NVIDIA and Kioxia Collaborate for New AI SSDs 100x Faster than Conventional Drives
With the realization of a new report, NVIDIA and its affiliates are ushering AI memory into the next phase, whereby Kioxia would soon make available commercial SSDs that probably can work up to 100 times faster than the usual configurations. Eventually, this is expected to replace High-Bandwidth Memory (HBM) in AI.
New Approach to AI Memory Architecture
According to a new report from Nikkei, the collaboration between Kioxia and NVIDIA establishes the path for specialized AI SSDs where high-speed storage devices will be mounted directly on the GPU, and include solutions providing far higher capacity than HBM.
To this end, NVIDIA's performance target for this new offering is an astounding 200 million IOPS (Input/Output Operations Per Second). Kioxia will most likely achieve this with two separate SSD configurations, potentially each reaching 100 million IOPS. Speed ratings for these drives require a whole new deal as far as the PCIe 7.0 connection is concerned, as well as an entire redesign of the memory subsystem.
The HBF and XL-Flash Technology
This movement away from HBM directly leads into more advanced NAND-based technologies, with the prime candidate being High-Bandwidth Flash (HBF), an original standard extended by SanDisk to address the problems with space and speed that classic NAND memories have.
High-Bandwidth Flash (HBF): The major benefit of HBF is the ability to attain terabytes in memory capacity by each device. This enormous pool of memory is particularly handy for data centers operating very large-scale AI inferencing workloads. Performance charts have shown HBF's capabilities being within a small percentage of what hypothetical HBM performance could achieve for AI tasks.
XL-Flash: Kioxia also has its own high-performance NAND technology, XL-Flash, which presents another potential path to achieving NVIDIA's ambitious performance targets.
AI Memory A Bright Future
This explains why HBM's inherent capacity limitations are being strategically overcome. The future of AI systems may well have greatly scaled and even more powerful technologies using advanced high-performance NAND solutions. Thus, NAND technology is fast going to be an important portion of the future foundation of AI memory.
