AMD Instinct MI350X product page spec table

VERBATIM 2026-10-04, curl -L -A 'Mozilla/5.0 (Windows NT 10.0; Win64; x64)' from amd.com. Every row of AMD's printed spec table, verbatim: 'GPU Architecture CDNA4' / 'Lithography TSMC 3nm | 6nm FinFET' / 'Stream Processors 16,384' / 'Matrix Cores 1,024' / 'Compute Units 256' / 'Peak Engine Clock 2200 MHz' 'Peak Microscaling Four-bit Precision Matrix (MXFP4) Performance 9.2 PFLOPs' 'Peak Microscaling Six-bit Precision Matrix (MXFP6) Performance 9.2 PFLOPs' 'Peak Microscaling Eight-bit Precision Matrix (MXFP8) Performance 4.6 PFLOPs' 'Peak Open Compute Project Eight-bit Precision Matrix (OCP-FP8) Performance (E5M2, E4M3) 4.6 PFLOPs' 'Peak ... OCP-FP8 Performance with Structured Sparsity (E5M2, E4M3) 9.2 PFLOPs' 'Peak Half Precision Matrix (FP16) Performance 2.3 PFLOPs' 'Peak Half Precision Matrix (FP16) Performance with Structured Sparsity 4.6 PFLOPs' 'Peak Half Precision (FP16) Performance 144.2 TFLOPs' 'Peak Single Precision Matrix (FP32) Performance 144.2 TFLOPs' / 'Peak Single Precision (FP32) Performance 144.2 TFLOPs' 'Peak Double Precision Matrix (FP64) Performance 72.1 TFLOPs' / 'Peak Double Precision (FP64) Performance 72.1 TFLOPs' 'Peak INT8 Matrix Performance 4.6 POPs' / '... with Structured Sparsity 9.2 POPs' 'Peak bfloat16 (BF16) Matrix Performance 2.3 PFLOPs' / '... with Structured Sparsity 4.6 PFLOPs' 'Transistor Count 185 Billion' / 'External Power Connectors 54V UBB' / 'Typical Board Power (TBP) 1000W (TBP)' 'Last Level Cache (LLC) 256 MB' / 'Dedicated Memory Size 288 GB' / 'Dedicated Memory Type HBM3E' / 'Infinity Cache Yes' / 'Memory Interface 8192-bit' / 'Memory Clock 8 GHz' / 'Peak Memory Bandwidth 8 TB/s' / 'Memory ECC Support Yes (Full-Chip)' 'GPU Form Factor OAM Module' / 'Bus Type PCIe 5.0 x16' / 'Infinity Fabric Links 7' / 'Peak Infinity Fabric Link Bandwidth 153 GB/s' / 'Scale-out AMD Infinity Fabric Links (Per GPU) 1' Confirms amd-instinct-mi350x field by field, including the subtle point that MXFP4 and MXFP6 print 9.2 PFLOPs with NO 'with Structured Sparsity' row, so they are DENSE, while OCP-FP8's 9.2 IS sparse (its dense counterpart is 4.6). MI350X is exactly 8/9 of MI355X on every matrix row (2.3/2.5, 4.6/5, 9.2/10.1) but only 144.2/157.3 on the vector rows - the 8-XCD/2200MHz vs 8-XCD/2400MHz relationship. FP64 72.1 vs 78.6 is NOT the 8/9 ratio (72.1/78.6 = 0.917, closer to MI350P's half), consistent with the record's claim that MI350X is the 4-XCD part. This page IS the source of 'Memory Interface 8192-bit' and 'Memory Clock 8 GHz', supporting the record's 8192-bit bus and its 7.81 Gbps-per-pin derivation (8000 GB/s x 8 / 8192). The MI355X page omits both lines.