Systems engineer - GPU inference systems, kernel/virtualization (ex-IBM Linux)
Focus: Inference, memory management, accelerated compute, quantization
Notes and consulting: lineargp.com
Selected Work
- 2026-10 Kern 808F el Bypass over PCIe: Doorbells & DMA . analysis . repo
- 2026-10 OpenBMC and Redfish on GB200 . analysis
- 2026-08 SACC: Synthetic PCIe Accelerator . analysis . repo
- 2026-07 KV Wire: RDMA over RoCE for Disaggregated Inference . analysis . repo
- 2026-07 GPU Passthrough: IOMMU, VFIO, and Thermal Debug . analysis . repo
- 2026-06 DiffusionBlocks AR: first language-model port, block-wise training, GPU decoupling . repo
- 2026-05 Roofline & Ridge Point: GPU and inference optimization . repo
- 2026-05 KV Mix: asymmetric K/V quantization, outlier-driven mixed precision . repo
- 2026-04 Deepseek V4: system design & implications . analysis
- 2026-04 Rootless Agent Sandbox: kernel primitives composition (ns, Landlock, seccomp) . repo
- 2026-03 KV Compression (TurboQuant): memory + demand implications . analysis
- 2026-03 Paged Attention: KV cache as virtual memory (CUDA) . repo
- 2026-02 Decode Attention: Multi-Query, Multi-head, Group Query . repo
- 2026-01 Kimi K2.5: methods + systems impact . analysis
- 2026-01 FP64 on Blackwell INT8 (Ozaki): POC . analysis . repo
- 2026-01 KV Offload (Engram): system design . analysis
Linux Kernel: Author of rpadlpar (mainline PCI Hotplug) and librtas (POWER userspace).
Also: quant-sampler (sampling in C, stable softmax, bit-packing) . rand-bitwise (uniform u32 RNG using Lemire's algorithm)
Background: 7 years Linux kernel at IBM. 15 years industry. Now building low-level systems/GPU/inference.
Contact: john@lineargp.com | LinkedIn