CryptoGPUs
Built to orderShips in 2–5 days
More from CryptoGPUs8 in AI workstations

CG AI Duo 48 Workstation (2x RTX 3090)

$2,615.57Was $2,906.19−10%

Black Friday price: you save $290.62 until Dec 1. No code needed.

Paid in Bitcoin, Monero, USDT or another coin you choose.

Free insured shipping. US delivery in about 5 to 11 business days. Shipping details

Overview

CG AI Duo 48 puts two refurbished RTX 3090s, 24GB each, on an ASRock X870E Taichi, the AM5 board whose two x16 slots sit four slots apart so the cards are not pressed together. A Ryzen 7 9700X, 64GB of DDR5, a Corsair HX1500i and a Fractal Meshify 2 XL complete the build.

Each card has 936GB/s of memory bandwidth, and the 48GB total holds a 70B model at 4-bit (about 40GB) or a 32B model at 8-bit with moderate context. Software splits the layers across the two cards, so generation runs at the pace of one card at a time rather than doubling. CUDA means llama.cpp, vLLM and Ollama run without extra work.

The trade-offs are age and heat. The 3090 is a 2020 design rated at 350W per card, so a pair needs the large case and 1500W supply listed here, and these cards are refurbished rather than new. Ampere also has no native FP8 or FP4 support, which newer RTX cards add.

  • 2x RTX 3090, 48GB VRAM total
  • 936GB/s memory bandwidth per card
  • Four-slot gap between GPU slots
  • 1500W power supply

Specifications

Performance
Graphics cards2x NVIDIA GeForce RTX 3090, 24GB GDDR6X each (refurbished, tested)
Total VRAM48GB (2 x 24GB), 384-bit bus per card
Memory bandwidth936GB/s per card
GPU detailsAmpere, 10,496 CUDA cores and 3rd-generation Tensor cores per card, 350W reference board power
ProcessorAMD Ryzen 7 9700X, 8 cores / 16 threads, up to 5.5GHz boost, 32MB L3, 65W TDP
Memory64GB DDR5 (2 x 32GB), up to DDR5-5600 with two modules
StorageNVMe SSD, capacity confirmed with your order
Local AI capacity70B at 4-bit, 32B at 8-bit, layers split across both cards
Platform
MotherboardASRock X870E Taichi, E-ATX, AM5, 4 DIMM slots, 5GbE, Wi-Fi 7
GPU slotsTwo PCIe 5.0 x16 slots four slots apart, x8/x8 with both populated; each 3090 links at PCIe 4.0 x8
CoolingArctic Freezer 36, dual tower, 2x P12 PWM PST fans, 159mm tall
Power supplyCorsair HX1500i (2025), 1500W, ATX 3.1, Cybenetics Platinum, 10-year warranty
Case and connectivity
CaseFractal Design Meshify 2 XL, 9 expansion slots, up to 530mm GPU length, 185mm cooler clearance
Case fans3x Dynamic X2 GP-14 preinstalled
Front I/O2x USB-A 5Gbps, 1x USB-C 20Gbps, audio and microphone jacks
Networking5Gbit Ethernet and Wi-Fi 7 with Bluetooth 5.4 on the motherboard
Software and warranty
Operating systemUbuntu 24.04 with Ollama, llama.cpp and Open WebUI, Ubuntu 24.04 with open models preloaded, or Windows 11 Pro (+$189.05)
Warranty3-year parts, lifetime support

Questions and answers

Still unsure? Visit the help center or message us.

Do the two cards act as one 48GB GPU?

Not as a single device. llama.cpp, vLLM and Ollama divide a model across the cards, so a 40GB model fits, but each card still holds only its own 24GB and the context cache is split too.

How fast can a 70B model generate?

Each token reads all the weights, and at 936GB/s a 40GB model has a ceiling near 23 tokens per second when the layers run one card after the other. Real speeds are lower than that ceiling.

Do two 3090s fit side by side?

The Taichi leaves four slots between its x16 slots. NVIDIA's own 3090 is a 3-slot, 313mm card, and we choose partner models no thicker than that.

Is an NVLink bridge included?

No. The 3090 supports NVLink, but inference software splits models over PCIe, and this build runs each card at PCIe 4.0 x8 from the CPU's x8/x8 slots.

Are the refurbished cards covered by warranty?

Yes. The cards are tested before assembly and covered by the same 3-year parts warranty and lifetime support as the rest of the build.

From the blog

All articles
Local AI

How much VRAM for local LLM work in 2026: a practical sizing guide

A practical 2026 sizing guide: add up weights, KV cache and overhead, check real GGUF sizes for current open models, and see what fits on 16GB to 512GB.

Pairs well with

Picked to go with this product. Check sizes and connectors for your build before you order.