VRAM
24 GB
TDP
355 W
MSRP
$999
Perf/W
0.27 img/min/W
Cost/1K tok
$0.11/M
Tested
2025-06-09
AMD Radeon RX 7900 XTX 24GB has a source-attributed result of 11.0 tok/s on Llama 3 70B Q4 (batch 1, 4096-token context, Q4_K_M; runtime not documented). This is a reference report, not an independently verified lab result. A 70B Q4_K_M artifact needs roughly 40GB for weights alone; smaller memory configurations require explicit offload or model splitting and do not establish full-GPU residency.
AMD Radeon RX 7900 XTX 24GB — RDNA 3 consumer flagship. 24GB GDDR6 at ~960 GB/s, 355W TDP. Competes with RTX 4080/4090 on VRAM at dramatically lower price. ROCm support improving but not at CUDA parity.
24GB VRAM at ~$900-1000 makes this the best $/VRAM consumer card if you can tolerate the ROCm ecosystem. Llama 3 8B Q4 runs at ~90+ tok/s via llama.cpp ROCm. Vulkan backends (llama.cpp) are slower but work universally. Best for Linux-native builders who want NVIDIA-tier VRAM at AMD pricing and are comfortable with the software tradeoffs. Not recommended for Windows AI workflows or production CUDA-dependent stacks.
Buy for Linux AI builds where $/VRAM is the priority and you're comfortable with the ROCm ecosystem. Skip for Windows, for production, or if you value plug-and-play over tinkering.
24GB VRAM for Linux-native builders who want NVIDIA-tier capacity at AMD pricing and are comfortable with software tradeoffs
ROCm driver compatibility — pin your kernel and ROCm versions once stable. HIP runtime errors are more common than CUDA errors in edge cases. Vulkan fallback is functional but 30-50% slower.
Ubuntu 22.04/24.04 (good with ROCm), Windows (Vulkan only, slower), macOS (unsupported)
355W TGP. 850W PSU minimum. ROCm power management less mature than NVIDIA. Higher idle power draw than equivalent NVIDIA cards.
AMD Radeon RX 7900 XTX 24GB with 24GB VRAM at 355W TDP, scored across 11 workloads with 17 benchmark records.
Reference workload
SDXL image gen
13 img/min
Quantization
FP16
· batch 1
| Workload | Score | Quant | Context | σ | Status | Tested |
|---|---|---|---|---|---|---|
| Llama 3 8B Q4 llm | 95.0tok/s | Q4_K_M | 4K | , | Curated Aggregate | 2025-06-09 |
| Llama 3 70B Q4 llm | 11.0tok/s | Q4_K_M | 4K | , | Curated Aggregate | 2025-08-13 |
| SDXL image gen image | 13.0img/min | FP16 | , | , | Curated Aggregate | 2026-03-11 |
| Qwen 2.5 14B llm | 48.0tok/s | Q4_K_M | 8K | , | Curated Aggregate | 2024-11-27 |
| Whisper transcription audio | 22.0x RT | FP16 | , | , | Curated Aggregate | 2024-10-18 |
| Mistral 7B llm | 95.0tok/s | Q4_K_M | 4K | , | Curated Aggregate | 2025-02-12 |
| Phi-3 Mini llm | 195.0tok/s | Q4_K_M | 4K | , | Curated Aggregate | 2025-03-08 |
| Gemma 2 9B llm | 58.0tok/s | Q4_K_M | 4K | , | Curated Aggregate | 2025-03-22 |
| DeepSeek-R1 7B llm | 78.0tok/s | Q4_K_M | 8K | , | Curated Aggregate | 2025-04-15 |
| Embedding throughput embedding | 5800.0emb/s | FP16 | 1K | , | Curated Aggregate | 2024-11-04 |
| Asure 12B llm | 52.0tok/s | Q4_K_M | 8K | , | Curated Aggregate | 2025-05-01 |
There is insufficient comparable evidence to score AMD Radeon RX 7900 XTX 24GB. A score needs results for multiple devices with matching workload, runtime version, quantization, context and batch size.
Planning estimates for 24GB: weights plus at least 2GB or 10% overhead. Actual KV cache depends on model, context and cache format; confirm the artifact before buying.
Curated from public sources (llama.cpp logs / vendor specs / vLLM community) — Automatic1111 ROCm.
View sourceHow we benchmark →Trust score
6/10
MyAI rating
Not scored
Runs
Not documented
Freshness
Stale
Source-linked row with explicit verification status.
Record date: 2026-03-11; 182 days old.
Owner-reported capture via ollama-local-api on 2026-06-01. Hardware identity and outputs are supplied by the owner; this is not independent replication.
Windows 10 build 26200 + ROCm 7.10 + Ollama 0.24.0 (HIPBLAS). 4 runs per model after a discarded warmup; decode_tps reported as per-run eval_count/eval_duration. Variance <= 0.4% on reproduced models. Captured by site owner on declared rig. Prompt SHA-256: 5e752b688e0761f321a5721d2fa6b57678bf1de1a1180bb4fc9b7e709bafb6a5.
llama-3.2-3b-instruct
3B / Q4_K_M
median 201.87 / min 201.31 / max 202.80 / n=4
qwen-2.5-7b-instruct
7B / Q4_K_M
median 115.44 / min 115.27 / max 116.04 / n=4
mistral-7b-instruct
7B / Q4_K_M
median 118.72 / min 118.38 / max 119.28 / n=4
gemma-4-31b-it
31B / Q4_K_M
n=4, throughput 0 tok/s
Decode failed: model prefilled then produced 0 tokens across all runs on AMD ROCm 7.10 + Ollama 0.24.0. A real, reproducible failure mode, not missing data.
Best place to start
Retailer we'd check first
Amazon
For mainstream AI hardware, Amazon usually updates street pricing, seller availability, and shipping speed faster than most comparison sites.
View current Amazon listingThere is insufficient comparable evidence for a purchase recommendation. Check model fit, software support and a current seller quote.
Affiliate note: this button opens the mapped Amazon product listing for this device. We may earn from qualifying purchases.