sFNX demonstrates how three used V100s can share an AI model's work

A small open-source engine reports strong code-edit throughput, with workload-specific results and unusually explicit hardware limits.

An open-source project called sFNX is testing an alternative to buying new accelerator hardware: make three older cards work simultaneously on each layer of an AI model. Its repository reports running Qwen3.8-Flash-Next on three Tesla V100-32GB GPUs costing approximately $1,800 used, connected over PCIe without NVLink.

The headline speed needs its workload attached. The developer reports 367–382 tokens per second on greedy code-edit tasks, while sampled prose reaches 139 tokens per second. Six concurrent conversations produce a combined 280–291 tokens per second with short prompts. These are the project's measurements, not an independent Intellisphere hardware test.

The implementation divides work within each layer across all three cards. That contrasts with assigning whole layers to different devices, where a single request can leave cards waiting for their turn. Its approach is specifically engineered around the communication constraints of this older hardware. The repository carries an MIT licence.

For small teams, the interesting question is whether a tightly matched software stack can extract more useful service from equipment already available. A repair assistant serving short code changes has a different demand pattern from a prose service supporting many simultaneous customers. The same peak number should not price both workloads.

The project also documents narrow compatibility and limited spare GPU memory. That makes it a focused engineering experiment with practical reuse potential. Anyone evaluating the economics should include the rest of the machine, energy, maintenance and their own representative workloads alongside the used-card purchase price.

Sources

- [sFNX repository and benchmark documentation](https://github.com/johnfink8/sfnx) — johnfink8 / GitHub