3 ms·

This is the reason why I went with the 7950X and not the 3D variant. Too much additional work involved to get the other cores to play well. Also I read that on

This is the reason why I went with the 7950X and not the 3D variant. Too much additional work involved to get the other cores to play well. Also I read that only one set of 6 cores as full access to cache while the other doesn't or has partial access.

_cenw 3y ago
The cache is on a single CCX, yes. Which is why the 7800X3D is so good: There is no second CCX. Another big advantage is that to get them to work, they need to use higher bin chips. CCX overheats easily because the stacked cache acts like a heat shield, so you get the most power efficient ones.
- kvemkon 3y ago
  The disadvantage of single CCD is only half of available write memory bandwidth (using at least 1 computing thread on each CCD) [1]. [1] 16 B/cycle write vs 32 B/cycle read per CCD https://www.servethehome.com/amd-ryzen-7000-series-platform-overview/amd-ryzen-7000-soc-architecture-soc-topology-with-2x-ccd-1x-iod/ https://www.servethehome.com/amd-ryzen-7000-series-platform-...
  - _cenw 3y ago
    I'd wager the number of workloads that'd actually take advantage of operating on entirely different data on a per-thread basis is much, much lower than ones that benefit from massive L3 cache. And note that the infinity fabric and memory controller don't run on the same clock. The fabric tops out at ~2200 (though testing shows sweet spot at 2033 usually) and the memory controller usually tops out at ~3100 (DDR5-6200).
    - kvemkon 3y ago
      Those might be not necessary 2 threads of one program, but 2 independent programs.
slowmovintarget 3y ago
The interesting thing was that in testing it turned out that having 3D cache for all cores didn't increase performance. The sweet spot for heat/performance was for half. Granted, I don't think I'm running anything so CPU intensive that I'll be able to tell. For gaming, I was more interested in the DDR5 memory than in worrying about the L3 cache. For everything else I do (including running LLMs and compilers) it's beyond adequate. For a mere $20K I could have had an overall slower gaming experience but be able to locally run things like the Goliath 120B model, but that seemed like a poor trade-off.

_cenw 3y ago

The cache is on a single CCX, yes. Which is why the 7800X3D is so good: There is no second CCX. Another big advantage is that to get them to work, they need to use higher bin chips. CCX overheats easily because the stacked cache acts like a heat shield, so you get the most power efficient ones.

kvemkon 3y ago
The disadvantage of single CCD is only half of available write memory bandwidth (using at least 1 computing thread on each CCD) [1]. [1] 16 B/cycle write vs 32 B/cycle read per CCD https://www.servethehome.com/amd-ryzen-7000-series-platform-overview/amd-ryzen-7000-soc-architecture-soc-topology-with-2x-ccd-1x-iod/ https://www.servethehome.com/amd-ryzen-7000-series-platform-...
- _cenw 3y ago
  I'd wager the number of workloads that'd actually take advantage of operating on entirely different data on a per-thread basis is much, much lower than ones that benefit from massive L3 cache. And note that the infinity fabric and memory controller don't run on the same clock. The fabric tops out at ~2200 (though testing shows sweet spot at 2033 usually) and the memory controller usually tops out at ~3100 (DDR5-6200).
  - kvemkon 3y ago
    Those might be not necessary 2 threads of one program, but 2 independent programs.

slowmovintarget 3y ago

The interesting thing was that in testing it turned out that having 3D cache for all cores didn't increase performance. The sweet spot for heat/performance was for half. Granted, I don't think I'm running anything so CPU intensive that I'll be able to tell. For gaming, I was more interested in the DDR5 memory than in worrying about the L3 cache. For everything else I do (including running LLMs and compilers) it's beyond adequate. For a mere $20K I could have had an overall slower gaming experience but be able to locally run things like the Goliath 120B model, but that seemed like a poor trade-off.