Tesla Referral M10 PACKS FOURINDEPENDENT GPUS ONTO A SINGLE CARD, EACH WITH ITS OWN 8GB, FOR 32GB TOTAL. ITIS ONE OF THE CHEAPEST WAYS TO PUT 32GB OF NVIDIA VRAM IN A CHASSIS.

• FOUR GPUS FROM ONE PCIE SLOT. PASS ONE TO EACH VM IN PROXMOX, ESXI, OR UNRAID.
• MULTI-GPU LLM INFERENCE. LLAMA.CPP AND OLLAMA CAN SPLIT A MODEL ACROSS ALL FOUR GPUS.
• FOUR PARALLEL 8GB WORKLOADS: TRANSCODING, CUDA JOBS, RENDER NODES, VDI SEATS.
• NVENC H.264 ACROSS FOUR GPUS FOR PLEX OR JELLYFIN STREAM DENSITY.
• HOMELAB LEARNING RIG FOR MULTI-GPU ORCHESTRATION WITHOUT THE MULTI-GPU BILL.

• THE 32GB IS 4X 8GB, NOT ONE 32GB POOL. SOFTWARE MUST BE MULTI-GPU AWARE TO USE ALL OF IT.
• PASSIVE CARD. IT NEEDS SERVER CHASSIS AIRFLOW OR A FAN SHROUD. IT WILL OVERHEAT IN A DESKTOP CASE WITHOUT ONE.
• POWER IS 8-PIN EPS (CPU STYLE), NOT PCIE 8-PIN. USE THE CORRECT ADAPTER CABLE.
• ENABLE ABOVE 4G DECODING IN BIOS.
• NO VIDEO OUTPUT. YOU NEED A SECOND GPU OR IGPU FOR DISPLAY.
• MAXWELL RUNS ON NVIDIA'S LEGACY DRIVER BRANCH. IT IS A CAPACITY AND DENSITY CARD, NOT A SPEED CARD, AND IT IS NOT FOR GAMING.