RequantTESTNET
Requant research

Lossless CPU-to-GPU Streaming Beyond VRAM

Abstract

All 960 output observations matched the CPU-replayed reference. ndzip and XBP1 reduced transferred bytes, while plain pinned transfer had the shortest host service in that run and led the tested SIFT cells.

Study 14 of the Consumer Blackwell programme. See the research methodology for timing and replay boundaries.

Across both tested SIFT reference epochs, neither ndzip nor XBP1 beat the stronger plain pinned-transfer control in any of the 20 matched cells at reuse counts 1, 10 or 100. On a distinct 20 GiB smooth synthetic stream, stock ndzip compressed the transfer to 14.1% of the original bytes and XBP1 to 77.2%. Plain transfer had the shortest measured host service and total route wall in this bounded streaming run.

Real data and controls

The SIFT results cover 400 recorded route cases and 163,680 timed frame fingerprint observations. Both independently replayed epochs compared plain pinned transfers at queue depths one and two against online and precompressed XBP1 and ndzip. One epoch used the ordered-copy fix; the other used the unmodified serial CPU encoder. Two deterministic input cohorts, FP32 and FP64 representations, and five frame budgets vary the workload within the same descriptor dataset. The comparison prices the named setup, validation and reuse costs. See the SIFT accounting and compact replay.

Bounded stream beyond VRAM

One fresh RTX 5060 Ti process ran the three online routes. Each consumed 320 successive 64 MiB frames from the same independently committed smooth FP32 stream. The logical input was 21,474,836,480 bytes per route. The selected board reported 17,103,323,136 physical framebuffer bytes and CUDA exposed 16,657,285,120 bytes. Only two source buffers were pinned at a time. The engine regenerated and checked each pair of source frames, submitted both nonblocking slots, drained both and checked the returned digest tables before reuse. The SHA-256 check covered every 4 KiB fragment of the decoded word sequence.

RouteLogical H2D wire bytesFraction of inputPreparation (s)Host service (s)Route wall (s)
Plain pinned21,474,836,480100.0%118.0493.425122.169
ndzip stock serial3,027,842,17614.1%122.3884.680127.551
XBP116,571,378,46477.2%124.2654.320129.193

Service includes online encoding and validation where applicable, transfer, GPU restore and hashing, synchronization, and host digest check. Preparation is charged separately. Route wall also includes setup, first use, output records and teardown. Setup took 0.145–0.169 s per route and first use 0.051–0.204 s. The owning process wall was 380.657 s. Separate source-buffer requests were 128 MiB per route. Slot requests are retained in the receipt as named allocations. Those figures do not establish physical memory peaks, opaque workspace size, or realized CPU/GPU overlap.

Independent verification matched all 960 fresh digest tables to the CPU-replayed reference bound to the 320 original frame identities. The owning wait returned zero, the monotonic clock bracket contained the observed interval, and the replay reconciled every route's host cost with the enclosing process wall. The result measures one shared-host run of one smooth stream and one queue depth. The size-dependent real descriptor comparison is reported separately above.

Retained failures and reproduction

The earlier whole-cache run was terminated under memory pressure before any route completed. Its worker wait retained the termination signal. The earlier expanded synthetic input mismatch remains a rejected collection. Neither attempt contributes to the timing comparison. The loader qualification initially exposed a lazy CUDA Runtime load; after pinning the actual runtime and NVRTC mappings, the six-observation native plain smoke passed under the canonical device lock.

The streamed receipt records the complete host, wire and named-memory census. The collector, owning wait, and independent verifier are the executed source paths in this release. Fresh reproduction uses the staged native libraries, the retained source-pinned reference manifest and its CPU-replayed witness tables. See the XBP1 format for its bit-preserving representation and raw-frame fallback. The shared-host figures describe the tested native XBP1 and external pinned ndzip variants.

The private source-bound full collection SHA-256 is 152765c08a095bb7d17cec5b9d25b68220e4b1d7ac02346c0036bee215e9778d. The independently verified compact readback SHA-256 is 8c22f584915e828d291043ccf8a0d41c48bfa666ab077eb883ff922030a26317.

Evidence and source revision

Cite this report

Requant. “Lossless CPU-to-GPU Streaming Beyond VRAM.” Technical report, 19 November 2026. Evidence revision d12bcf1c3270f28fe85828725fb2f2c5c790b986. Canonical report.

This is an organizational technical report, not a peer-reviewed publication.