- GPU-like PCIe card offers 10PFLOPs FP4 compute power and 2GB of SRAM
- SRAM is typically used in small amounts for cache in processors (L1 to L3)
- Leverages LPDDR5 memory instead of costly HBM
Emerging from Silicon Valley, d-Matrix, a startup backed by Microsoft, has introduced a groundbreaking chiplet-based solution designed to accelerate AI inference in enterprise environments without the need for traditional GPUs. The Corsair, the company’s flagship product, introduces a novel approach by leveraging an all-digital compute-in-memory architecture, utilizing modified SRAM cells for unprecedented speed and efficiency.
The Corsair PCIe card integrates two d-Matrix Application-Specific Integrated Circuits (ASICs) on a full-height, full-length card, each featuring four chiplets. This design delivers 9.6 PFLOPs FP4 compute power with 2GB of SRAM-based performance memory, addressing the memory wall challenges often associated with existing AI architectures. Notably, the Corsair uses LPDDR5 memory instead of the more expensive HBM, enabling up to 256GB per card to handle larger models and batch inference workloads efficiently.
A Leap in AI Compute Performance
d-Matrix claims that the Corsair offers a 10x improvement in interactive performance, 3x enhancement in energy efficiency, and 3x cost-performance ratio when compared to leading GPU alternatives like NVIDIA’s H100.
Innovative Memory-Compute Integration
Sree Ganesan, d-Matrix’s head of product, explained in an interview with EE Times, “Today’s architectures hit a memory wall. We need to innovate on memory bandwidth and directly compute within memory to overcome these challenges. This approach is key to building a sustainable AI infrastructure.” By integrating computation directly within memory, the Corsair achieves an impressive 150 terabytes of bandwidth per second, significantly mitigating latency issues.
Betting on Inference and Transformers
CEO Sid Sheth discussed the company’s origins, stating, “d-Matrix was founded in 2019 on the insight from hyperscalers that inference was the future. It was a leap of faith, and we placed our bets on transformer networks early on. This foresight was validated following the surge in interest triggered by ChatGPT in 2022.”
Future Developments
The Corsair is set to enter mass production in Q2 2025, positioning d-Matrix at the forefront of the AI hardware market. The company is already planning its next generation of ASICs. The Raptor will incorporate 3D-stacked DRAM, aiming to support reasoning workloads and larger memory capacities, expanding the horizon for what the company can achieve.
Sid Sheth’s vision for d-Matrix is clear: “Our focus is on creating a scalable and efficient AI infrastructure that can power the next generation of intelligent applications. With the Corsair, we’re just getting started.”
Commentary and Engagement
The arrival of the Corsair from d-Matrix heralds a new era in AI compute, offering significant advantages in performance, efficiency, and affordability. As the market continues to evolve, solutions like the Corsair will play a critical role in shaping the future of the industry. Stay tuned for more updates from d-Matrix and the broader AI landscape.
We invite you to share your thoughts and experiences related to d-Matrix’s Corsair and AI hardware advancements. Your insights can inspire others and contribute to valuable discussions.
Like, share, and subscribe to Archynetys for the latest in tech news, innovations, and expert opinions. Your support helps us deliver high-quality content that fuels your curiosity.
