Two DGX Spark units run a large model fast enough for agents
Two DGX Spark units linked together can run a large MoE model such as at speeds that may work for AI agent use with . The setup needs two DGX Spark machines and a special cable that costs about $180, which gives a much faster connection between the machines.
In the shared benchmark, two DGX Spark units using vLLM and FP8 reached about 41 for one request and about 350 in total when handling many requests at once. reached about 46.9 for one request, while a M2 Ultra with 192GB memory reached about 29.7 .
One DGX Spark alone reached about 14 . The two-unit DGX Spark setup stands out because FP8 helps speed, and it can serve many requests at the same time.
Key points
- Two DGX Spark units are needed to reach the more useful agent-serving speed.
- The setup mentions a special connection cable costing about $180.
- Two DGX Spark units with vLLM and FP8 reached about 41 for one request.
- Total throughput reached about 350 with many requests running at once.
- One DGX Spark alone reached about 14 , much slower than the two-unit setup.