AMEVA-Forge Documentation

Release 2.0.0 Initializing WebGPU... [ Live WebGPU Studio ] [ GitHub Repository ]

6. FlashAttention-2 vs Standard Attention Profiler

Profiles on-chip SRAM tiling and evaluates real-time VRAM scaling from O(N^2) to O(N) across varying sequence lengths.

Device Target: WebGPU
VRAM Reduction: 0.0%
Speedup Factor: 1.0x
Flash Latency: 0.0 ms

Transformer Attention Parameters

SRAM Tiling Formula:
Q_block @ K_block.T -> Online Softmax -> @ V_block
Zero intermediate $N imes N$ matrix materialized in global VRAM.

VRAM Memory Materialization (Bytes)

Standard Attention:
0 MB
FlashAttention-2:
0 MB

WebGPU Kernel Dispatch Latency (ms)

Standard Attention:
0.0 ms
FlashAttention-2:
0.0 ms
Select parameters and click Profile to execute live attention comparison.

Technical & Mathematical Deep-Dive