AMD Radeon AI PRO R9700 product specifications page (AI 9000 Series)

Vendor primary spec page for the RDNA4 workstation part. No publication date printed, so published is null. Verbatim fields used: GPU Architecture 'AMD RDNA 4', Ray Accelerators 64, AI Accelerators 128, Stream Processors 4096, Compute Units 64, Boost Frequency Up to 2920 MHz, Game Frequency 2350 MHz, Peak Pixel Fill-Rate Up to 373.76 GP/s, Peak Single Precision (FP32 Vector) Performance 47.8 TFLOPs, Peak Half Precision (FP16 Vector) Performance 47.8 TFLOPs, Peak Half Precision (FP16 Matrix) Performance 191 TFLOPs, Peak Half Precision (FP16 Matrix) Performance with Structured Sparsity 383 TFLOPs, Peak 8-bit Precision (FP8 Matrix) Performance (E5M2, E4M3) 383 TFLOPs, Peak 8-bit Precision (FP8 Matrix) Performance with Structured Sparsity (E5M2, E4M3) 766 TFLOPs, Peak 8-bit Precision (INT8 Matrix) Performance 383 TOPs, Peak 8-bit Precision (INT8 Matrix) Performance with Structured Sparsity 766 TOPs, Peak 4-bit Precision (INT4 Matrix) Performance 766 TOPs, Peak 4-bit Precision (INT4 Matrix) Performance with Structured Sparsity 1531 TOPs, ROPs 128, Transistor Count 53.9 Billion, External Power Connectors 12V-2x6, Total Board Power (TBP) 300W, Minimum PSU Recommendation 750W, GPU Memory: Dedicated Memory Size 32 GB, Dedicated Memory Type GDDR6, AMD Infinity Cache Technology 64 MB, Memory Interface 256-bit, Peak Memory Bandwidth 640 GB/s, Memory ECC Support 'Yes (Linux Only)', Board Specifications Bus Type PCIe 5.0 x16, Cooling Active, Board Width Double Slot, DisplayPort 'Configuration varies by partner design. GPU supports up to 4x display out.', Display Configurations 'Up to: 4x 4096 x 2160 (4K DCI) @ 120Hz with DSC, 2x 6144 x 3456 (6K) 12-bit HDR @ 60Hz Uncompressed, 2x 7680 x 4320 (8K) 12-bit HDR @ 60Hz with DSC, 1x 12288 x 6912 (12K) @ 120Hz with DSC', TPP 3062, TPP Density 8.91. Note the two qualifications: memory ECC is Linux-only, and display support is 'up to' 4 outputs because it varies by partner board. This is the only AMD Radeon page in this repo's source set that publishes FP8 matrix throughput. Accessed 2026-10-03.