Given CUDIMM support potentially relieving timing constraints I wonder: does Arrow Lake’s CUDIMM support enable dual dimm per channel support at meaningful DDR5 speeds?
Being able to run productivity workloads with 192GB (or even 128GB) might be a bonus for a CPU which has little else to go for it…
Especially since the competition seems to be sitting that one out.
I have not yet seen any dual rank CUDIMM modules. It remains to be seen. Perhaps CUDIMM does not solve the bottleneck for for dual rank modules?
According to https://www.intel.com/content/www/us/en/content-details/832586/intel-core-ultra-200s-series-processors-datasheet-volume-1-of-2.html (Table 49. DDR Support Matrix Table): the officially supported spec for 2DPC is 4800 for 1R DIMMs and 4400 for 2R DIMMs regardless of UDIMM or CUDIMM.
That’s up from 4000/3600 on Raptor Lake.
Oh, and 5600 is officially supported on a single DIMM on 2DPC motherboards now, up from 4400 
.
I got up to 8400 using cudimms on a unify with 2 dimms on 1:2 mode with really tight timings driving 120GB/s
I need an additional identical set to test 2DPC, I believe it has a really strong memory controller
120GB/s doesn’t sound half bad, perhaps I wouldn’t quite need that much, but the 40-50Gb/s you get 2DPC on Ryzen 4 certainly are unerwhelming.
If you do get another set, I’d love to hear your results!
Thanks for the link!
That indeed looks rather sad, and as if there was basically zero advantage via CUDIMM for the 2DPC angle. I guess it’s not really the clock, but the impedence, drive currents, capacitance or the sheer amount of additional metal you need to herd electrons through, which limit what you can do there.
Kinda pointless as you never run 2DPC with a single DIMM 
He meant only using 1 of each channel, but you knew that.
Raptor Lake and prior were officially only good for 4400 or something on any MB that had 4 holes.
Yes, 4400. I’ve a lot of workloads which go DDR bound with tightly timed 5600 and a 9900X, so Arrow’s relevant in ways Raptor… just isn’t.
Assume you meant GB/s. I got 50 GB/s with untightened 4x48GB B-die. Haven’t specifically benched with Raphael but, from entry level benches at 74 GB/s with Granite Ridge at DDR5-5600, I’d expect 2x48 M-die tightened at 5200 with 3.9 μs tREFI’d be over 60.
Haven’t seen a lot for Arrow overclocks yet but it’s looking like probably ~120 GB/s at 8000. Granite Ridge’s more like ~95 GB/s. So ~94% unidirectional bus utilization from Arrow up in the 1R CUDIMM zone rather than ~73%. That’s not a new disparity in the benches and the past several Zen generations have been competitive anyways, so I’m unsure how much importance to place on the numbers.
Also, my experience is latency and bandwidth reported by DDR benches are fairly responsive to refresh. So they’re partially a measure of DIMM cooling and thus whether it’s a dual or quad DIMM build with single tower air, conventional dual air, setback dual tower (Fuma 3, Royal Knight 120), or CPU water with DIMM prevailing air, downflow air, water, or occasionally something more exotic.
Which CPU’s under the cooler’s mostly orthogonal to DIMM temperatures and thus to tREFI, though Arrow being more air friendly than Intel 7 parts is potentially a factor.
2R non-CUDIMM isn’t getting to speeds where 1R really benefits from CUDIMM so, yeah, probably not. Maybe 4x24GB 1R CUDIMM would do something meaningfully different 2x48GB 2R non-C but the odds of improvement don’t seem great from what I know.
This topic was automatically closed 273 days after the last reply. New replies are no longer allowed.