Parallel Paths to High-Bandwidth Memory for ML/AI: Specific Purpose Me...
summary
UEFI tags HBM as EFI_MEMORY_SP (UEFI 2.8), the kernel maps that to E820_TYPE_SOFT_RESERVED and calls memblock_reserve so Linux’s DAX/HMEM auto‑claims the regions; NVIDIA instead calls add_memory_driver_managed() to hot‑plug HBM as a NUMA node (IORESOURCE_SYSRAM_DRIVER_MANAGED) so malloc/numa_alloc_onnode target it directly. Current problem: DAX steals SPM, breaking GPU IO mappings; temporary workaround is efi=nosoftreserve. Proposed fixes include ACPI device ownership, a DAX claim API or CDAT/CXL‑style tables and backward compatibility for older kernels.
AI-written summary · about
chapters (4)
- 0:00Introduction and problem overview
- 7:53Proposed solutions for SPM
- 18:13Kernel integration challenges
- 21:03Q&A and wrap-up
related talks
02mSession: ISMM - ISMM: Keynote 2 - Dr. Euicheol Lim
58mKeynote: Enabling the AI Revolution: Next-Generation Memory Solutions for AI Computing(…)
21mRemoving Obstacles before Breaking Through the Memory Wall: A Close Look at HBM...
30mGPU Sharing at CERN: Cutting the Cake Without Losing a Slice!
05mHami: Dynamic, Smart, Stable GPU-Sharing Middleware In Kubern...
01h 05mReal In-Memory Processing
56mReduce, Reuse, Recycle: Is this Mantra under attack with Proliferation of Machine(…)
28mSimplifying AI for Edge Compute with HAMi