Correct. The 2400 chips should be limited to x80 because of their superior monolithic manufacture.
There are a couple of other BCLK settings to help coax the BCLK higher like it initially being set at a lower frequency earlier on in POST and then ramping up to the desired value, or a setting that applies a different clock during a cold boot.
For me PCIe devices start having trouble around 105.8MHz… but I hadn’t adjusted by skew rates or voltages associated with them which would probably let me go higher.
Here’s a pic of Atlas, I don’t believe I posted it earlier as I was still waiting for parts. Ran short on tubing. There are a few too short runs, and longer runs due to not having spring to prevent collapse. Also not the case I wanted to use. Made it work. And work it does. Well.
Is that test mostly measuring cache throughput, or was that overclocked to break the HPET/ACPI timers? I don’t see how greater than 307GB/s is possible with 8 channels of DDR5-4800.
Nice. I bought a Define 7 and had a hard time fitting a radiator to the top as the VRM heatsink on the ASRock MB is so tall. Just barely squeezed it in there.
I recently bought a w9-3495X + Asus Pro W790E Sage SE. In one evening of tweaking I got, a very respectable given the density, 6800C32 running (VDD/VDDQ 1.4v, MC 1.35v) with 8x green stick (Kingston 48GB Hynix M 1Rx4 [KSM56R46BS4PMI-48HMI] RDIMMs but hoping to push further. A bus width of x4 doesnt seem to effect overclocking as theorised by other users in this thread, IMO single rank DIMMs are more important :
I printed my own bracket for cooling the RDIMMs as they run fairly hot. It’s a shame that there are no commercially available RDIMM compatible water blocks.
I’ve noticed a couple of major differences to regular DDR5: tREF is 32k max vs 256k, tWRWR_sg has to be ~= to tWRRR_sg for training to complete and there’s large discrepancy between read and write even after tuning tertiaries:
I also couldn’t get ASRock Timing Conf or MemTweakIt to read out the timings so apologies for the bad photos. Additionally, OCTool1121 even though it has a W790 sub menu doesn’t seem to function correctly.
With that in mind are there any resources/tools for W790/RDIMMS that people are using?
You’ve definitely delved deeper into memory OC than I have.
I haven’t found anything that can read memory/SPD info correctly. HWiNFO, CPU-Z. Memtest86 can’t seem to display anything and was reporting my read speeds as something like 33GB/s. Even if they claim to support SPR it doesn’t seem to work.
I had been using skatterbencher’s youtube videos for references, and he goes into some timing related things in his series on these cpus.
Actually hardcore overclocking also has a general ‘DDR5’ video. It’s not specific to these RDIMMS, but he goes over Hynix settings, which a lot of these modules are using. It seems the speed limits here are more related to the RCD than the quality of the chips themselves. Especially when the RCD gets hot.
Same for me. Skatterbencher’s videos have been invaluable.
Short update from me; I’ve been having issues with AVX2 workloads (prime95 etc.) on this board and couldn’t get them stable at all with low thread count workloads, whatever voltage or frequency I used. I think this might be bios related as Asus have released “PRO WS W790E-SAGE SE BIOS 1401”. For me the AVX offsets are bugged, where I need to set the AVX512 ratio offset to get AVX2 workloads to throttle down and if I set the AVX2 ratio it effects SSE workloads. Has anyone else had similar issues?
As for my memory overclock I got improved performance with 6800C34 using the following optimised timings that are both y-cruncher VST (12 hours) and memtest stable at 1.4v VDD, 1.3v VDDQ, 1.3v MC (on bios version 1202):
Anything over 6800 isn’t fully stable but that’s to be expected I suppose as this board doesn’t look like it was designed for more. Here’s the final performance:
So if you’re looking at memory for this platform considering 48GB Hynix M 1Rx4 DIMMs if you want to trade off performance vs capacity.
