KIOXIA’s Maulik Sompura joins Brian for an in-depth, informative dialogue about all issues flash and different business occasions. The subject is well timed and related, given the large enlargement of AI and trendy workloads.
Maulik Sompura is the Senior Employees Director—Product Planning and Administration at KIOXIA, with over 13 years of expertise in NAND, reminiscence, and SSDs, together with the final six years at KIOXIA. Moreover, his profession consists of stints with Intel and Toshiba.
The dialog extends past storage and reminiscence, transitioning to liquid cooling, hyperscalers, and kind elements. In addition they focus on open-source software program, new fabs, issues about scaling, and fears of one other “Covid-boom-and-bust” occasion.
It’s an informative dialog that may immediate you to consider one thing in the long run. We’ve damaged the podcast into five-minute segments, so you’ll be able to bounce to the part that’s most related to you and your atmosphere. Nevertheless, it’s nicely definitely worth the time to observe this in its entirety.
Brian and Maulik kick off the dialogue with how radically SSDs and flash storage have advanced, particularly underneath the stress of AI and hyperscale knowledge facilities.
0:00–5:00: Hyperscalers Quietly Operating the Present
- SSD portfolios have exploded in complexity: many kind elements, endurance ranges, and efficiency tiers.
- The AI growth has moved storage proper behind GPUs/HBM as a first-class precedence.
- Hyperscalers (Meta, Google, Amazon, and many others.) dictate detailed drive necessities, and people customized designs in the end form what enterprises get.
- OCP specs assist unify and simplify the chaos by giving everybody a standard goal on high of NVMe.
- The business is actively attempting to shrink the variety of SKUs and converge on trendy kind elements like E3, at the same time as legacy kind elements linger.
5:00–10:00: E3, E2, and Petabyte-Class Drives
- E3 is turning into the popular enterprise kind issue for Gen6+ because of its higher energy and thermal traits.
- 2.5″ drives will stick round for years because of lengthy product lifecycles, however new designs are shifting to EDSFF.
- E2 is a bigger, capacity-first SSD kind issue, seen as a real exhausting drive alternative at excessive capacities.
- With superior NAND stacking, a single E2 drive might finally attain 1 PB.
- Hyperscalers mitigate the dangers of big drives utilizing sharding, erasure coding, and predictive analytics; rebuild instances for enormous SSDs are nonetheless a lot better than for giant HDDs.
10:00–15:00: Liquid Cooling Comes for SSDs Too
- AI coaching methods are driving widespread adoption of liquid cooling, and that now extends all the way down to storage.
- SSDs in E3 kind elements are being redesigned for direct contact with chilly plates (floor flatness, supplies, and connector particulars).
- Liquid-cooled Gen5 SSDs will pave the way in which for Gen6, the place energy and warmth go even increased.
- Enterprises are cautious about bringing water into the info heart, however rising energy prices and GPU density will make liquid cooling more durable to keep away from.
15:00–20:00: AI Storage: Capability, Efficiency, and Cloth Stress
- AI pushes storage alongside two axes: large capability and really excessive efficiency.
- KIOXIA’s 245 TB-class drive (with 32‑excessive QLC stacks) targets knowledge lakes and enormous repositories, not simply generic workloads.
- Inside a contemporary field, storage can push 250–280 GB/s; the bottleneck more and more turns into the community cloth, not the drives.
- Hyperscalers are racing to 400/800 GbE and smarter NICs/DPUs, whereas mid-market prospects battle to maintain up with that stage of material funding.
- Conventional databases and Fibre Channel stay, however vector databases and AI-centric knowledge shops are becoming a member of the combination.
20:00–25:00: Smarter Software program and a True Flash Tremendous‑Cycle
- KIOXIA isn’t simply transport drives, it’s investing in software program (e.g., ISAAC) to cut back HBM/DRAM wants whereas preserving AI efficiency.
- The technique is holistic: enhance price and effectivity on the system stage, not simply via sooner NAND.
- New fabs like Kitakami (K2) are ramping up, however distributors are cautious after Covid-era boom-and-bust cycles that price the business ~$30B.
- Scaling NAND (extra layers, finer lithography) is getting dearer and faster-paced, so that offer will develop, however in a managed method.
- Many see this as a multi-year “tremendous‑cycle” pushed by AI and knowledge development, not only a regular 1–2-year flash-upturn.
25:00–Finish: AI In every single place, Flash vs. HDD, and Studying from HPC
- Knowledge retains rising and is never deleted; AI is shifting to wherever that knowledge lives, together with the sting, backups, and secondary tiers.
- Excessive-capacity SSDs can ship ~100 PB per rack, versus a number of racks of HDDs for a similar capability, with much better efficiency and effectivity.
- HDD efficiency doesn’t scale with capability; SSDs ship orders-of-magnitude higher IOPS/GB and throughput, bettering TCO and favoring flash at scale.
- Reveals like OCP and Supercomputing (SC) are the place the long run is seen first: extensive racks, 800V busbars, unique cooling, and energy designs.
- Hyperscaler and HPC improvements are steadily shaping what “regular” enterprise servers and knowledge facilities will seem like, with resiliency and effectivity.
