Kioxia Unveils 5TB Flash Prototype for Faster AI and Edge Computing
▼ Summary
– Kioxia has developed a prototype 5TB flash memory module with 64GB/s bandwidth, intended for edge computing and as a potential DRAM alternative.
– The module uses a daisy-chain design that maintains high bandwidth even as capacity increases, overcoming previous architectural limitations.
– It employs PAM4 signaling and high-speed transceivers to double bandwidth and achieve 128Gbps throughput per transceiver.
– The prototype consumes under 40W and features controller prefetching and low amplitude signaling to reduce latency and improve speed.
– This technology targets servers in mobile edge computing for real-time IoT and AI workloads, with plans for commercialization in big data and generative AI.
Kioxia has introduced a groundbreaking 5TB flash memory prototype designed to meet the escalating demands of AI and edge computing applications. This innovation delivers an impressive 64GB/s bandwidth, positioning it as a robust alternative to traditional DRAM in high-performance environments. Developed as part of Japan’s national research initiative, the module addresses the persistent challenge of balancing storage capacity with data transfer speeds.
The prototype employs a daisy-chain architecture where individual flash units, referred to as beads, are linked sequentially, each equipped with its own controller. This configuration allows for significant capacity expansion without sacrificing bandwidth, a notable improvement over earlier designs constrained by bus-based connections.
To achieve such high throughput, Kioxia integrated advanced PAM4 signaling technology, which encodes two bits per signal, effectively doubling data transmission rates compared to conventional binary methods. The memory interface operates at 4.0Gbps, supported by low-amplitude signaling and distortion correction mechanisms. Additionally, controller-level enhancements like prefetching contribute to reduced read latency, further optimizing performance.
Power efficiency remains a key strength, with the module consuming under 40W even at peak operation across an eight-lane PCIe Gen 6 interface. This balance of high throughput and manageable energy use makes the technology particularly suitable for memory-intensive tasks typically reliant on DRAM.
Targeted at mobile edge computing servers, the prototype responds to the growing need for localized data processing driven by 5G and 6G networks. Its capabilities align well with real-time IoT applications and AI workloads, where cost, capacity, and speed must be carefully balanced.
Kioxia aims to advance this technology toward commercial deployment, with potential applications spanning IoT infrastructure, big data analytics, and generative AI systems.
(Source: TechRadar)
