Linear Group systems research

Inference, systems, and chips

Research notes on low-level systems and GPUs: inference, memory, and the hardware underneath.

John Rose — ex-IBM Linux kernel, author of rpadlpar (PCI hotplug, POWER) · github.com/jnros · john@lineargp.com

Consulting

Scoped engagements: KV cache and memory optimization, quantization, multi-tenant GPU isolation, driver-level bring-up and performance.

Research notes

2026-08 SACC: A Synthetic PCIe Accelerator QEMU device model plus Linux driver, written from a spec. Kernel-bypass submission, doorbells and BARs, arbitrary PCIe topology. writeup repo
2026-07 RDMA for Disaggregated Inference RoCE implementation; KV cache transport and tradeoffs. writeup repo
2026-07 GPU passthrough: VFIO, IOMMU, and Thermal Debug NVIDIA passthrough to a KVM guest on AMD. ~30 °C thermal spike from GPU driver contention (resolved). Build steps and mitigations. writeup
2026-06 DiffusionBlocks AR: first language-model port Block-wise training with GPU decoupling. repo
2026-05 Roofline & ridge point: GPU inference optimization Decode, roofline, and ridge-point benchmarks; H100 characterization found <30% of peak memory bandwidth utilized. repo
2026-05 KV Mix: asymmetric K/V quantization Outlier-driven mixed precision; K and V tolerate different bit-widths. repo
2026-04 DeepSeek V4: system design & implications Analysis of the methods and their systems impact. gist
2026-04 Rootless agent sandbox from kernel primitives Isolation composed from namespaces, Landlock, and seccomp — no root, no containers. repo
2026-03 KV compression (TurboQuant): memory & demand implications What KV compression does to serving capacity per GPU. gist
2026-03 Paged attention: KV cache as virtual memory (CUDA) From-scratch CUDA implementation treating KV cache as paged virtual memory. repo
2026-02 Decode attention: MHA / GQA / MQA Benchmarked attention variants across memory- and compute-bound regimes. repo
2026-01 Kimi K2.5: methods + systems impact Analysis of training methods and inference-side consequences. gist
2026-01 FP64 on Blackwell INT8 tensor cores (Ozaki scheme): POC Proof of concept for emulated double precision on INT8 hardware. repo gist
2026-01 KV offload (Engram): system design Design for KV cache offload across memory tiers. gist

Background

Seven years in Linux kernel at IBM: architected I/O virtualization for POWER (rpadlpar, still in mainline), created librtas, led the international virtualization team, upstreamed across 12 kernel releases. Fifteen years of technology strategy and capital allocation at AT&T.