HARDWARE BENCHMARK & VRAM ENVELOPE โข CONSUMER GPU20 GB VRAM โข 800 GB/s
What Models Can Run on AMD Radeon RX 7900 XT (20GB GDDR6)?
20 GB VRAM800 GB/s Bandwidth315W TDPMSRP: $699Cloud: $0.29/hr on Vast.ai
AMD Radeon RX 7900 XT (20GB GDDR6) provides 20 GB of usable memory with 800 GB/s memory bandwidth at 315W TDP. Sub-$700 GPU with 20GB VRAM fitting 14B Q8 and 27B Q4 models.
โก PARTIAL OFFLOAD21.93 GB (110%)~6 t/s
Qwen 3.8 27B Instruct on AMD Radeon RX 7900 XT (20GB GDDR6)
โ ๏ธ Partial Offload โ Slow CPU/RAM Spillover
โก PARTIAL OFFLOAD
Required VRAM
21.93 GB
110% of hardware capacity (20 GB)Est. Decode Speed
~6 tok/s
Dense model @ 800 GB/sAvailable Headroom
-1.93 GB
Needs +1.93 GB more VRAMMemory Allocation & Headroom
21.93 GB allocated of 20 GB capacity
110% utilized
Weights: 16.32 GB
KV: (32K): 4.29 GB
Overhead: 1.32 GB
Deficit: -1.93 GB
โก Architecture & Throughput Transparency: Single-user decode speed (~6 tok/s) is memory-bandwidth bound at batch size = 1 on AMD Radeon RX 7900 XT (20GB GDDR6) (800 GB/s). Prompt prefill / TTFT (~87 tok/s) is compute-bound and utilizes GPU Tensor cores. Actual latency varies with FlashAttention-2 and runtime engine (Ollama/vLLM/llama.cpp).
OLLAMA_FLASH_ATTENTION=1 OLLAMA_KV_CACHE_TYPE=q8_0 ollama run qwen3.8:27bโ ๏ธ Out of VRAM? Don't downgrade model quality.
Out of VRAMSpin up an on-demand cloud GPU instance in 60 seconds with pre-installed Ollama and vLLM. (Qwen 3.8 27B Instruct @ 21.93 GB VRAM)
RunPod Cloud GPUsBest 1-Click Ollama & vLLM
$0.34/hr (RTX 4070 Ti) โ $2.39/hr (H100)Vast.ai GPU MarketplaceCheapest Dual RTX 3090 / 4090
$0.18/hr (RTX 3060) โ $0.55/hr (RTX 4090)Lambda GPU CloudEnterprise 8x H100 Clusters
$1.99/hr (GH200 / H100)Quantization Precision Matrix
32K ctx โข KV: Q8_0| Quantization | Weights | KV Cache | Total VRAM | Est. Speed | Fit Status |
|---|---|---|---|---|---|
| FP1616 bpw | 54.4 GB | 4.29 GB | 61.34 GB (307%) | 0 t/s | OOM |
| Q8_08.5 bpw | 28.83 GB | 4.29 GB | 34.88 GB (174%) | 0 t/s | OOM |
| Q6_K6.56 bpw | 22.3 GB | 4.29 GB | 28.12 GB (141%) | 0 t/s | OOM |
| Q5_K_M5.68 bpw | 19.31 GB | 4.29 GB | 25.03 GB (125%) | ~5 t/s | PARTIAL |
| Q4_K_M4.8 bpw | 16.32 GB | 4.29 GB | 21.93 GB (110%) | ~6 t/s | PARTIAL |
| FP44.25 bpw | 14.42 GB | 4.29 GB | 19.96 GB (100%) | ~30 t/s | FITS |
| IQ3_XS3.3 bpw | 11.15 GB | 4.29 GB | 16.58 GB (83%) | ~38 t/s | FITS |
Export Full Architecture & Infrastructure Spec
Get an executive-ready PDF report with exact memory math, hardware benchmarks, and cost comparisons.
