Flux Computing offers a compute platform that combines SRAM memory cells with on‑chip photonic waveguides, delivering ultra‑low‑latency, high‑bandwidth data transfer while reducing energy use. This silicon‑photonic architecture surpasses traditional HBM‑based GPU memory in throughput per milliwatt and total cost of ownership, enabling scalable, interactive AI inference and training for hardware manufacturers, data centers, and enterprise AI providers.
Funding
Funding not disclosed
Founders
Product
Problem
Current GPU memory architectures, based on HBM and silicon‑only SRAM, are constrained by physical design and packaging limits that prevent simultaneous high throughput, low latency, and low cost. These constraints hinder the scaling of AI workloads and limit accessibility of advanced compute capabilities.
Solution
Flux Computing proposes a new compute paradigm that integrates SRAM‑based memory with photonic interconnects, creating a scalable architecture that can exceed the performance‑per‑watt and total cost of ownership of traditional HBM solutions. By leveraging photonics for data movement, the design reduces energy consumption and latency while maintaining high bandwidth. The approach enables higher interactivity for AI models and supports larger workloads without the cost penalties of existing silicon‑only designs. This architecture is positioned to deliver the throughput needed for next‑generation AI applications at a more affordable price point.
Target Audience
Primary customers are AI hardware manufacturers, data center operators, and enterprise AI solution providers seeking high‑performance, cost‑effective compute platforms for large‑scale machine learning workloads.
Features
- SRAM memory cells combined with on‑chip photonic waveguides for ultra‑low‑latency data transfer
- Photonic interconnects that provide higher bandwidth per milliwatt compared to electrical HBM links
- Scalable modular design allowing easy expansion of memory capacity without proportional cost increase
- Integrated silicon‑photonic packaging that overcomes traditional GPU memory physical constraints
- Optimized for AI inference and training workloads requiring high interactivity and low latency