Silicon Motion is taking a portfolio of enterprise SSD controllers to Nvidia's GTC and mapping them onto five tiers of AI storage, without announcing a new part, a performance figure or a date.
The mapping is the whole release. MonTitan SM8466 and SM8366 cover both the KV cache extension tier and the near-GPU tier, SM8308 joins them for compute-optimized storage inside AI nodes, SM8366 and SM8388 handle nearline and warm data with TLC and QLC support, and SM8008 sits under the boot drive. Reference design kits come in E1.S, E1.L, E3.S, E3.L and U.2.
"NAND storage is becoming a strategic layer in AI infrastructure," said Wallace Kou, President and CEO of Silicon Motion. Read that as a positioning statement about where the margin is going. Silicon Motion does not sell drives; by its own description its customers are NAND flash vendors, enterprise storage providers, module makers and OEMs, so every tier it names here is a design win it has to win through somebody else's product.
Kou also set out the specification he wants the tier judged on: "As inference context continues to expand, storage must deliver not only high capacity, but also deterministic latency, sustained bandwidth, and predictable quality-of-service." Determinism and tail latency, rather than peak throughput, is the correct axis for KV cache offload, because a GPU stalled waiting on a straggler read is idle no matter how good the average looks. It is also the axis on which the release supplies nothing measurable.
Count the numbers in the announcement and there are none of the kind that matter. No IOPS figure, no bandwidth in gigabytes per second, no latency in microseconds, no drive writes per day, no capacity point, no price and no availability date. Endurance and capacity appear three times each, always as adjectives. The digits in the body are a booth number, five form factor names and five part numbers. PerformaShape, the patented technology credited with managing latency distribution and workload isolation, is named once and never described.
There is also a naming divergence worth flagging. Silicon Motion's release calls the Nvidia programme "NVIDIA Inference Context Memory Storage (ICMS) initiative" and uses that acronym four times. Nvidia's own announcement of the platform at the same show calls it the "NVIDIA CMX™ context memory storage platform" and the string ICMS does not appear on that page at all. Kioxia's GTC release, describing the same problem of extending the memory hierarchy past GPU memory, writes "NVIDIA’s Context Memory Storage (CMX)". Three documents from the same show, three descriptions of the same tier, and one of them is carrying an acronym the vendor does not use. Nothing in Silicon Motion's release explains the difference.
The boilerplate is doing some work too. Silicon Motion describes itself as the global leader in NAND flash controllers and states that it ships more SSD controllers than any other supplier worldwide. No unit figure, no period and no source accompanies that, so it belongs in the column marked vendor claim rather than in the column marked market share.
Bootnote: the VAST Data executive quoted in Silicon Motion's release is not a random endorsement. Nvidia's own STX announcement lists VAST Data among fifteen storage providers and manufacturing partners building on its modular reference designs, a group that also takes in Dell Technologies, DDN, HPE, IBM, NetApp, Nutanix, Supermicro and WEKA, and that is the guest list this controller pitch is aimed at.
Leave a Reply