Is Quad channel Memory in NUMA mode still REAL quad channel bandwidth?

Since activating the numa mode in threadripper platform kind of isolate memory access/location of some applications. Can we still consider that Quand channel if the numa node are using local memory set. I believe with my limited knowledge that should be named 2xDual channel even if a thread of a foreign node can have access to local memory of an other thread. Am I still wrong or each MCH of the dies have a quad channel memory bandwith within the cosmp-astrologic fabric… or we, human, just considering the memory bandwith as a whole out of the cpu package. I see plenty of video explaining that we can see ram channel as a bandwith value in a point in time, and dual channel adding a two times wider bandwith. Then in the server domain, the quad channel appeared but where controlled “independently” by two or more cpu, after godzillions of threads cycles, we, human, can now have QC in our machines at our homes. Isn’t it amazing? I still believe it is 2xDual Channel Memory called Quad for marketing.

I added 2x16GB 3200Mhz for a 64GB total. I may think I notice an improvement in general but I’m sure is not blazingly faster. I reduced my OS nvme 1TB SSD usage down to 600GB ish from 920GB usage. I think that help a lot too. it is a samsung evo 970 1TB.

Can I use a second NVMe (maybe not 1TB tough) to use as a dedicated system cache or I have to assign it to a HDD with the tool mentionned in YT L1 channel. I may just wait to the 1tb to be at better pricing.

Thank you for your reading, hope to get responses. ask me if I know this or that I you think I missed something. I’ll be happy to educate myself

1 Like

I understand your confusion, but it is still called Quad. Zen is more complicated, but similar things happen when you have dual cpu motherboards while they are running at 4, 6, 8 and so on channels. The channel quantity are more relatively called by how many separate paths there on motherboard rather than what cpu is doing with them.

While in fact the TR will be acting as 2x dual channel; it is still a quad channel CPU as whole. In terms of bandwidth your whole CPU still receives whole bandwidth of quad channel.

In terms of your os nvme, there is little to no difference from memory related performance. Any drive for different reasons will have performance boost when there’s less data on it. Less or more obvious.

2nd nvme
Sure you can get anything you want… in terms of windows that would be system page file… Smaller, and much faster nvme would be best for that role for “dedicated system cache” - don’t waste money on anything bigger than 120G unless it has to be it or you need that page file bigger. - Assign the whole drive min/max as almost total size of the nvme; and disable page file on other drives.
In terms of performance you may notice it sometimes - you may not; other than that if you already use nvme and want just to use as cache for your first nvme it makes no sense at all.

You may see some performance gain if you pin some of the system processes / or your software to certain cores from either of the chiplets vs run on all. Just make sure its running on one that is holding the memory data for it. ~ so check your spec sheet and figure out which part of the cpu is handling 2nd nvme. (it also may be more beneficial to move your system to smaller nvme and keep it with page file there; while everything else would be sitting on bigger file nvme on 2nd slot.

1 Like