← All machines

GPU desktop (prebuilt) · tiny corp

tiny corp tinybox red flash

tiny corp's AMD box with six 32GB Radeon cards in a water-cooled EPYC server chassis, made to order. It gives 192GB of VRAM spread over six GPUs; the blower fans are audible.

Memory for models

184.5 GB

of 192 GB in the largest config

Memory bandwidth

640 GB/s

per card; 6 cards take turns on each token

Maker price

~$22,000

Largest at 4-bit

Hy3

173.6 GB at 8,192 tokens of context

gpt-oss-20b, 4-bit

~80–117 tok/sestimate

no measurement on this hardware yet

Memory configurations

One page covers every memory option the maker sells. Prices are the maker’s list price for the cheapest configuration with that much memory.

MemoryPriceMemory the GPU can use
192 GB~$22,000
184.5 GB gpu-usable memory

6 x 32GB VRAM across six cards (tiny corp configurator); models must be split across GPUs. Less 1.5 GB on each of the other 5 cards for its own runtime buffers. How GPU memory limits work on unified-memory machines.

What it runs at 4-bit

Every model in the catalog at its 4-bit file with 8,192 tokens of context. 32 of 41 fit. Speeds are the median of measured runs on the same GPUs where any exist; otherwise an estimate from memory bandwidth. Try other quants and context lengths.

ModelNeedsFitGeneration speed
Kimi K32.8T, 104B active · UD-Q4_K_XL1406.8 GBcontext memory estimatedDoes not fit
DeepSeek V4 Pro1.6T, 49B active · Q4_K_M890.5 GBcontext memory not countedDoes not fit
Kimi K2.61T, 32B active · UD-Q4_K_XL545.7 GBcontext memory estimatedDoes not fit
DeepSeek V4.1 Flash552B, 16B active · Q4 (DwarfStar)484.5 GBDoes not fit
GLM-5.3753.3B · UD-Q4_K_XL437.4 GBcontext memory estimatedDoes not fit
DeepSeek V3.2671B, 37B active · Q4_K_M379.6 GBcontext memory estimatedDoes not fit
MiniMax M3428B, 23B active · Q4_K_M244.8 GBcontext memory not countedDoes not fit
Qwen3.5 397B-A17B397B, 17B active · Q4_K_M229.1 GBcontext memory estimatedDoes not fit
GLM-5.3 Flash320B, 18B active · UD-Q4_K_XL187.6 GBcontext memory estimatedDoes not fit
Hy3295B, 21B active · Q4_K_M173.6 GBFits~12–18 tok/sestimate
MiniMax M2.7228.7B · Q4_K_M132.7 GBFits–
Qwen3.8 Flash-Next125B, 6B active · Q4_K_M113.1 GBcontext memory estimatedFits~28–40 tok/sestimate
Nemotron 3 Super 120B-A12B120B, 12B active · Q4_K_M82.6 GBcontext memory estimatedFits~18–27 tok/sestimate
Qwen3.5 122B-A10B122B, 10B active · Q4_K_M73.0 GBcontext memory estimatedFits~25–37 tok/sestimate
Mistral Medium 3.5 128B128B · Q4_K_M74.0 GBFits~4.2–6.6 tok/sestimate
GLM-4.5-Air106B, 12B active · Q4_K_M70.9 GBFits~19–28 tok/sestimate
Mistral Small 4 119B (2603)119B, 6.5B active · Q4_K_M69.3 GBcontext memory estimatedFits~40–59 tok/sestimate
Llama 4 Scout 17B-16E Instruct109B, 17B active · Q4_K_M62.4 GBcontext memory not countedFits~16–23 tok/sestimate
gpt-oss-120b117B, 5.1B active · MXFP4 (native)60.2 GBcontext memory estimatedFits~58–85 tok/sestimate
Qwen3 Coder Next80B, 3B active · Q4_K_M46.9 GBcontext memory estimatedFits~88–128 tok/sestimate
Llama 3.3 70B Instruct70B · Q4_K_M43.6 GBFits~7.4–12 tok/sestimate
Nemotron 3.5 Lightning 30B-A3B30B, 3B active · Q4_K_M25.3 GBcontext memory estimatedFits~63–91 tok/sestimate
Qwen3.6 35B-A3B35B, 3B active · Q4_K_M22.4 GBcontext memory estimatedFits~83–122 tok/sestimate
Ornith 1.5 35B-A3B35B, 3B active · Q4_K_M21.9 GBcontext memory estimatedFits~86–125 tok/sestimate
Qwen3 Coder 30B-A3B Instruct30.5B, 3.3B active · Q4_K_M19.5 GBFits~79–116 tok/sestimate
Gemma 4 31B IT30.7B · Q4_K_M18.9 GBcontext memory estimatedFits~17–27 tok/sestimate
Granite 4.2 30B29.3B · Q4_K_M20.3 GBFits~17–27 tok/sestimate
Qwen3.8 27B27B · Q4_K_M18.2 GBcontext memory estimatedFits~18–28 tok/sestimate
Gemma 4 26B-A4B IT25.2B, 3.8B active · Q4_K_M17.4 GBcontext memory estimatedFits~62–91 tok/sestimate
Qwen3.6 27B27B · Q4_K_M17.7 GBcontext memory estimatedFits~19–29 tok/sestimate
Devstral Small 2 24B (2512)24B · Q4_K_M16.1 GBFits~22–34 tok/sestimate
gpt-oss-20b21B, 3.6B active · MXFP4 (native)12.5 GBcontext memory estimatedFits~80–117 tok/sestimate
Phi-414.7B · Q4_K_M11.3 GBFits~35–55 tok/sestimate
Ministral 3 14B Reasoning (2512)13.9B · Q4_K_M10.4 GBFits~38–60 tok/sestimate
Gemma 4 12B IT11.95B · Q4_K_M8.2 GBcontext memory estimatedFits~44–69 tok/sestimate
Qwen3.5 9B9B · Q4_K_M7.0 GBcontext memory estimatedFits~55–87 tok/sestimate
Granite 4.2 8B8.8B · Q4_K_M7.9 GBFits~57–89 tok/sestimate
Gemma 4 E4B IT8B · Q4_K_M6.5 GBcontext memory not countedFits~103–162 tok/sestimate
Qwen3 8B8.2B · Q4_K_M7.3 GBFits~62–98 tok/sestimate
Llama 3.1 8B Instruct8B · Q4_K_M7.1 GBFits~64–100 tok/sestimate
Qwen3.5 4B4B · Q4_K_M4.3 GBcontext memory estimatedFits~114–180 tok/sestimate

Measured speeds

Nobody has published a run on this GPU combination with a stated model, quant and backend that we could find. The speeds above are estimates.

Specifications

Chip
6x AMD Radeon AI PRO R9700 (32GB each) + 32-core AMD EPYC
Memory
GDDR6 on GPUs (32GB x6); DDR4 system RAM
Memory bandwidth
640 GB/s per card (6 cards)
FP16 compute
tiny corp lists 640 GB/s per card (3,840 GB/s aggregate across six cards, not one pool); no TFLOPS figure published
Software
ROCm, Vulkan
Ships with
not stated
Power
1700 W (1700W PSU (tiny corp); GPUs limited to 225W each)
Form factor
Tower
Networking
2x 1GbE + OCP 3.0 slot
Maker page
tinygrad.org

Sources

  • tinygrad.org: made to order, ships 1-8 weeks, 1700W PSU, network, chassis
  • tinygrad.org: preset prices, R9700 32GB 640 GB/s per card, 225W limit, noise note; RTX 6000 configs flagged out of stock

Similar machines