UPDATE Looks like my old setting for BIOS 1204 aren’t stable now , I’ve had to reduce the IMC to 1.275v and increase VDD to 1.42v to stop WHEA IMC errors from cropping up while VST testing for 1401. Not sure if that is some weird DDR5 training issue, I’ve experience those before pushing for 8000Mhz+ on other platforms
Makes sense, the DIMMs I have are using Renesas RG5R256 (Gen 2 5600 MT/s RCD), they seem to have no issues running at high speeds but reading the spec sheet they are rated up-to 70C so would assume that’s why as you shove more VDD into them they start to cause stability issues. Agree with other poster’s assertions that the impact of running dual rank DIMMs on IMC stability has, in my experience with overclocking other systems, been more of a limiting factor, so doubt it matters what RCD you have if you keep things relatively cool. Saying that I don’t seem to be having issues when running Y-Cruncher VST AVX512 at worse case temperatures:
Interestingly I bet the “micro heatsinks” improve temperatures because the RCDs seem to have a hotspot in the very centre of the IC package:
I hope they release Emerald Rapids for this platform the idle power consumption for these multi tile 3500 cpus are terrible. My system idles at 180-200w and thats without any add in cards and power saving fully enabled.
Wendel found that the first generation DDR5 RDIMMS(like the GSkill Zeta R5 6400 I have) start having stability issues around 60C - he mentioned it in his SPR video.
My system with W7-2495x, 8x64GB RAM at 4800MHz, RTX4090, one expansion card (sound card) and 8 case fans draws ~ 160w from wall socket when idling in Balanced windows power plan. Switching to high performance plan adds ~ 20 more watts.
I suspect the RAM modules contribute a lot. From what I’ve managed to google, single stick can use up to 10w of power.
I pulled half my DIMMs to see what they used. Going from 8->4, dropped idle by about 10W. I would imagine that if you have dual rank you’d be looking at more than that.
This is the highest memory bandwidth I’ve seen from any of the workstation systems, Threadripper included.
Shameless plug for the CFD benchmark thread:
If you’ve got the extra time, I’d be really curious to see what score your system could get on the CFD-EM benchmark; I’ve got a feeling you’d be able to beat out the 24 memory channel dual socket EPYC 9374F system with with yours due to the superior memory coherency plus still relatively high bandwidth.
Does anyone make heatsinks that fit these yet? It was weird that XMP DRAM didn’t come with any.
According to HWiNFO, mine are around 48C most of the time (at 6400MHz), using 60W, but HWiNFO’s numbers are dubious.
If they make a refresh without fixing these CPU’s issues, I want my money back.
BTW, what are you using for a thermal camera?
I’d love to borrow the 30Hz FLIR I use at work but I don’t think they’d let me take home a camera that goes for $10k on the used market.
I’ll take a look once I’ve got my per core overclock sorted. I think those 12 channels EPYC and 8 channel 7000 series threadripper CPUs have more synthetic bandwidth from what I’ve seen. I was fairly disappointed with this platform in that regards tbh. But I’m locked into the OneAPI so shrug.
Currently it’s 51x4, 50x16, 48x34, 44x42, 41x56 and -4x AVX2/512 with half my cores unable to run >4.6ghz AVX workloads
I’ve been testing for 2 weeks trying to get things 100% stable with corecycler for my use cases. Slow progress as a small FFT AVX2 run in prime95 take 16Mins per core. That’s 15hours for one iteration :<
I wouldn’t worry too much as the tRef or refresh interval for RDIMMs is max 32k vs UDIMMS at 256K (auto by most bioses I’ve seen is around 6-8k) so you don’t really have to worry about temperature until your >80C at load. It’s only if you’re hammering overclocked DIMMs and pushing for low latency with your timings that temperature matters.
I’d love a water block for my DIMMs though as I would be able to reduce the latency further shoving 1.65v VDD into them to see what the limit is.
I’m using a cheap Seek Thermal usb-c camera for my smart phone at around £200, you can get the Flir ones with a camera that are probably better. It’s probably not very accurate but does the job. Recommend them.
Apparently several of the existing DDR5 watercooling heatspreaders already work, but just require different thermal pads than what they come with. I have a friend that is using XTIA and another friend that is using Byski.
, I’ve had to reduce the IMC to 1.275v and increase VDD to 1.42v to stop WHEA IMC errors from cropping up while VST testing for 1401. Not sure if that is some weird DDR5 training issue, I’ve experience those before pushing for 8000Mhz+ on other platforms 
