EdgeCortix Looks To Chiplets For Third-Gen Reconfigurable AI Chip | AI With Sally
Why It Matters
EdgeCortix’s chiplet‑based, dynamically reconfigurable AI processor delivers edge‑centric performance and flexibility, giving manufacturers a resilient, cost‑effective solution as AI models and supply‑chain dynamics evolve.
Key Takeaways
- •EdgeCortix focuses on latency‑centric AI inference for edge devices.
- •DNA architecture dynamically reconfigures compute blocks to boost performance per watt.
- •Sakura 2 supports mixed precision from BF16 to 4‑bit integer formats.
- •Chiplet‑based third‑gen design aims to improve scalability and flexibility.
- •Partnerships with NASA and Japanese ecosystem bolster market credibility.
Summary
EdgeCortix’s latest announcement centers on its third‑generation, chiplet‑based AI inference processor, extending the Sakura line with a reconfigurable data‑flow architecture dubbed DNA. The company positions the new silicon as a solution for latency‑critical edge workloads, emphasizing performance‑per‑dollar and performance‑per‑watt metrics rather than raw throughput. The DNA engine introduces a dynamic compute bus that can be statically compiled yet re‑wired at runtime, allowing individual compute blocks to be enabled or disabled per layer. This flexibility supports mixed‑precision inference—from BF16 down to 4‑bit integer—delivering up to 30 TFLOPs of BF16 compute and achieving more than 20 tokens per second on 8‑billion‑parameter LLMs while preserving power efficiency. Saki Das Gupta highlighted real‑world deployments, including NASA space‑flight applications and a strategic foothold in Japan’s semiconductor ecosystem. He noted that Sakura 2’s dynamic reshaper engine and chiplet integration enable higher utilization of on‑chip resources, reducing off‑chip memory traffic and improving deterministic latency compared with GPU‑style dynamic compilation. The move toward chiplet modularity and runtime reconfigurability positions EdgeCortix to address the rapidly evolving transformer landscape, offering customers a scalable, future‑proof edge AI platform that can adapt to new model formats without costly hardware redesigns.
Comments
Want to join the conversation?
Loading comments...