Radeon RX 7900 XT
AMDconsumer| Overview | |
|---|---|
| Architecture | RDNA 3 (2022) |
| Launch date | 2022-12-13 |
| Transistor count | 58B |
| GPU target (gfx) | gfx1100 |
| Memory | |
| VRAM | 20 GB GDDR6 |
| Memory bandwidth | 800 GB/s |
| Cache | 80 MB |
| Compute — vector | |
| FP32 | 51.6 TFLOPS |
| FP16 | 51.6 TFLOPS |
| Compute — matrix / tensor | |
| FP16 | 103 TFLOPS |
| INT8 | 103 TOPS |
| INT4 | 206 TOPS |
| Cores & clocks | |
| Compute Units | 84 |
| Shader cores | 5,376 |
| Matrix / Tensor cores | 168 |
| Peak clock | 2,400 MHz |
| Board & system | |
| TDP | 315 W |
| Interconnect | PCIe only (no Infinity Fabric multi-GPU link) |
Compatible ROCm versions
Same Navi 31 die (gfx1100) as the RX 7900 XTX, so it shares that card's ROCm support from 6.0.0 onward. 20 GB of VRAM makes it the second-roomiest consumer Radeon in this collection after the XTX's 24 GB — relevant for local LLM work, where VRAM capacity binds before compute does.
fp16TFLOPS (103, dense) is AMD's directly published "Peak Half Precision (FP16 Matrix) Performance" figure, distinct from the 51.6 TFLOPs vector rate captured as fp16TFLOPSVector. No structured-sparsity variant is listed (RDNA 3 lacks the 2:4-sparsity feature RDNA 4 parts have), so there is only one figure per precision to report. fp8TFLOPS omitted — RDNA 3 has no native FP8 matrix support; AMD's page instead publishes INT8/INT4 matrix rates (int8TOPS/int4TOPS) via its 168 "AI Accelerators" (used as matrixCoreCount).
memoryBandwidthGBs uses the raw GDDR6 figure (800 GB/s), not any Infinity Cache-inflated "Effective Memory Bandwidth" number, for consistency with the rest of this collection. peakClockMHz uses the published Boost Frequency ("Up to 2400 MHz"), not the 2000 MHz Game Frequency. AMD publishes no MSRP on this page, so msrpUSD is omitted. Note the URL has no hyphen before "xt" (same quirk as the 7900 XTX page). Source: amd.com/en/products/graphics/desktops/radeon/7000-series/amd-radeon-rx-7900xt.html ("Expand All" spec accordion), pulled 2026-09-20.