Register-Transfer Level (RTL) Learning Lab / SystemVerilog

Tiny-GPU Architecture Explorer

Understand the hardware mental model: kernel → blocks → cores → threads → memory. Click any hardware block to discover what it does and where it lives in the RTL.

↗ Open source repository
2Compute cores (default)
4Thread lanes per core
16-bitInstruction width
8-bitData path width

One kernel, multiple cores

The control path launches work; memory controllers service requests.

click a block to inspect
ControlComputeMemoryState / dataSimplified signal-flow — not a physical floorplan

🧠 The right mental model

01Dispatcher operates at block granularity. A core processes one block at a time, not a continuously scheduled pool of warps.
02Fetcher / decoder / scheduler are shared per core. All active lanes follow the same decoded instruction.
03ALU / LSU / register file / PC logic are per thread lane. Same instruction, different operands and results.
04Memory channels are shared. A request can finish after variable latency; the core waits before moving on.

⚠️ What this tiny GPU deliberately omits

No pipelined instruction overlap. FETCH → DECODE → REQUEST → WAIT → EXECUTE → UPDATE is a sequential control-state machine.

No warp scheduler or scoreboard. The simple core doesn't need out-of-order hazard tracking because it finishes the current instruction before fetching the next.

No branch divergence handling. The scheduler selects next_pc[THREADS_PER_BLOCK-1] as the shared next PC.

No RTL cache in the current src/ list. README discusses caching as a learning direction; the connected design uses memory controllers directly.

Quick reference

Full names for every component & abbreviation

Read a short label in the diagram, then find its complete name here. Click any diagram box to inspect its RTL and signals.

Name guide
GPUGraphics Processing Unit
DCRDevice Control Register
FSMFinite-State Machine
RFRegister File
ALUArithmetic Logic Unit
LSULoad/Store Unit
PCProgram Counter
NZPNegative / Zero / Positive condition flags
RTLRegister-Transfer Level
ISAInstruction Set Architecture
SIMDSingle Instruction, Multiple Data
CHMemory Channel
LDRLoad Register (load from data memory)
STRStore Register (store to data memory)
CMPCompare
ADDAddition
rd / rs / rtDestination / first source / second source register fields
blockIdx / blockDim / threadIdxBlock index / threads per block / local thread index