Search For sale
Use the down arrow to enter the dropdown. Use the up and down arrows to move through the list, and enter to select. To remove the current item in the list, use the tab key to move to the remove button of the currently selected item. Use Escape to close the dropdown and return to the search box.
No results found
Navigate Categories by tabbing to the major areas. Use the Down key to open the subcategory and then use the Down key once more to enter it. Navigate the subcategory using the Up and down keys. Once on the subcategory you are interested in. Use the Enter key to navigate to that page.

Take action

$2,000

$2,500

AI server. 3x MI50 32gb, 96gb total

Last updated 2 days ago in Santa Ana, CA

Listed in categories:

Chat securely on the app

Image #1

Description

96GB VRAM AI server for running large language models locally. Fully set up and working - powers on straight into a web chat interface with three models loaded and ready. SPECS GPUs: 1x EVGA GT 1030 2gb (for display), 3x AMD Vega 20 32GB = 96GB HBM2 total (2x Radeon Pro Vega II, 1x Instinct MI50 - same gfx906 chip, all three running at full x16) CPU: Intel i7-6900K, 8 core / 16 thread, 3.2GHz Motherboard: ASUS X99-E WS, true triple x16 via PLX switch RAM: 32GB DDR4-2400 (4x8GB, 4 slots open, takes up to 128GB) Storage: 2TB Sabrent Rocket Q NVMe PSU: 1x Superflower 1300w, 1x Superflower 800w Cooling: 40mm fan on each card, holds 48-54C under sustained full load SPEEDS I MEASURED ON THIS MACHINE GPT-OSS 20B - 106 tokens/sec Qwen3 30B - 74 tokens/sec Gemma 4 26B - 70 tokens/sec Qwen3.6 35B - 67 tokens/sec Qwen3.8 27B - 22 tokens/sec Running Ubuntu 24.04 with ROCm and llama.cpp already built and configured for these cards, plus Open WebUI web interface. I'll include my full setup documentation - saves you a lot of trial and error.

All content is available to screen readers from the outset. The See more button is for visual users only to expose content incrementally that is already available to you

Details

Condition

Used (normal wear)

All content is available to screen readers from the outset. The See more button is for visual users only to expose content incrementally that is already available to you

Item location map

Map is approximate to keep seller’s location private.

Similar items

Tim's other offers