Research

The paper-stage side of the lab: designs we argue out in writing before committing hardware and code to them.

• Intel Arc Pro B70
• AMD EPYC
• llama.cpp
• SYCL

Run AI workloads as a small shop and the bills are only half the problem. API spend scales with usage and rate limits cap your throughput, but the part that should bother you is that every request ships your data (and your customers' data) to someone else's datacenter. The standard fix is NVIDIA hardware, which means datacenter cards at datacenter prices, or a consumer-card build that pulls 3,400W and needs its own electrical work.

This paper works out a third option: eight Intel Arc Pro B70s in a purpose-built ASUS 4U GPU server on dual AMD EPYC. 256 GB of VRAM for about $45K at public-reseller pricing, drawing 2,200 to 2,600W sustained. It runs frontier open-weight models (GLM-5.2, DeepSeek-V3, 70B dense) at production quality, and every byte stays in the building. At $2,000 a month of API spend, the hardware pays for itself in about 23 months.

Inside: the full bill of materials, the power and thermal budget, the llama.cpp SYCL software stack, throughput projections, risks, and a rollout plan that validates one card before assembling all eight.

Read the white paper (11 pages) → · Download

• QUIC
• IPsec
• Networking

A concept exploring QUIC as the outer transport for IPsec-encapsulated traffic over high-latency satellite links: connection migration instead of tunnel re-establishment, congestion control tuned for 500 ms+ RTTs, and no head-of-line blocking. Paper-stage thinking, for now; the write-up lands here when it's ready.