Parasail is deploying d-Matrix Corsair inference accelerators alongside NVIDIA Hopper and NVIDIA Blackwell GPUs to speed up AI inference on its platform. The company is targeting up to 10x faster token generation and “more cost-efficient” inference for customers by splitting inference work across different types of silicon.
The deployment pairs NVIDIA GPUs with Corsair in a heterogeneous, disaggregated inference setup. Parasail plans to use GPUs for compute-heavy prefill, then hand off latency-sensitive decode to d-Matrix Corsair accelerators. Parasail describes the approach as a production, commercial-scale example of running NVIDIA AI infrastructure and purpose-built inference accelerators together, with each handling the phase of inference it’s best suited for.
For data center operators, the practical implication is straightforward: if Parasail can shift decode off GPUs without creating new bottlenecks, it can raise useful throughput per deployed GPU and stretch existing capacity while new data center approvals and buildouts take time. But the headline gains here are still “up to” claims, and the real test will be workload-by-workload results once the first deployments are in and measured.
Parasail says its automatic kernel optimization technology routes workloads to the right hardware at the right time across the heterogeneous fleet, with the goal of extracting better performance and value per model at scale.
d-Matrix attributes Corsair’s performance to its Digital In-Memory Compute (DIMC) chiplet architecture, which integrates compute and memory on the same silicon to avoid repeated transfers between separate processor and memory chips during inference. d-Matrix says Corsair can deliver up to 10x faster interactive inference and up to 3x better energy efficiency versus “traditional approaches.” Corsair is built on TSMC’s N6 process, and d-Matrix lists organic substrates and LP-DDR5 memory as part of the platform.
“We run large Hopper and Blackwell fleets today and pairing them with Corsair lets us deliver our customers the edge they need while also extending the life of the hardware we’ve already deployed,” said Mike Henry, founder and CEO of Parasail.
The companies say they’ll share detailed performance results and case studies after the first series of deployments. They also say they plan to explore expanded integration across Parasail’s global fleet of more than 40 data centers in 15 countries.
d-Matrix says the Corsair inference platform is available now for select, qualified customers, and Parasail says its inference services are available now.
Source: d-Matrix
