ScaleFlux develops controllers to tackle AI memory bottleneck
ScaleFlux says its new controller chips for PCIe Gen 6 solid‑state drives and Compute Express Link memory aim to keep up with fast GPUs and TPUs, which are creating a memory bottleneck in AI data centers.
15 Oct 2026 0 reads Compiled with AI
#ScaleFlux #GPU #TPU #PCIe #Compute Express Link
AI data centers have poured money into faster accelerators, denser racks and liquid cooling, but the data pipeline that feeds the processors has not kept up. As a result, expensive compute often waits for memory and storage. The gap between fast GPUs and the slower data feed is becoming a central planning issue for data centers.
ScaleFlux co‑founder Hao Zhong says GPU and TPU performance has advanced quickly—over a thousand times in eight years—but the data pipeline must keep pace. If it does not, the efficiency gains from the processors will be limited. Zhong notes that the shift from training to inference and agentic workloads is increasing the need for data close at hand.
To address this, ScaleFlux unveiled two new controller chips this year. One is designed for PCIe Gen 6 solid‑state drives, the other for Compute Express Link memory. These chips act as small processors that manage the external DRAM and solid‑state drive stages of the pipeline.
Zhong describes AI infrastructure as a multi‑stage pipeline: the first stage is processors with high‑bandwidth memory, followed by external DRAM and solid‑state drives. Making the second and third stages faster and more efficient is key to keeping AI workloads running smoothly. The new controllers aim to reduce the memory bottleneck that threatens future AI performance.