings. I can't test many games for this, as my bench system has a lowly HD7850 in it right now, so getting into CPU limited scenarios in games is a bit difficult until I incorporate the Ryzen 3k into my main system. However, Shadow of the Tomb raider shows an "If the CPU were the limiting factor" FPS report... and in this report, mild tuning of the tRC, tRFC, tWR, tFAW and the 2 xSC_L sub-timings resulted in ~6-7% improvement in both average and 5% low FPS at both DDR4-3600(1800 Fabric) and DDR4-3800 (1900 Fabric). It was actually larger than the benefit in going from 3200 to 3600.

AIDA latency didn't really change from the subtimings, but read / copy performance got marginally better.

Sensitivity is obviously higher in my specific testing, as that benchmark completely ignores GPU limitations, and in most real-world conditions we'll be GPU limited some, if not most, of the time, but it was still surprising to me how big the difference was from sub-timings.

 

[Head1985]

Hmm, those subtimings must be pretty loose. I used the latest 1usmus DRAM calculator for DDR4-3733, dialed clocks back to DDR4-3600 (due to weird system behavior above that speed), and got:

CPU CPU Clock Motherboard Chipset Memory CL-RCD-RP-RAS Read Speed
12x Ryzen 9 3900X HT 4375 MHz [ TRIAL VERSION ] X570 Dual DDR4-3600 14-16-14-28 CR1 56456 MB/s

CPU CPU Clock Motherboard Chipset Memory CL-RCD-RP-RAS Write Speed
12x Ryzen 9 3900X HT 4375 MHz [ TRIAL VERSION ] X570 Dual DDR4-3600 14-16-14-28 CR1 55492 MB/s

CPU CPU Clock Motherboard Chipset Memory CL-RCD-RP-RAS Copy Speed
12x Ryzen 9 3900X HT 4375 MHz [ TRIAL VERSION ] X570 Dual DDR4-3600 14-16-14-28 CR1 57870 MB/s

CPU CPU Clock Motherboard Chipset Memory CL-RCD-RP-RAS Latency
Ryzen 9 3900X 4375 MHz [ TRIAL VERSION ] X570 Dual DDR4-3600 14-16-14-28 CR1 65.9 ns

Also watch out for performance degradation at speeds above 3600. My CBR20, 15, 11.5, and 10 results all tanked hard even when I booted with stability using DDR4-3666. I think if you push IF speed too high, things get . . . weird.

Subtimings are for hynix c-die dual rank memory.Its even little better with cl16-17-17-19 and 450TRFC and alt.But rest subtimings is same like in dram calculator.

 

[lopri]

Wait, so your sticks are not even b-die stuff? Wowowow.

 

[Head1985]

I have kingston predator 2x16GB 3333mhz dual rank ram.I dont think samsung b-die even exist as dual rank 16GB.They are only single rank max 8GB.
Btw dual rank is like 5% faster than single rank at same clock so i am not complaining

 

[DrMrLordX]

@Head1985

There is some dual-rank 16GB b-die, but it's not that interesting compared to Samsung M-die or upcoming A-die. So at this point I would not waste money on it.

 

[Kedas]

Looks like performance is much more dependent on good memory sub-timing than the IF speed.

So no matter the memory speed you have trying some better sub-timings settings can give you a boost if your game is CPU limited. DDR4-3000 can get DDR4-3600 performance in some cases with better settings, obviously you can also tweak the DDR4-3600 again to be better.

 

[Head1985]

He is using 3800mhz cl14-14-15-28-303.Thats crazy insane timings.I am not surprised he gets some perf gain vs all auto timings

 

[Zstream]

Well these cheap ballistic sports are at 3600@1.375v. No idea on timings, as I played with a few and the damn thing refused to boot. I thought it would at least reset and ask me to go back in the bios...