Observed arrival · 2026-08-21
Wahid Semiconductor’s 3D Near-Memory LPU
Wahid Semiconductor proposes stacking DRAM vertically on a compute unit to accelerate the memory-bound decode phase of large-model inference.
Why it surfaced
This is an unusually candid semiconductor thesis: the page distinguishes shipped silicon, established facts, and unbuilt design targets, while openly stating that no Enlight silicon exists yet. Its proposed 10–40 TB/s bandwidth, DDR interface, KV-cache compression, and planned April 2027 tape-out give readers concrete engineering claims to interrogate rather than another vague AI accelerator pitch.
Wahid Semiconductor proposes stacking DRAM vertically on a compute unit to accelerate the memory-bound decode phase of large-model inference.
Observed signals
Read the marks
Editorial observations of this landing page, not a rating.
One card from the complete issue