PoCX Plotter v2: Ring Buffer Pipeline

Dev Diary #1 — Interactive visualization of the fused scatter+compress architecture
All numeric values shown are examples for an RTX 2000 Ada. Your GPU will auto-tune these at startup.

The Pipeline

The GPU ring buffer holds hashed nonces as a continuous fill — not aligned to warp boundaries. The hash kernel adds ~1.4 warps per dispatch, while compress drains 2 warps at a time. The ring capacity (~3.25 warps) is just enough that the two kernels never collide.

Compress pulls a pair (A + B), transposes B, and XORs them into the GPU scatter buffer (1 GiB). For X2/X4, multiple pairs are XOR-accumulated before the result transfers to the host write buffer. The host never sees raw nonce data — no plot buffer needed.

Ring Buffer Pipeline

Select a compression mode and hit Play.

Speed:

Empty

Ring fill

Warp A (plain)

Warp B (transpose)

XOR result

Writing

Memory Comparison

v2 GPU memory depends on your card's compute units and kernel workgroup size — the plotter auto-tunes at startup.

Scale (-x): Escalate (-e): Disks: v2 mode: Double buffer (v2 only)

GPU cores: KWS:

	Plotter v1	Plotter v2
Host plot buffer
Host write buffer
Host GPU buffer
Total host memory
GPU memory