Intel Gaudi 3 AI Accelerator Technical Paper Table 1 - Gaudi 2 vs Gaudi 3 feature comparison (VERBATIM, July 2025 V1 Rev 3)

VERBATIM ROW, as printed: 'BF16 MME TFLOPS | 432 | 1678', 'FP8 MME TFLOPS | 865 | 1678', 'BF16 Vector TFLOPS | 112 | 28,7' (OCR of '28.7'), 'MME Units | 2 | 8', 'TPC Units | 24 | 64', 'HBM Capacity | 96 GB | 128 GB', 'HBM Bandwidth | 2.46 TB/s | 3.7 TB/s', 'On-die SRAM Capacity | 48 MB | 96 MB', 'On-die SRAM Bandwidth (read/write) | 6.4/6.4 TB/s | 12.8/6.4 TB/s', 'Networking (bidirectional) | 600 GB/s | 1200 GB/s', 'Host Interface | PCIe Gen4 X16 | PCIe Gen5 X16', 'Host Interface Peak BW | 64 GB/s (32 GB/s per direction) | 128 GB/s (64 GB/s per direction)', 'Media Decoders | 8 | 14'. (Gaudi 2 column first, Gaudi 3 column second.) PCIe CONVENTION, VERIFIED EXPLICITLY: Intel prints BOTH numbers in one cell. 'Peak BW' is the BIDIRECTIONAL AGGREGATE and the parenthetical is PER DIRECTION. 32 GB/s per direction on Gen4 x16 = 16 GT/s x 16 / 8 = 32 GB/s; 64 GB/s per direction on Gen5 x16 = 32 GT/s x 16 / 8 = 64 GB/s. Both check out arithmetically, so Intel's aggregate figures (64 and 128 GB/s) are correct as aggregates and must NOT be halved. Intel never states dense or sparse for any Gaudi TFLOPS figure. Confirms: vram_gb 96/128, memory_bandwidth_gbps 2460/3700, all three Gaudi 2 flops rows (432 MME BF16, 865 MME FP8, 112 BF16 VECTOR as a distinct unit), both Gaudi 3 flops rows (1678 BF16 MME, 1678 FP8 MME), 600/1200 GB/s bidirectional networking. Confirms MME/TPC unit counts and 8 media decoders (Gaudi 2) / 14 (Gaudi 3).