Ryzen 3000

[zombie67]

Got it. Thanks!

[zombie67]

10980XE is no where to be found yet.

So what's the latest on Cascade Lake-X availability? I have been thinking of getting a second AVX-512 machine. They really do outperform by quite a bit on projects where it is supported, and being first counts (PG). All I can find are 9th generation Skylake-X. And when I search on Cascade Lake-X availability, all the articles are back from November.

[zombie67]

So what's the latest on Cascade Lake-X availability? I have been thinking of getting a second AVX-512 machine. They really do outperform by quite a bit on projects where it is supported, and being first counts (PG). All I can find are 9th generation Skylake-X. And when I search on Cascade Lake-X availability, all the articles are back from November.

I guess Intel isn't talking?

[John P. Myers]

It was launched months ago but Intel doesn't want to sell any because they're having to sell them at half the price of the 9980XE, which is all the 10980XE is anyway. Just renamed to sound new.

[zombie67]

It was launched months ago but Intel doesn't want to sell any because they're having to sell them at half the price of the 9980XE, which is all the 10980XE is anyway. Just renamed to sound new.

Heh. When you're losing an argument, I guess it's best to just shut up.

The good news is that the new threadripper, or at least the 3970X that I have, out-performs my i9-9820X with AVX-512. And this is on LLR tasks at PG, which take advantage of AVX-512. And I am not talking about just total production. I am talking about returning tasks faster, and more of them at a time. For example, running PPS-Mega (LLR):

i9-9820X (10 core, HT off, OC to 4ghz (but AVX-512 may negate the OC)): 3-thread 5-thread and 10-thread per task take the roughly same time per task. Less threads per task, and the times go up in a linear fashion. So I run 3 3-threaded tasks at a time. Each tasks takes about 19 minutes. I experimented with both windows and linux. No difference in run times.

3970X (32 core, HT off, no OC): Running 4 threads per task, each task takes about 14 minutes. Going smaller per task and times increase linearly. Going larger per task and there is no reduction in time. But here is the important part: This was with linux. With Windows, tasks ran 2-3 times longer with the same settings. Something is clearly wrong with the way windows deals with these threadrippers, and I assume it has to do with not keeping MT tasks associated to the individual chiplets.

Summary: That is 3 tasks at 18 minutes each vs 8 tasks at 14 minutes each. Even if you equalize for the number of cores, the threadripper wins. Particularly if returning your copy of the task first is important.

[rhihannisu]

The threadrippers are monsters. Even my 2920 is a beast. so glad AMD remembered how to make good processors again.

[zombie67]

1) Regarding windows and 128 threads:

https://www.anandtech.com/show/15483...3990x-review/3

Looks like SMT/HT + 128 threads + AVX2 + windows is a problem. See the NAMD graph and comments. And the the summary:



Does Linux have this problem?

2) I beginning to think that Windows is too dumb to understand the chiplet architecture used in Zen2. For example, there are 4 chiplets of 8 cores each on a 3970X. Each chiplet contains its own cache. Scenario: (SMT/HT off) running 4 MT tasks, each task using 8 cores, it would be best if each task was contained in a single chiplet. That way cache calls would be fastest. Otherwise the task has to go out to the bus to retrieve the data elsewhere. But the poor task times I am seeing are making me think windows is not doing the smart thing.

Is there a way to make windows keep MT tasks tied to a single chiplet?

Is linux any better at this?

3) https://www.reddit.com/r/linuxmaster..._scheduler_vs/

I am going to have to try linux to see the difference. No VM this time, a separate boot device.

Turns out that linux is actually much better at affinity WRT chiplets, keeping the MT tasks within each chiplet.

http://www.primegrid.com/forum_threa...ap=true#139003