Amplifying PCIe Bandwidth 6x with NVENC

A PoC that compresses activations using NVENC on the RTX 5090 before passing them over PCIe was measured against real workloads. The README claims this approach can effectively amplify GPU-to-GPU transfer on 4090/5090 setups without NVLink by the compression ratio.

• In a capture of 1,735 diffusion activations, it recorded 6.1× lossless compression and cosine 0.991 LOO.
• MultiEngineDirectBackend showed 0.180 ms/frame encode and 0.262 ms/frame decode on 668 frames of real FLUX activations.
• Running poc/17's 30×4096² fp16 GEMM concurrently with 64-frame encode achieved a parallel wall-clock of 26.0 ms versus a serialized 40.1 ms, a 1.34× speedup, reaching 67% of the theoretical maximum overlap.

The README suggests this combination can reach an effective bandwidth of ~180 GB/s, but cross-GPU peer-to-peer wiring still remains as the next integration step.