IT Engineer Shares 6-Month Review of 256GB VRAM 'Data Center on Wheels'
SweetHomeAbalama0 · reddit · 2026-08-03
An IT infrastructure engineer shared a detailed 6-8 month operational review of their custom 256GB VRAM AI server (8x3090 + 2x5090).
Hardware & Goals
- Features an AMD Threadripper 64-core CPU, 512GB RAM, and a combined 2900W dual-PSU setup.
- Designed to bypass API limits, enabling simultaneous frontier MoE inferencing and ComfyUI image generation for internal business tasks like data analysis and design.
Performance
- Thanks to chassis airflow modifications and hybrid cooling, idle temps hover in the mid-40s°C, peaking around the 60s°C under sustained inference loads.
- The author notes this setup is ideal for power users or creative professionals who frequently hit API limits, but not recommended for casual users or model training.
More from Infra
- K3 2.8T Model Hits 947 tokens/s Decoding on Single B300 Node — casper_hansen_ · 2026-08-03
- Teknium: DeepSeek V4 Flash Hits 70 tok/s on Local Inference — Teknium · 2026-08-03
- Debate: Open-Source Models Are Good Enough, OpenAI's Real Moat is Gigawatts of Compute — yacineMTB · 2026-08-03
- Fluidstack Mass Hires Data Center Engineers to Build Gigawatt-Scale AI Facilities — MxMnr · 2026-08-03
- Fully Offline: Running a Custom Q/A Model on ESP32S3 Microcontroller — slvDev_ · 2026-08-03
- NVIDIA Partners with Safe Superintelligence to Expand Compute Access via Vera Rubin — thione · 2026-08-03