Skip to content
Friendly disclaimer: flozi00 TechHub is a solo side-project next to a full-time job — personal learning notes, no official statements. Verify critical steps yourself.
Knowledge Base

flozi00TechHub

Deep dives into servers, GPUs, AI infrastructure and modern IT systems — practical notes from real deployments.

Browse Documentation

Navigate structured guides on hardware, AI infrastructure, Linux and networking.

Find Anything Fast

Full-text search across every guide — specs, configurations and benchmarks included.

in the header — just start typing

Explore by Topic

Guides(54)

Latest Articles

6 min read

How we serve GLM on B300 in production: MTP k=5 tuning

Production notes from serving a GLM-class MoE on 8x B300: the acceptance-rate math behind speculative decoding, why the measured log_stats chain says k=2-k=3 captures most of k=5, and why NVLS failures are silent by default.

llm-inferencespeculative-decodingmtpb300
12 min read

Agent-Fleet Token Economics Is Cache Economics: Decomposing the 60x Dollar Spread

The self-reported day of 44 agents, 4 billion tokens and $1,300 versus $80,000 — recomputed line by line against live price sheets, decomposed into cache discount, model spread and looping fraction, and rebuilt as a capacity-planning checklist for anyone running an internal agent fleet.

aimachine-learningllminference
15 min read

BOOST and the End of Prefetch: Why Grace Hopper Wants Both Memory Tiers at Once

BOOST (arXiv:2609.13592) decomposed to the bandwidth ledger: why prefetch tiering structurally burns HBM writes during decode, why concurrent proportional access adds the host tier instead of stealing from it, the alpha-math behind +31% throughput and 4.3% TPOT at iso-batch, and the NVLink-C2C scope guard that keeps PCIe x86 hosts out of the claim.

aimachine-learninggpugpu-memory
10 min read

The Cache-Read Price War of September 2026 That Wasn't Three-Sided

The press told a story of three frontier labs cutting prices in unison on September 21-22, 2026. The price sheets say otherwise: one vendor cut cache reads to 0.05x, one shipped new SKUs with an untouched cache multiplier, and one did nothing at all. Decomposed to arithmetic, with per-vendor Barrie-shaped day costings in Python.

aimachine-learningllminference
20 min read

CXL-SSDs For LLM Prefix Caching: Byte-Addressability Buys Nothing Without Chunk Awareness

A September 2026 Sogang University paper (arXiv:2609.26828) asks whether CXL-SSDs can replace NVMe for LLM prefix caching and starts with a resounding no: a stock CXL-SSD is about 3× slower than local DRAM and no faster than an NVMe SSD, and a generic next-n prefetcher leaves TTFT essentially unchanged while burning NAND bandwidth. The interesting half is the constructive one — LM-CXD, a co-designed CXL-SSD that makes KV chunks device-visible I/O units, exposes NAND-to-DRAM progress to the serving engine, uses device DRAM as a GPU-accessible buffer, and pipelines layerwise KV movement under GPU compute, reaching up to 2.6× (compute-asynchronous) and 4.03× (layerwise) lower TTFT than the stock device, within 1.5× of local DRAM. This guide prices every step of the interface stack in bandwidth-and-latency arithmetic, works through why CXL's memory semantics do not make loads free, why prefetching without engine knowledge must speculate and lose, and what remains unverified because no CXL-SSD is commercially available yet.

aimachine-learningllminference