Cross-Domain Co-Design: Decode-to-Display-to-Inference

Thesis

Where do the boundaries between video decode, display rendering, and AI inference align such that the same silicon serves all three workloads? The textbook answer is that they do not. In practice, modern consumer SoCs ship with separate decoder cores, dedicated display compositors, and NN accelerators sitting on limited shared substrate. Each domain carries its own clock tree, its own memory access pattern, and its own scheduling discipline. The integration cost is paid in DRAM bandwidth, in cross-block synchronization headers, and in software glue that exists only because the hardware was specified one block at a time.

TsugiCinema's filed work spans those three domains. Trinity (US Prov. 63/987,139) sits on the decode side, with a hardware sideband that phase-locks plesiochronous discrete decoder ICs at scanline granularity. DLC (US Prov. 64/054,446) sits on the compression side, with a pre-encode pixel-domain transform that splits a high-precision master into two non-viewable layers and orchestrates two concurrent hardware decoder contexts on a single consumer-class SoC. Infinity (US Prov. 64/055,093) lifts the same plesiochronous architecture to chassis-scale gradient consensus, with hardware elastic buffers at gradient-tensor granularity and a hysteretic phase-correction sideband on the training substrate. K-Pool LoRA (US Prov. 64/060,315) is the deliberate software analog at the inference-time adaptation layer, with K-snapshot retention, sign-quantized active update, and gradient-detached regime routing mapped from Infinity's apparatus claim. The cross-domain co-design research line asks whether the substrate underneath those four filings can be unified at the system-architecture layer, and what falls out if it can. Specific claims and architectures from this line are forward-looking. No patent has been filed under this work-item.

Why this matters

• The cross-domain thesis is what makes the four filed surfaces (Trinity, DLC, Infinity, K-Pool LoRA) more than the sum of their parts. The portfolio is assessed as worth more bundled than as single assets, because the value of any one filing depends on the surfaces it reinforces across the other three.
• The cross-domain story is the portfolio's strategic anchor. The unbundling thesis (decode-only, training-only, inference-only) is structurally weaker than the bundled thesis on the same IP because the apparatus-to-software mapping between Infinity and K-Pool LoRA, and the scale-shifted plesiochronous architecture between Trinity and Infinity, are load-bearing once the four are read together.
• Status: Research. The work is forward-looking and reflects internal architecture conviction, not filed claims. The substantive weight rests on the four filed surfaces this line aggregates rather than on the line itself.

Status and what is next

Active research. The cross-domain story is the lens that informs filing strategy across the portfolio, and is the framing applied when adjacent surfaces (memory pooling, sideband protocol generalization, multi-vendor accelerator pools) are evaluated for prior-art clearance and claim scope. Specific claims and architectures from this line are not in any current filing. Honest disclosure: this is the most speculative work-item on the labs site. The evidentiary anchor is the four filed provisionals it aggregates, plus the apparatus-to-software mapping documented in the K-Pool LoRA specification's Section 2.5. The cross-domain line itself contributes integrative posture and filing-strategy discipline; it does not contribute claim scope.