d-Matrix CEO on why 3D-stacked memory compute could break the AI memory wall
firstadopter · x · 2026-10-19
Tae Kim interviews d-Matrix founder and CEO Sid Sheth on tackling the AI industry's "memory wall" — the widening gap between compute speed and data feed rate — with 3D-DRAM.
- Founded in 2019, d-Matrix is reportedly ahead in 3D-DRAM, shipping it in next-gen Raptor products slated for end of next year
- Recent moves: Squadron multi-rack inference cluster with Arista, Broadcom, and Supermicro announced at OCP; Raptor XPUs will adopt Nvidia NVLink Fusion, with MGX racks expected Q4 2027
- The conversation covers the company's early pivots and why inference needs memory-centric compute
- Caveat: execution remains the test, but success could reshape the industry.
More from Infra
- Anthropic Built Scion, Its Own Source Control System, as Merged Changes Grew 12x to 30,000/Day — Pavan_Belagatti · 2026-10-19
- Dwarfstar on 128GB M5 Max Mac Studio hits 16 tok/s with SSD offload via DeepSeek Flash 4.1 — antirez · 2026-10-19
- HBM shake-up: SK Hynix HBM4 packaging delays reported, prices seen up 30-45% in 2027 — tengyanAI · 2026-10-19
- Researcher warns of ~1,000 cloud inference endpoints potentially compromised in active risk — matthew_d_green · 2026-10-19
- AWS open-sources TorchNeuron, a native PyTorch backend that makes Trainium a standard PyTorch device — PyTorch · 2026-10-19
- Container machines: persistent Linux dev VMs unique to Apple's container runtime — AIFlow_ML · 2026-10-19