Graphics cards Apple Calculate VRAM for your model, context and concurrency →
A growing selection of desktop and workstation cards. Model capacity is a memory estimate; speed records retain their original test conditions.
Using a Mac? Apple
Choosing a GPU for a 70B LLM? Estimate memory and compare GPU options →
Compare GPU models
Chip maker / GPU Model capacity Estimated · 4K context Published output speed Different tests; compare conditions Purchase USD · single card Rent USD · single GPU-hour
14B Qwen3 14B · Q4_K_M Memory estimate Models & memory Qwen3 8B : within budget · 7.2 GiB estimated
Qwen3 14B : within budget · 11.0 GiB estimated
Qwen3 32B : over budget · 21.4 GiB estimated
Llama 3.3 70B : over budget · 42.9 GiB estimated
No sourced test yet
$549 Launch starting price · reference NVIDIA announcement · US reference Price evidence Historical official starting price announced 6 Jan 2025. Not a current retailer offer; board SKU, tax, shipping and availability vary.
Current purchase price and stock not verified · observed 7 Nov 2026
Inspect listing ↗ No matching quote
14B Qwen3 14B · Q4_K_M Memory estimate Models & memory Qwen3 8B : within budget · 7.2 GiB estimated
Qwen3 14B : within budget · 11.0 GiB estimated
Qwen3 32B : over budget · 21.4 GiB estimated
Llama 3.3 70B : over budget · 42.9 GiB estimated
No sourced test yet
$429 Launch starting price · reference NVIDIA announcement · US reference Price evidence Historical official starting price announced 15 Apr 2025. Not a current retailer offer; board SKU, tax, shipping and availability vary.
Current purchase price and stock not verified · observed 7 Nov 2026
Inspect listing ↗ No matching quote
32B Qwen3 32B · Q4_K_M Memory estimate Models & memory Qwen3 8B : within budget · 7.2 GiB estimated
Qwen3 14B : within budget · 11.0 GiB estimated
Qwen3 32B : within budget · 21.4 GiB estimated
Llama 3.3 70B : over budget · 42.9 GiB estimated
No sourced test yet
No quote collected
No matching quote
32B Qwen3 32B · Q4_K_M Memory estimate Models & memory Qwen3 8B : within budget · 7.2 GiB estimated
Qwen3 14B : within budget · 11.0 GiB estimated
Qwen3 32B : within budget · 21.4 GiB estimated
Llama 3.3 70B : over budget · 42.9 GiB estimated
No sourced test yet
No quote collected
32B Qwen3 32B · Q4_K_M Memory estimate Models & memory Qwen3 8B : within budget · 7.2 GiB estimated
Qwen3 14B : within budget · 11.0 GiB estimated
Qwen3 32B : within budget · 21.4 GiB estimated
Llama 3.3 70B : over budget · 42.9 GiB estimated
~150 tok/s Llama 3 8B · INT4 NVIDIA test · 2 Oct 2024 Test conditions llama.cpp, CUDA backend. Input 100 tokens, output 100 tokens. Manufacturer-reported historical throughput.
Exact engine commit, model artifact revision and full concurrency/latency definition are not provided in this record.
Original test ↗ No quote collected
32B Qwen3 32B · Q4_K_M Memory estimate Models & memory Qwen3 8B : within budget · 7.2 GiB estimated
Qwen3 14B : within budget · 11.0 GiB estimated
Qwen3 32B : within budget · 21.4 GiB estimated
Llama 3.3 70B : over budget · 42.9 GiB estimated
No sourced test yet
No quote collected
No matching quote
NVIDIA
32 GB GDDR7
Consumer · Blackwell
Model fit, tests & prices → Hardware source Chip designer: NVIDIA. Desktop 32 GB reference specification; laptop and regional D variants are separate.
Manufacturer specifications ↗ Checked 7 Nov 2026. Specifications do not establish current retail stock.
32B Qwen3 32B · Q4_K_M Memory estimate Models & memory Qwen3 8B : within budget · 7.2 GiB estimated
Qwen3 14B : within budget · 11.0 GiB estimated
Qwen3 32B : within budget · 21.4 GiB estimated
Llama 3.3 70B : over budget · 42.9 GiB estimated
No sourced test yet
$1999 Launch starting price · reference NVIDIA announcement · US reference Price evidence Historical official starting price announced 6 Jan 2025. Not a current retailer offer; board SKU, tax, shipping and availability vary.
