Home /Modules/cpu/SPECIFICATION.md

SPECIFICATION.md

lib/cpu/nvidia-grace/SPECIFICATION.md · 75 line(s) · markdown

NVIDIA Grace (ARM Neoverse V2) Specification

1. Identity

Model: NVIDIA Grace (ARM Neoverse V2) Architecture class: 64-bit server CPU (ARM Neoverse) Data path: 64-bit Addressing: 48-bit external address space (256 TiB) Register model: RISC general-purpose register file

The profile exposes a programmer model while using the shared SLeeLa CPU objects (clock, register file, ALU, decoder, bus) to sequence each instruction.

2. Programmer-visible state

AArch64 register model; many Neoverse-V2 cores with a coherent mesh and large shared cache.

The SLeeLa model stores these in SLRegisterFile and exposes 32 addressable register slots to the runnable model.

3. Address and data path

4. Clock

Rated model clock: about 3400 MHz (overclockable in the model via setSpecClock).

The model records nominal frequency, cycle number, bus direction, and completion phase. Clock frequency is not instruction latency: one instruction may consume multiple cycles.

5. Execution model

The SLeeLa model sequences each instruction as:

  1. instruction fetch;
  2. opcode decode;
  3. operand/address acquisition;
  4. ALU/register operation;
  5. memory read/write;
  6. next-instruction transition.

6. Cache

cache coherence protocol: modeled where multi-core

7. DMA

The base CPU has no integrated DMA controller; DMA is a system-level bus-master concern and SLDMAController is left uninstantiated for the base profile.

8. Interrupts

The model records interrupt request, recognition, vector/entry and timing through SLInterruptController without inventing undocumented internal behavior.

9. MMU/TLB

The processor integrates an MMU/TLB; address translation is modeled as virtual -> physical through that unit.

10. Evidence labels

Sources