Reply to thread

Message

[QUOTE="Queen, post: 85653, member: 27"]
SOCAMM memory is breaking out of NVIDIA orbit as AMD and Qualcomm chase fixes for AI memory choke points.

SOCAMM stops being NVIDIA-only
[LIST]
[*]SOCAMM starts life as a NVIDIA-focused memory play.
[*]AMD and Qualcomm start circling the same tech.
[*]Agentic AI workloads push memory limits hard.
[/LIST]
What SOCAMM actually is
[LIST]
[*]The design pulls from LPDDR-style DRAM.
[*]The modules stay removable, not soldered down.
[*]It complements HBM rather than replacing it.
[/LIST]
Why AI systems care
[LIST]
[*]Agentic AI needs fast short-term memory.
[*]SOCAMM allows massive per-CPU capacity.
[*]Millions of active tokens stay resident.
[/LIST]
Different design path from NVIDIA
[LIST]
[*]AMD and Qualcomm test square module layouts.
[*]DRAM chips sit in dual-row arrangements.
[*]The goal is cleaner power control.
[/LIST]
Power management changes
[LIST]
[*]PMIC shifts directly onto the SOCAMM module.
[*]Boards lose some power circuitry complexity.
[*]Stable high-speed operation becomes easier.
[/LIST]
Performance and efficiency tradeoffs
[LIST]
[*]Bandwidth trails behind HBM.
[*]Power draw stays relatively low.
[*]The balance still works for memory-heavy tasks.
[/LIST]
What comes next
[LIST]
[*]NVIDIA plans SOCAMM 2 with Vera Rubin clusters.
[*]AMD and Qualcomm evaluate future AI racks.
[*]Wider adoption looks likely if tests hold.
[/LIST]
[/QUOTE]

Name