All Open-Weights Models Tested on AMD Radeon RX 7900 XT (20GB GDDR6) (Q4_K_M @ 32K Context)
Standard Q4_K_M precision with 32K context window and Q8_0 KV cache
| Model | Architecture | Total VRAM Needed | Headroom | Est. Decode Speed | Compatibility | Detailed Config |
|---|---|---|---|---|---|---|
| Qwen 3.8 27B Instruct Alibaba Cloud | 27.2B Dense | 21.93 GB(110%) | -1.93 GB | ~6 tok/s | โก PARTIAL | Configure CLI โ |
| Qwen 3.6 35B-A3B MoE Alibaba Cloud | 35B MoE | 24.1 GB(121%) | -4.1 GB | ~46 tok/s | โก PARTIAL | Configure CLI โ |
| DeepSeek-V4.1-Flash (552B Engram MoE) DeepSeek | 552B MoE | 345.59 GB(1728%) | -325.59 GB | 0 (OOM) | โ OOM | Configure CLI โ |
| Meta Muse Spark Open (816B-A42B MoE) Meta AI | 816B MoE | 512.86 GB(2564%) | -492.86 GB | 0 (OOM) | โ OOM | Configure CLI โ |
| Google Gemma 4 26B MoE Google | 26.4B MoE | 20.23 GB(101%) | -0.23 GB | ~31 tok/s | โก PARTIAL | Configure CLI โ |
| Google Gemma 4 E4B (Edge 4B) Google | 4.1B Dense | 4.24 GB(21%) | +15.76 GB | ~172 tok/s | โ FITS | Configure CLI โ |
| DeepSeek-R1 Distill Qwen 32B DeepSeek | 32.8B Dense | 25.41 GB(127%) | -5.41 GB | ~5 tok/s | โก PARTIAL | Configure CLI โ |
| DeepSeek-R1 Distill Qwen 14B DeepSeek | 14.8B Dense | 13.16 GB(66%) | +6.84 GB | ~48 tok/s | โ FITS | Configure CLI โ |
| Meta Llama 4 Scout (109B-A17B MoE) Meta AI | 109B MoE | 71.66 GB(358%) | -51.66 GB | 0 (OOM) | โ OOM | Configure CLI โ |
| Meta Llama 3.3 70B Instruct Meta AI | 70.6B Dense | 49.96 GB(250%) | -29.96 GB | 0 (OOM) | โ OOM | Configure CLI โ |
| Meta Llama 3.1 8B Instruct Meta AI | 8B Dense | 7.87 GB(39%) | +12.13 GB | ~87 tok/s | โ FITS | Configure CLI โ |
| Microsoft Phi-4 Mini 3.8B Microsoft | 3.8B Dense | 3.92 GB(20%) | +16.08 GB | ~187 tok/s | โ FITS | Configure CLI โ |
| Qwen 2.5 Coder 32B Instruct Alibaba Cloud | 32.5B Dense | 25.22 GB(126%) | -5.22 GB | ~5 tok/s | โก PARTIAL | Configure CLI โ |
| DeepSeek Coder V2 236B MoE DeepSeek | 236B MoE | 151.34 GB(757%) | -131.34 GB | 0 (OOM) | โ OOM | Configure CLI โ |
| Meta Llama 3.1 405B Instruct Meta AI | 405B Dense | 260.72 GB(1304%) | -240.72 GB | 0 (OOM) | โ OOM | Configure CLI โ |
| Mistral Large 2 (123B Instruct) Mistral AI | 123B Dense | 83.04 GB(415%) | -63.04 GB | 0 (OOM) | โ OOM | Configure CLI โ |
| Google Gemma 2 27B Instruct Google | 27.2B Dense | 23.81 GB(119%) | -3.81 GB | ~6 tok/s | โก PARTIAL | Configure CLI โ |
| Google Gemma 2 9B Instruct Google | 9.2B Dense | 12.1 GB(61%) | +7.9 GB | ~71 tok/s | โ FITS | Configure CLI โ |