Radeon AI PRO R9700
AMDworkstation| Overview | |
|---|---|
| Architecture | RDNA 4 (2025) |
| Transistor count | 53.9B |
| GPU target (gfx) | gfx1201 |
| Memory | |
| VRAM | 32 GB GDDR6 |
| Memory bandwidth | 640 GB/s |
| Cache | 64 MB |
| ECC | Yes |
| Compute — vector | |
| FP32 | 47.8 TFLOPS |
| FP16 | 47.8 TFLOPS |
| Compute — matrix / tensor | |
| FP16 (dense / sparse) | 191 / 383 TFLOPS |
| FP8 (dense / sparse) | 383 / 766 TFLOPS |
| INT8 (dense / sparse) | 383 / 766 TOPS |
| INT4 | 766 TOPS |
| Cores & clocks | |
| Compute Units | 64 |
| Shader cores | 4,096 |
| Matrix / Tensor cores | 128 |
| Peak clock | 2,920 MHz |
| Board & system | |
| TDP | 300 W |
| Form factor | Double-slot desktop add-in card |
| Cooling | Active |
| PCIe | PCIe 5.0 x16 |
| Interconnect | PCIe only (no dedicated multi-GPU link) |
Compatible ROCm versions
The largest-VRAM AMD card outside the Instinct line, and the reason it matters here: 32 GB on gfx1201, the same Navi 48 target as the RX 9070 XT, so it inherits that target's ROCm support from 6.4.0 (2025-04) onward. Anyone hitting the 16 GB ceiling on RDNA 4 consumer cards gets double the capacity without changing gfx target, ROCm version, or toolchain — see /rocm-consumer-gpus/ for the consumer-side picture.
Added 2026-09-21 after a reader pointed out the site covered NVIDIA's RTX PRO 6000 Blackwell but had no AMD workstation entry at all — a vendor-coverage asymmetry, not an oversight about one card.
Compute specs are identical to the RX 9070 XT's full Navi 48 configuration (64 CUs, 4096 stream processors, 128 AI accelerators, 53.9 B transistors) at slightly lower clocks (2920 MHz boost vs 2970) and 300 W instead of 304 W. What actually differs is memory capacity (32 GB vs 16 GB on the same 256-bit bus at the same 640 GB/s) and ECC support, which AMD publishes as "Yes (Linux Only)" — recorded as `eccSupport: true` with that OS restriction noted here rather than in a field, since the schema has no place for it. Bandwidth being identical to the 9070 XT's means this card fits much larger models without feeding them any faster.
fp16TFLOPS (191, dense) and fp8TFLOPS (383, dense) are directly published, each with an explicit "with Structured Sparsity" variant (383 / 766) in the ...Sparse fields — AMD publishes dense as the headline and sparsity separately, the opposite of NVIDIA's convention, so no halving is applied. int8TOPS/int8TOPSSparse mirror the FP8 figures (383/766); int4TOPS is listed but its sparsity variant (1,531 TOPs) has no schema field. matrixCoreCount uses the page's "AI Accelerators" count (128).
AMD also sells a spec-identical R9700S at amd-radeon-ai-pro-r9700s.html — every compute, memory and power figure matches this card exactly; the only published difference is Cooling (Passive vs Active), i.e. an OEM/SI variant for chassis with their own airflow. Deliberately not given its own entry: it shares this card's gfx target, capacity and throughput, so it would add a row without adding a compatibility fact. (Contrast the RX 9060 XT 8GB/16GB pair, which do get separate entries, because VRAM capacity changes which models fit.)
peakClockMHz uses the published Boost Frequency ("Up to 2920 MHz"), not the 2350 MHz Game Frequency. tdpWatts is the published "Total Board Power (TBP)". No Launch Date field is published on this page, so launchDate is omitted; releaseYear (2025) is unaffected — AMD announced the card at COMPUTEX 2025 (2025-05-20) without naming an availability date. AMD publishes no MSRP on the spec page, so msrpUSD is omitted. Source: amd.com/en/products/graphics/workstations/radeon-ai-pro/ai-9000-series/amd-radeon-ai-pro-r9700.html ("Expand All" spec accordion), pulled 2026-09-21.