NVIDIA Grace (ARM Neoverse V2) Architecture
Topology
SLNvidiaGraceCPU
-> SLClock -> SLRegisterFile -> SLALU -> SLInstructionDecoder -> SLIOBus -> SLIOTiming -> SLInterruptController
Register section
The architectural register section is a RISC general-purpose register file: AArch64 register model; many Neoverse-V2 cores with a coherent mesh and large shared cache.
Internal latches and datapath elements are modeled only where their observable timing or behavior is useful.
Control section
The control side is represented as:
- instruction register;
- instruction decode;
- control/timing sequencer;
- interrupt logic;
- clock coordination.
Datapath
registers -> internal bus/latches -> ALU -> register/memory destination
Address formation uses the architectural 48-bit address path.
Bus
The external interface exposes 48 address lines, a 64-bit data path, read/write control, clock/phase signals, and reset/interrupt controls. Each observable memory transaction is modeled as a bus event (see BUS.md).
Design principle
The SLeeLa representation is compositional: a NVIDIA Grace (ARM Neoverse V2) variant replaces only the couplers, I/O, timing, cache or opcode components that differ, rather than duplicating the whole CPU hierarchy.
Sources
- https://www.nvidia.com/en-us/data-center/grace-cpu/
- https://en.wikipedia.org/wiki/ARM_Neoverse