Exemplo (inglês) · Esta página técnica está disponível em inglês no momento.
OpenHBM + HBM4
Avalotl Omni uses OpenHBM as the open research integration path for an HBM4-class memory architecture target.
Memory target
Memory architecture
HBM4 target
Integration path
OpenHBM
Use
High-bandwidth model weights, activations and tensor data movement
Status
Research and architecture target; not a fabricated HBM4 subsystem
Why it matters
Accelerator throughput depends on moving operands fast enough to the compute fabric. HBM4-class bandwidth is therefore treated as a first-class architectural constraint alongside matrix compute.
Hierarchy
The memory path is designed around local register files and SRAM near the tensor engines, chiplet-level SRAM, system cache, and the OpenHBM/HBM4 external high-bandwidth memory interface.
Design targetPublished HBM4 references here describe the Avalotl Omni architecture target, not benchmark results from manufactured silicon.