Parasail Combines NVIDIA GPUs with D-Matrix Accelerators for Faster Inference

Parasail is deploying D-Matrix Corp. Corsair accelerators with NVIDIA Hopper and Blackwell GPUs to deliver faster and more efficient inference services, announced in a press release. The integration aims to achieve up to ten times faster token generation and improved cost performance for Parasail customers.

The deployment marks one of the first large-scale examples of heterogeneous disaggregated inference, where NVIDIA infrastructure handles compute-intensive prefill tasks and D-Matrix Corsair accelerators manage latency-sensitive decode operations. Parasail said this approach allows it to extend the lifespan and utilization of its existing GPU fleet across data centers.

D-Matrix Corsair accelerators use Digital In-Memory Compute architecture that combines compute and memory on the same chip. This design reduces data movement between components, improving both speed and energy efficiency. According to the companies, Corsair enables up to ten times faster interactive inference and three times better energy efficiency than traditional GPU setups.

Parasail plans to scale this configuration across its network of more than forty data centers in fifteen countries. D-Matrix's Corsair platform is available to select customers, while Parasail's inference cloud services are already in production.