Big GPUs don't need big PCs
a day ago
- #Performance
- #eGPU
- #Raspberry Pi
- The Raspberry Pi 5 can run multiple GPUs despite having only 1 lane of PCIe Gen 3 bandwidth.
- Performance tests included Jellyfin transcoding, GravityMark rendering, LLM/AI tasks, and multi-GPU applications.
- The Pi often matches or slightly underperforms a modern PC but excels in efficiency, sometimes by a large margin.
- A setup with four Nvidia RTX A5000 GPUs on a Pi achieved nearly the same performance as a high-end server for LLM tasks.
- Cost comparison shows a Pi eGPU setup costs $350-400 vs. $1500-2000 for a comparable PC.
- Idle power consumption is significantly lower on the Pi (4-5W) compared to the PC (30W).
- Jellyfin transcoding works well on the Pi for most use cases, though raw throughput is better on a PC.
- GravityMark benchmarks showed the Pi can sometimes outperform a PC with older GPUs like the RX 460.
- AI performance on the Pi is competitive, especially with Nvidia cards, and often more efficient.
- Multi-GPU setups on the Pi are possible but suffer from performance overhead due to PCIe limitations.
- The Pi is ideal for GPU-bound workloads where efficiency and cost are priorities over peak performance.