Haven't seen much discussion on project configuration on Ryzen.

Did a little benchmarking recently. Stock speed 1700x (8 core 16 thread) DDR4 running at 2133 dual channel and SMT on.
Running 16 workunits produces ~67% of the work that running 8 does. I suspect running 16 properly is handicapped by memory bandwidth.

I'll do some more testing with SMT off
as well as Ram running at a more modern speed

I am doing a test now to compare virtual cores vs. real cores on two Intel Haswell machines running Win7 64-bit.
On an i7-4771, I am running WAH2 on four real cores.
On an i7-4790, I am running WAH2 on eight virtual cores.

The results thus far is that it is a dead heat. That is, the real cores are essentially twice as fast as the virtual cores, so the total output is the same. That is probably because CPDN uses code that was developed before hyperthreading came along, and can't do anything with it.

And they probably compile their stuff with Intel in mind, so AMD gets even less support for virtualization. But it would be interesting to see how well AMD does on full cores. They could be faster than Intel in that case. (I have a Ryzen 1700 that often does better than Haswells per core, but those machines are on Ubuntu.)

Windows 10 on my Ryzen.. there are plenty of benchmarks out there to prove the windows scheduler holds them back a bit.

I've got a Haswell on win10 as well.. does more per core but it's also at 4.2ghz

Be a while before I report on my next round of Ryzen tests.

Anyone else tested the impact of DDR4 speed or SMT?

[/quote]I am doing a test now to compare virtual cores vs. real cores on two Intel Haswell machines running Win7 64-bit.
On an i7-4771, I am running WAH2 on four real cores.

Check Task Manager, Performance. I think you will find that Windows automatically treats ALL cores as Real and that the computing is equally distributed to 8 cores. That is what I read in some technical discussion board.

That is what it does for my i7-4790. During working hours I tell BOINC to use 50% of the CPUs, but I see all 8 running at just over 50%. At night I specify 100% of CPUs and see all 8 running at 100%.

Check Task Manager, Performance. I think you will find that Windows automatically treats ALL cores as Real and that the computing is equally distributed to 8 cores. That is what I read in some technical discussion board.

That is a bit different. To run real cores, I disable hyper-threading in the BIOS. Then, there are only four real (really real) cores available to the OS.

In this manner, I can check the benefits of hyper-threading, and don't see them in the case of CPDN (at least WAH2). Most other projects do see an overall increase in performance with hyper-threading.

I would recommend still enabling hyper-threading, and just limit CPDN to four cores in the BOINC Manager, by setting "Use 50% of the cores" for example (or an app_config.xml file). In that manner, you still can get the benefits of HT for the other projects you run (or other desktop applications) that can use it.

I didn't bother running CPU jobs on Bulldozer/Piledriver. The poor CPU floating point performance meant that they were better employed running GPU jobs like Einstein. I ran CPU jobs on Intel.

Switching to a Ryzen 2200G has been kind of interesting. Not only is the FP performance in the same ballpark as Intel, but it also meant in this case a healthy jump in memory throughput (to 3000 MT/s). And my hypothesis is that given acceptable FP performance, CPDN is mainly memory bound, not CPU bound. Given that it's a giant matrix that has no hope of fitting in any data cache, that wouldn't be too surprising.

For various reasons, I haven't tested this (CPDN outages, other projects running). But if it's true, it leads to two further observations: first, SMT/hyperthreading will mainly help other (non CPDN) processes, that are not memory bound. They can execute during the numerous cache misses caused by traversing a giant data structure. Second, running your CPU at full speed all the time for CPDN is a waste of power. My results seem (anecdotally) to be similar whether the 2200G is running at 3.6GHz or throttled to 2.4GHz, at a third of the power.

And yes, the 2200G has no SMT. With a lot of experience running BOINC and VMs, and BOINC in VMs, I'm not a big fan.

---

I have to agree with you Evans. I've now done a little experimenting with cpu clocks and memory clocks on my Ryzen 1700x

Work done scaled linearly at 1:1 as memory clocks went up.
CPU clocks saw some gains but nothing like the memory clocks.

I've got mine running at 2933mhz w/mediocre timings and I'm currently the #5 host.

3 of the top 10 hosts are Ryzen 7 1st gens... CPDN certainly works great on Ryzen.

I'd love to see what a 2nd gen Ryzen w/really fast RAM could produce.