Discussion Hello again!

[AshlayW]

I disappeared for a while due to reasons, including, mainly, financial issues. Anyway, I think I may return and help out with some of that crunching that needs doing. My system(s) have changed a lot, but with the new hardware being released I'm really excited to return to at least some form of contribution to WCG and FAH.

Right now I have an R7 5800X3D and 2x RTX 3060 is my main system used as a blender OptiX machine, and a R9 3900XT and another 3060, and a 3050 together as a secondary (also OptiX) machine. My main excitement was to see how good the 5800X3D is in WCG since the huge L3 V-cache should allow many projects to fit entirely in L3 cache (as it's unified), boosting IPC and perf/W, too.

I was thinking about getting back into FAH with the 4 RTX GPUs and maybe the V56 in my laptop. This always comes around this time of year, when the machines also heat my room (I live in an attic-room at the top of a 3 storey house), so the power bill has dual purpose :^)

There's so many things I need to catch up on. I was just skimming some of Mark's posts on the 7950X and nerding out over Zen4, so it's really exciting. I long for a 7950X + RTX 4090 system but I'd rather not miss my rent for 6 months to save for it lol.

Anyway, I hope everyone is OK, and doing well. How is everyone doing?

On a side note I'm a little bit worried about AT since Dr. Cutress left, but there's no other community I'd rather be part of tbh.

-Ash

edit: fixing typos

 

[crashtech]

Glad you are back! I had wondered what happened, iirc you were a force to be reckoned with in WCG. The 5800X3D is definitely interesting! I'd ruled it out for my purposes because of the need to increase density. It's not possible for me to add more computers, so my focus has to be on increasing core count per machine. If AMD releases a high core count X3D CPU, I find it likely that I would go into debt to obtain one!

 

[StefanR5R]

7773X, 7573X, 7473X, and 7373X are available in retail and in stock, here in Germany at least. :-)

There was one WCG project which was presumed to desire a lot of cache: Microbiome Immunity Project. This project has finished a while ago though.

WCG's African Rainfall Project, and the few Climateprediction.net projects are other ones of which I have read users suspect high cache demands. I never saw sound evidence of that though. These would probably be very hard to benchmark.

Cache demands of PrimeGrid's applications are known in great detail though. Its biggest LLR based project, Seventeen or Bust, wanted 25 MBytes cache per task when I last checked (March 2021). So that is still served very well with plain Zen 3 and Zen 4 CCXs. Genefer-21 and Genefer-22 presumably want more than 32 MB (perhaps somewhat more than 40 MB). But when I squeezed two of such tasks onto a Broadwell-EP with 55 MB L3$, the drop in throughput versus just on task was just about 10%. So Zen 3/4 without 3D cache might muddle through with one task on a CCX quite reasonably too. And besides, Genefer-21/-22 run more efficiently on recent GPUs anyway, compared to CPUs.

Though cache-heavy science applications aside, I wonder if the big 3D cache helps in scenarios where you run moderately RAM- and cache-demanding applications on almost all threads, and at the same time a GPU task (or several, in case of a multi-GPU computer). The RAM accesses of the GPU feeder task might bear down heavily on the CPU's caches, depending on the access patterns and how the processor's caching algorithms deal with it, to the detriment of the other tasks.

 

[Markfw]

7773X, 7573X, 7473X, and 7373X are available in retail and in stock, here in Germany at least. :-)

There was one WCG project which was presumed to desire a lot of cache: Microbiome Immunity Project. This project has finished a while ago though.

WCG's African Rainfall Project, and the few Climateprediction.net projects are other ones of which I have read users suspect high cache demands. I never saw sound evidence of that though. These would probably be very hard to benchmark.

Cache demands of PrimeGrid's applications are known in great detail though. Its biggest LLR based project, Seventeen or Bust, wanted 25 MBytes cache per task when I last checked (March 2021). So that is still served very well with plain Zen 3 and Zen 4 CCXs. Genefer-21 and Genefer-22 presumably want more than 32 MB (perhaps somewhat more than 40 MB). But when I squeezed two of such tasks onto a Broadwell-EP with 55 MB L3$, the drop in throughput versus just on task was just about 10%. So Zen 3/4 without 3D cache might muddle through with one task on a CCX quite reasonably too. And besides, Genefer-21/-22 run more efficiently on recent GPUs anyway, compared to CPUs.

Though cache-heavy science applications aside, I wonder if the big 3D cache helps in scenarios where you run moderately RAM- and cache-demanding applications on almost all threads, and at the same time a GPU task (or several, in case of a multi-GPU computer). The RAM accesses of the GPU feeder task might bear down heavily on the CPU's caches, depending on the access patterns and how the processor's caching algorithms deal with it, to the detriment of the other tasks.

I have the motherboard for those, and 2 7663 64 cores Milans. I want the 7773x's but they are $5,500 each ! That would be $11,000, and thats way out of my budget.

 

[Orange Kid]

Welcome back.
Run'em hard and keep warm.

 

[Assimilator1]

Welcome back

On a side note I'm a little bit worried about AT since Dr. Cutress left, but there's no other community I'd rather be part of tbh.

Any info on that?