d-Matrix is developing AI inference processors based on digital in-memory compute (DIMC) architecture, which performs calculations directly where data is stored rather than shuttling data between memory and compute units. This approach is specifically designed to deliver low-latency, energy-efficient inference for large language models.
Core Products
Corsair is d-Matrix’s first-generation AI inference processor using DIMC architecture. The chip uses a chiplet design that combines compute and memory in a single package, eliminating the memory bandwidth bottleneck that limits traditional GPU inference. The Nighthawk platform provides the system-level solution for deploying Corsair in data centers.
Competitive Position
d-Matrix’s in-memory compute approach attacks the fundamental bottleneck in LLM inference: moving data between memory and compute. By performing operations in place, the architecture can theoretically achieve both lower latency and lower power consumption than GPUs. Strategic investments from Microsoft, Samsung, and SK Hynix validate the approach and provide manufacturing and deployment partnerships.
Recent Developments
d-Matrix raised $110 million in 2023 to bring Corsair to production. The company has been working with early customers to validate performance on real LLM workloads. Manufacturing partnerships with memory companies (Samsung, SK Hynix) ensure access to advanced process technology. The company has begun sampling chips to select customers.
Outlook
d-Matrix is at an earlier stage than Groq or Cerebras but targets a similar opportunity: purpose-built inference hardware that outperforms GPUs for LLM workloads. The in-memory compute approach is theoretically compelling, and strategic investor backing provides validation. Success depends on achieving production-quality chips and demonstrating real-world advantages over rapidly improving GPU inference.