> Source: developer community teardown video transcript (timestamped) | Follow-up in our on-device deployment series | Six-way verification against public sources
Who is Thor-U: the same ADAS chip as Li Auto's L series
TechInsights' cockpit and ADAS SoC analysis records that Li Auto adopted NVIDIA Thor U starting May 2025 on some L-series models, at roughly 730 TOPS (TechInsights' figure, precision unspecified). The 1,800 RMB 64GB teardown board is thus a part that flowed out of an automotive-grade ADAS domain controller—same chip lineage as shipping vehicles.
That explains the price. The Jetson AGX Thor T5000 devkit costs USD 3,499 (128GB LPDDR5X, 273GB/s, 2070 sparse FP4 TFLOPS); DGX Spark costs USD 4,699 (128GB unified memory, 273GB/s, 1000 TOPS FP4). A teardown part brings a platform that would cost tens of thousands of RMB down to 1,800—automaker procurement volume paid the scale premium once on developers' behalf.
The "half a Spark" reality: one silicon, two cuts
"Half" is not rhetoric. GB10's device name in CUDA is literally NVIDIA Thor—mmapeak's sample output shows compute capability 11.0, 20 SMs, 2,560 CUDA cores, and the T5000's GPU is also a 2,560-core Blackwell. Desktop Spark and automotive Thor are two packages of the same silicon with identical compute capability, so the CUDA ecosystem (llama.cpp, PyTorch, ComfyUI) runs directly on the teardown board.
The two cuts: memory halved (64GB vs 128GB), and compute reduced to ~730 TOPS vs GB10's 1000 TOPS (~70%). The video's measured result—27B decode speed at ~70% of Spark—matches exactly: decode is bandwidth-bound, so the memory subsystem is the main constraint; MTP spreads bandwidth cost across compute by verifying multiple tokens at once. The long-context prefill gap is wider, consistent with prefill being compute-bound (730 vs 1000 TOPS) plus an immature automotive software stack.
Two puzzle pieces on the same day
Together with this morning's TensorFold post on reproducible speed, the pieces line up: software makes correctness a contract, hardware prices compute at 1,800 RMB. Qwen3.8-27B appears again, adding another deployment-collapse data point.
Updated price-anchor spectrum: gaming PCs on a 284B model (FreeToken), M5 Ultra bandwidth math, Jetson Thor FP8 on-device at 24.3Hz (APXInf), and now an automotive teardown board at 1,800 RMB.
Teardown board economics: paying weekends for the savings
The transcript's most honest line: "The board is cheap, but power, cooling, and the software environment are on you—the savings come out of your weekends." No warranty, no complete BSP, no ecosystem; DRIVE OS's automotive base must be hand-adapted to run inference frameworks. The video author contributed mmapeak support for the board—buyers writing their own benchmark tools is standard open-source hardware practice.
Two falsifiable predictions:
1. Within 12 months, a third-party accessory ecosystem (power baseboards, cooling shells, flashing guides) will form, turning "Thor-U on the desk" from tinkering into tutorials. 2. Mainstream inference stacks like llama.cpp will add official support for automotive Thor SKUs, stripping the software-stack portion out of the prefill gap—what remains will be pure hardware difference.
---
*Materials: mmapeak repository (Z841973620/mmapeak, CUDA MMA microbenchmark, sm_110, sample output from GB10); TechInsights ADAS SoC analysis (Thor U ≈ 730 TOPS, Li Auto adoption 2025-05); NVIDIA official and three reseller spec pages for T5000/Spark; six-way search cross-check. The video itself could not be traced (uploader unconfirmed); measured figures are per the video; the precision of the 730 TOPS figure is unstated in public sources.*