Post AVvhL5DextD7ygA4SO by niconiconi@mk.absturztau.be
(DIR) More posts by niconiconi@mk.absturztau.be
(DIR) Post #AVvhL5DextD7ygA4SO by niconiconi@mk.absturztau.be
2023-05-23T00:22:03.342Z
1 likes, 0 repeats
FDTD electromagnetism simulation via openEMS, AMD Ryzen 9 7950X vs Apple M1 Ultra. The results are baffling, M1 Ultra outperforms Ryzen by 150% to 200% in a few simulations, as expected due to its 1024-bit memory bus. But in other simulations it's 50% as slow. There are definitely some non-DRAM bottlenecks in the code that prevent scaling.
(DIR) Post #AVwioZkso4JgC2Hvvs by counteractor@pawoo.net
2023-05-23T11:32:38Z
0 likes, 0 repeats
@niconiconi Why the thread count differences?
(DIR) Post #AVwiotOZmiMB6e0OwK by niconiconi@mk.absturztau.be
2023-05-23T12:13:14.501Z
0 likes, 0 repeats
@counteractor@pawoo.net It's memory-bound, increasing thread count beyond a point will make it slower and slower. On different machines the thread count needed for reaching the peak speed is different.