Used condition, pulled from working service
-
🔥 GENUINE ORIGINAL NVIDIA — NOT SXM2, NOT modded. Part Number TCSV100M-32GB-PB (or equivalent NVPN 699-2G500-0200 series). Native PCIe 3.0 ×16 edge connector — drops directly into any standard PCIe ×16 slot, no SXM2-to-PCIe adapter needed (unlike the $50 cheaper SXM2 variant that requires $60+ adapter + separate cooler).
-
💾 32GB HBM2 @ 900 GB/s — True 32GB usable. 4096-bit memory bus, 900 GB/s bandwidth — the highest memory bandwidth in its price class for 2026. Handles 30B Q4_K_M LLMs, 14B full precision, 70B Q2 comfortably. Q4_K_M on 30B models is the recommended sweet spot.
-
🧠 5120 CUDA + 640 Tensor Cores (Volta). 14 TFLOPS FP32, 112 TFLOPS FP16 tensor — solid for local LLM inference, diffusion models, scientific HPC. Compute capability 7.0, fully supported by CUDA Toolkit, PyTorch, TensorFlow, Ollama (sm_70).
-
🏗️ Passive-cooled, server/HPC grade. No on-board fan — relies on chassis front-to-back airflow (designed for 1U/2U/4U server racks). ECC enabled by default for 24/7 reliability. If installing in a desktop PC, must add blower fan ducting — this is NOT a plug-and-play gaming card.
-
⚠️ HEADLESS — No display outputs. Zero HDMI/DP/DVI. Cannot drive a monitor. Must pair with integrated GPU or a cheap display card (GT 710/1030) for video out. This is a pure compute accelerator, not a graphics card.