stats.md

Running tflow.py with 25 epochs and 4096 batch size

AD102 NVidia 4090 Ada Dual 384bit 1008GB/s 24G+24G PCIe 8x only - 32768 cores

98ms 70c 270x2w 91-70%

AD102 NVidia 4090 Ada 24G PCIe 8x - 16384 cores

137ms 66c 360w 97-93%

GA102 NVidia A4500 Ampere Dual 320bit 640GB/s PCIe x8 40G NVLink - 14336 cores

• 20% faster than a single A6000
• 38% faster than a single A4500
• 381ms 90c 400w 97-57%
• 372ms on 13900KS
• 608ms for batch 8192

GA102 NVidia A6000 Ampere 384bit 768GB/s 48G PCIe x16 - 10752 cores

• The motherboard will affect the performance of an A6000 with up to 30% performance degradation - making an A6000 behave like an A4500
• 390ms 100c tt 308w 100-90% on 790 maximus with full power availablity 1600w psu
• 510ms 92c 280w 92-70% - on 790 hero with lower power availability 1500w psu
• 388ms-397ms on 14900k
• 757ms for batch 8192

GA102 NVidia A4500 Ampere 320bit 640GB/s 20G PCIe x8 - 7168 cores

• 511ms 90c 205w 99-92%
• 524ms 90c 209w 98-90%
• 512ms 13900KS
• 989ms for batch 8192

GA102 NVidia A4500 Ampere 320bit 640GB/s 20G PCIe 16x - 7168 cores

505ms 90c 205w 99-96%

GA104 NVidia A4000 Ampere 256bit 448GB/s 16G PCIe 16x - 6144 cores

778ms 92c 140w 92-70%

AD104 NVidia 3500 Ada Mobile 192bit 432GB/s 12G - 5120 cores

552ms 110c 96w 100%

TU104 NVidia RTX-5000 Mobile 256bit 448GB/s 16G PCIe 3.0 16x - 3072 cores

990ms 118w 99-92%

GP107 NVidia P1000 Pascal 4G - cores

M1Max Macbook Pro

32 core GPU is 8x faster than CPU

strategy = tf.distribute.OneDeviceStrategy(device="/gpu")

47000ms 80% GPU

strategy = tf.distribute.OneDeviceStrategy(device="/cpu")

24000ms 450% CPU