06On-device
One Laptop, Three Engines
The match data could not leave our hands, so the compute came to the data. One Snapdragon X Elite laptop: its CPU decodes a full match in under 10 seconds, and its GPU and NPU are the headroom for local models.
01This laptop
Where every byte stayed
- DeviceSnapdragon X Elite laptopthe only machine the match data ever touched
- ProcessorSnapdragon X EliteX1E78100 · Qualcomm Oryon CPU
- Cores12Oryon CPU cores
- Memory32 GBsystem memory
- GraphicsAdreno X1-85Qualcomm Adreno GPU
- Storage1 TBinternal drive
- SystemWindows 11 ProARM64 · 64-bit
How we read the specs
> 'Win32_ComputerSystem','Win32_Processor','Win32_PhysicalMemory', 'Win32_OperatingSystem','Win32_VideoController','Win32_DiskDrive' | ForEach-Object { Get-CimInstance $_ }
Published: device class, processor and core count, memory, graphics, storage, OS edition and architecture. Left out on purpose: serial numbers, user and host names, and anything else that identifies the device.
02Three engines
CPU, GPU and NPU on one chip
- Oryon CPU · 12 cores on our machine. Runs our decoder today.
- Adreno GPU · X1-85 on our machine. Next: vision models.
- Hexagon NPU · up to 45 TOPS. Next: local LLMs.
- Shared memory · one pool for all three engines.
03Today
The CPU decodes a match
- < 10 sfull-match decode, numpy only
- 1 of 3engines used: the Oryon CPU
- arm64native Linux container, offline
- 0bytes sent off the machine
Our Initial Challenge solution ran on the Oryon CPU alone. The Adreno GPU and the Hexagon NPU sat idle.
04Next
Headroom for local models
-
Vision on device
A light ball detector for the frames the tracker misses, running on the GPU or NPU, beside the data. It is the on-device vision engine on our roadmap.
-
Local LLMs
Language models that turn possession numbers into tactical text, run locally with nothing sent anywhere.
The Hexagon NPU in our laptop is rated at up to 45 TOPS: trillions of operations every second, set aside for neural networks alone. It sat idle through the Initial Challenge. That idle silicon is our headroom.
What we vouch for is what we ran ourselves: a full match decoded in under 10 seconds, on the CPU alone, with the network switched off.
Why hardware
Privacy is a place
The match data may not go to any cloud platform, external service or hosted model, so renting a GPU or calling a hosted API was never an option. A laptop strong enough to do the work means nothing has to leave it. Privacy and speed come from the same choice: computing locally.