09_runtime.md

Runtime and Launch

The phoenix_runtime.py acts as the simulated PCIe Host Driver for the GPU.

The Principle of Data Locality

In a distributed chiplet architecture, moving data is extremely expensive. Phoenix-XM operates on the principle of Move Compute, Not Data.

Rather than transferring terabytes of matrix data across the inter-tile fabric to a specific SM, the Runtime and Global Scheduler allocate the data into physical memory banks located on a specific Tile. The Global Scheduler then guarantees that the compute blocks assigned to operate on that data are launched on the same Tile.

Kernel Launch Sequence

1. Compilation: The Python Assembler compiles the RISC-V assembly into machine code.
2. Memory Allocation: The Runtime writes the machine code into the GPU's Instruction Memory via the simulated host pins (host_imem_wr_en).
3. Data Transfer: The Runtime writes initial vectors/matrices into the GPU's Data Memory via (host_dmem_wr_en).
4. Launch: The Runtime pulses host_kernel_valid and passes the Program Counter, Block Count, and Threads Per Block.
5. Wait: The host waits until the hardware asserts host_kernel_done.
6. Retrieval: The Runtime reads the final computed results back out of GPU memory.