FRONTPAGE AMD Ryzen 9 3900X and Ryzen 7 3700X CPU Review

[Johan45]

8086k stock, delid+LM, Noctua D15 under same load runs around 55C. I could try a bigger cooler on the 3600. I have other Noctuas in other systems so maybe a straight swap to bigger would help.

The new AGESA might make that better Mack

 

[mackerel]

The new AGESA might make that better Mack

I'd try it if either of my Ryzen mobos offer it. I'm checking on a daily basis, although I'm more hopeful of a ram speed increase as currently it is running worse than Zen+.

 

[Blaylock]

I'm trying to digest this memory bus/infinity fabric decoupling a little further. From the DDR4 optimal settings graph it mentions the decouple takes place at 3600Mhz. This is the recommended price/performance spot. It also goes on to show lower latency at 3733Mhz and claims this as the "sweet spot".

Are they recommending overriding the 2:1 mode and OCing to 3733Mhz for maximum performance? At these speeds is the extra 133Mhz and 2ns really noticable or would this be primarily for benchmarking results?

Sorry if this was answered before.

 

[EarthDog]

I believe it changes at 3733 and goes to 1:2 with the IF. I believe AMD said DDR4 3600 is the sweetspot??

 

[Woomack]

2ns is like an error margin. Run benchmark a couple of times and you will see. If IMC is not scaling much above 3600 then not much will help with significant improvement. So far most memory setting comparisons were showing barely any differences. I mean something like 3200 CL14 or 3600 CL16 will be 5FPS better than 2666 CL16 ... as long as you play in 1080p or below but it's like 150 vs 155 FPS so not many users really care.
I guess that everyone has to test it on their own motherboard as there will be also differences in max clock or additional tune options.

 

[Blaylock]

"Once the memory speed surpasses 3600 MHz the memory bus and infinity fabric “decouple” and operate in a 2:1 ratio, this is automatic."

 

[Woomack]

I see no problem here ... pass ~5200 mem clock and should be faster

 

[Johan45]

I'm trying to digest this memory bus/infinity fabric decoupling a little further. From the DDR4 optimal settings graph it mentions the decouple takes place at 3600Mhz. This is the recommended price/performance spot. It also goes on to show lower latency at 3733Mhz and claims this as the "sweet spot".

Are they recommending overriding the 2:1 mode and OCing to 3733Mhz for maximum performance? At these speeds is the extra 133Mhz and 2ns really noticable or would this be primarily for benchmarking results?

Sorry if this was answered before.

I believe it changes at 3733 and goes to 1:2 with the IF. I believe AMD said DDR4 3600 is the sweetspot??

It goes 2:1 after you pass 3600 MHz, it is forceable as I posted yesterday but depends on the CPU and board I would assume. As far as what's the "best" memory. Like Woomack said 3600-3800 XMP would make very little difference in performance day to day. It's not the Infinity fabric, which ties everything together, that changes speeds but the memory controller and running 2:1 has a noticeable but slight impact.
Another test with 4x8GB at 3600. Popped them in hit XMP and bam right into Windows. Not stability tested but ran GB4 without a problem and with good results.

 

[EarthDog]

So to be clear, at 3733 it changes the memory controller to 2:1, not 3600, correct? Does it do it at 3700 if that is an option?

 

[Johan45]

It does it when it passes 3600, that's the line AFIK. I can do some testing this morning (my day suddenly opened up) to be sure of the cut-off but the way I read it at 3601 it decouples
Right now testing 4x8 GB at 3800 CL16

 

[EarthDog]

Gotcha, I thought it was the memory multiplier that did it and why I mentioned 3733.

But I believe AMD themselves said 3600 sweetspot. After it decouples, it is then slower needing a more significant bump for more performance.

I wonder if this is by board or in the CPU somewhere (3601 vs 3733 like AMD slide shows).


PS - Boards and processors land today for me... I am really excited to play with this stuff!!!!!

 

[mackerel]

Can someone do a set of bandwidth measurements at various ram speeds? Even if the latency increases, I'm wondering if the bandwidth continues to scale in 2:1 region.

 

[Johan45]

Gotcha, I thought it was the memory multiplier that did it and why I mentioned 3733.

But I believe AMD themselves said 3600 sweetspot. After it decouples, it is then slower needing a more significant bump for more performance.

I wonder if this is by board or in the CPU somewhere (3601 vs 3733 like AMD slide shows).


PS - Boards and processors land today for me... I am really excited to play with this stuff!!!!!

I know that it can be board dependant and relies on good BIOS coders. When I was tesing on the X470 CHVII it would decouple well before 3600 which to me was just wrong. So I let them know, not that it'll change right away and it's easily worked around by forcing the controller to run 1:1. The MSI and ASUS are slightly different in how that's done but they both have the ability.
One more thing of note, when pushing the IF/mem controller past 1800 MHz (3600 DDR) it can become unstable and the IF has error correction. What you'll see is a drop in MT benchmark scores even though you're running faster RAm. There are two voltage adjustment that can help with that. One is the SOC and the other which feeds from the SOC is the VDDG. It's best to keep the SOC 100 mV over the VDDG since it feed from that rail. It took me a bit to get 4x8 gb 3800 Cl16 behaving properly but between balancing those rails and the DRAM voltage I finally have a result and currently Testing stability with AIDA64.
So to answer Blay's question. This is the difference between 3600 and 3800 same timings just a speed bump.



- - - Auto-Merged Double Post - - -

Can someone do a set of bandwidth measurements at various ram speeds? Even if the latency increases, I'm wondering if the bandwidth continues to scale in 2:1 region.

It does it's just a drop in latency

 

[Johan45]

4x8 GB at 3800 is going to take some ore tweaking to stabilize, lasts about 10 minutes in AIDA before it errors out. That'll be for another day

 

[Kenrou]

But I believe AMD themselves said 3600 sweetspot. After it decouples, it is then slower needing a more significant bump for more performance.

No, AMD said the sweet spot is 3733mhz which confused me to no end because of the controller memory changes.

 

[Johan45]

No, AMD said the sweet spot is 3733mhz which confused me to no end because of the controller memory changes.

For stability and ease it's set at 3600 in BIOS if you're adventurous you can get 3733 as I believe that AMD felt in "most" cases it would be stable but it takes a manual setting to keep it at 1:1

 

[bassnut]

Give me a couple days and we'll see how asrock handles it too. I got lucky and grabbed 1 of 6 .... 3900X at canada computers last night will have it Saturday and a X570 TIACHI landing tomorrow.

 

[Dolk]

I've only been on the x370 boards, I don't remember seeing a RAM Ratio or NB frequency manipulator. Did something change with the latest boards or am I just getting old?

 

[Blaylock]

It's internal in the CPU. Above 3600Mhz the memory controller and infinity fabric decouple.Just read this page and it spells it all out.

 

[Dolk]

Thats what I got from here, I just wanted to confirm I didn't miss anything.