I put a datacenter GPU in my gaming PC
blog.tymscar.com
May 31, 2026
14 min read
🔥🔥🔥🔥🔥
63/100
Summary
A datacenter GPU was installed in a gaming PC using an adapter, providing a total of 32GB of VRAM across two GPUs. This setup allows for running a 27 billion parameter model at 32 tokens.
Key Takeaways
- A Tesla V100 SXM2 GPU with 16GB of HBM2 memory was purchased for about £150 and adapted for use in a gaming PC, providing a total of 32GB of VRAM when combined with an RTX 4080.
- The V100 GPU offers 900 GB/s of memory bandwidth, surpassing the bandwidth of newer consumer GPUs, including the RTX 4080 and Apple’s M3 Max.
- An SXM2-to-PCIe adapter was used to connect the V100 GPU to the gaming PC, costing around £50, allowing for significant cost savings compared to purchasing a high-end consumer GPU.
- The V100's cooling fan operates at 82 decibels and cannot be controlled, making it unsuitable for quiet environments.
Community Sentiment
Positives
- Decommissioned datacenter GPUs like the NVIDIA V100 and AMD MI50 are becoming accessible for local experimentation, fostering a community of enthusiasts who keep these models relevant.
- The cottage industry of 3D-printed fan shrouds for datacenter GPUs enhances cooling efficiency, which is crucial for maintaining performance during intensive tasks.
- The AMD MI250X offers impressive specifications with 128GB of HBM2E memory, making it a compelling option for high-throughput AI applications, despite connection challenges.
Concerns
- The NVIDIA V100's lack of bfloat16 support highlights its aging hardware features, which could limit its effectiveness for modern AI workloads.
- Slow prefill times in AI models can severely hinder performance, particularly for applications requiring rapid response, which is a significant drawback for local deployments.
- The complexity of integrating advanced GPUs like the MI250X into standard systems may deter users, as proprietary requirements can complicate accessibility.