Current purchase price and stock not verified · observed 7 Nov 2026
Inspect listing ↗
32B Qwen3 32B · Q4_K_M Memory estimate Models & memory Qwen3 8B : within budget · 7.2 GiB estimated
Qwen3 14B : within budget · 11.0 GiB estimated
Qwen3 32B : within budget · 21.4 GiB estimated
Llama 3.3 70B : over budget · 42.9 GiB estimated
Up to 51.8 tok/s Qwen3.8 27B · Precision unpublished AMD test · Aug 2026 Test conditions Preliminary average token-generation throughput over at least three runs. llama.cpp, Vulkan, MTP=2, Windows 11 Pro 25H2; Ryzen 9 9950X, 64 GB RAM, Adrenalin 26.7.1.
Quantization, input/output lengths, model revision and exact engine commit are not published in the cited test description. Not comparable to the 4090 reference test.
Original test ↗ No quote collected
No matching quote
70B Llama 3.3 70B · Q4_K_M Memory estimate Models & memory Qwen3 8B : within budget · 7.2 GiB estimated
Qwen3 14B : within budget · 11.0 GiB estimated
Qwen3 32B : within budget · 21.4 GiB estimated
Llama 3.3 70B : within budget · 42.9 GiB estimated
No sourced test yet
No quote collected
70B Llama 3.3 70B · Q4_K_M Memory estimate Models & memory Qwen3 8B : within budget · 7.2 GiB estimated
Qwen3 14B : within budget · 11.0 GiB estimated
Qwen3 32B : within budget · 21.4 GiB estimated
Llama 3.3 70B : within budget · 42.9 GiB estimated
No sourced test yet
No quote collected
NVIDIA
96 GB GDDR7 ECC
Workstation · Blackwell
Model fit, tests & prices → Hardware source Chip designer: NVIDIA. Workstation Edition. Server Edition, Max-Q and MIG partitions are separate variants.
Manufacturer specifications ↗ Checked 7 Nov 2026. Specifications do not establish current retail stock.
70B Llama 3.3 70B · Q4_K_M Memory estimate Models & memory Qwen3 8B : within budget · 7.2 GiB estimated
Qwen3 14B : within budget · 11.0 GiB estimated
Qwen3 32B : within budget · 21.4 GiB estimated
Llama 3.3 70B : within budget · 42.9 GiB estimated
No sourced test yet
No quote collected
No matching quote
No GPUs match these filters. Try a smaller model, shorter context or another maker.
What these numbers tell you. Capacity is a planning estimate. We add the exact Q4_K_M file size, an FP16 KV cache for the selected total context, and 2 GiB for the runtime. One request, all weights on one GPU. Nominal manufacturer GB labels use a conservative decimal-GB budget (1 GB = 10⁹ bytes); actual free VRAM and engine allocations can differ. A fit is a candidate to validate, not a tested deployment or a maximum model size.
Driver, operating system, quantization and engine support must also be checked. CPU offload, multi-GPU and training are outside this estimate.
Model files & calculation KV bytes = 2 × layers × KV heads × head dimension × total context tokens × 2 bytes. GiB = 2³⁰ bytes. Context includes prompt and generated tokens. Architecture is pinned with the model metadata; allocation overhead remains an assumption.
Qwen3 8B · Q4_K_M — 4.68 GiB of files. Qwen/Qwen3-8B-GGUF; revision 7c41481f57cb. KV architecture ↗
Qwen3 14B · Q4_K_M — 8.38 GiB of files. Qwen/Qwen3-14B-GGUF; revision 530227a7d994. KV architecture ↗
Qwen3 32B · Q4_K_M — 18.40 GiB of files. Qwen/Qwen3-32B-GGUF; revision 938a7432affa. KV architecture ↗
Llama 3.3 70B · Q4_K_M — 39.60 GiB of files. bartowski/Llama-3.3-70B-Instruct-GGUF; revision b6c5c9f176f3. KV architecture ↗
Speed belongs to a test. Each speed record names the model, engine and published conditions. These two tests use different models and settings, so they are not a GPU ranking. Missing results stay unknown. Manufacturer reports are external evidence; we have not run these benchmarks.
Prices keep their conditions. Purchase references distinguish asking prices from official starting prices and show stock evidence. Rentals link to their dated configuration, cloud type and recorded availability. Taxes, shipping, storage and other fees may be extra. A workstation card does not inherit a similarly named server card’s rental quote.
Put the hardware in a workload budget. 