GGayatri Panse·16h agoBackprop 01: Where Did the Memory Go?Why the same Gemma 4 12B fit 1.83 users on vLLM and about 13 on llama.cpp24 min read·gradientdecent.hashnode.dev
CCyber Knight·2d agoMior and Skyfall: The Attacks That Broke AMD's "Even the Host Can't Look Inside" PromiseTwo research teams showed a malicious cloud host can read a confidential VM's memory — and forge the signed report that says everything is fine.5 min read·nulltrace.hashnode.dev
GGayatri Panse·Oct 31Epoch 2: It Serves. Does It Train?Fine-tuning Gemma 4 with Unsloth on two AMD Radeon RX 7900 XTX cards, and how much of the pain was actually AMD's Slug34 min read·gradientdecent.hashnode.dev2
SShraddha bhat·Oct 17The $8.2 Billion Bet on the Third Dimension: How AMD Just Changed the AI Race ForeverThe AI Industry Has Been Missing a Dimension. AMD Just Paid $8.2 Billion to Add It.9 min read·aiinsights1909.hashnode.dev
SServers99·Oct 16KNOD Explained: Offloading Linux Network Traffic to AMD GPUsHow the In-Kernel Network Offload Device bypasses CPU bottlenecks and processes up to 70 million packets per second without userspace overhead.8 min read·servers99.hashnode.dev
GGayatri Panse·Oct 13Epoch 1: Try It, Time It, Then Buy ItHow three rounds of Gemma 4 experiments shaped our choice of GPU, inference engine and model23 min read·gradientdecent.hashnode.dev1
MMIG servers·Oct 8AMD EPYC 9006 "Venice" CPUs: Architecting the Agentic AI StackHow AMD EPYC 9006 Processors Power Scalable, High-Performance Infrastructure for Agentic AI Workloads5 min read·migservers.hashnode.dev
TThrottle·Oct 6The FP8 trap: my GPU bill dropped 47% because the model was printing "!!!!!!"FP8 on an AMD MI300X looked 41-47% cheaper per token. Then I read the answers. What really happened, and the fix that's a real 32% saving.4 min read·throttle.hashnode.dev
VVedant Jadhav·Oct 3What 57 LLM Benchmark Runs on an AMD Instinct MI300X Taught Me About VRAM, Throughput, and QuantizationI built AMD LLM Lab to understand LLM inference performance before running the model.7 min read·vedantjadhav.hashnode.dev13A
JJosh Green·Aug 1A 30B Model Crawled on My Mini PC, and the Weights Were InnocentA mixture-of-experts model fell off a cliff on an AMD mini PC. The fix was a different runtime, not a different model.5 min read·joshgreen.hashnode.dev21A