<?xml version="1.0" encoding="utf-8" standalone="yes"?><rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom"><channel><title>Rocm on kmarble.dev</title><link>https://kmarble.dev/tags/rocm/</link><description>Recent content in Rocm on kmarble.dev</description><generator>Hugo -- gohugo.io</generator><language>en</language><copyright>© 2026</copyright><lastBuildDate>Fri, 05 Jun 2026 15:55:14 -0500</lastBuildDate><atom:link href="https://kmarble.dev/tags/rocm/index.xml" rel="self" type="application/rss+xml"/><item><title>Gemma 4 QAT Benchmark: Same Quality, Faster Generation, Less VRAM</title><link>https://kmarble.dev/posts/gemma-4-qat-benchmark-same-quality-faster-less-vram/</link><pubDate>Fri, 05 Jun 2026 00:00:00 +0000</pubDate><guid>https://kmarble.dev/posts/gemma-4-qat-benchmark-same-quality-faster-less-vram/</guid><description/></item><item><title>Qwen3.6-35B vs Gemma4-26B: Real Workload Benchmarks on Radeon 7900 XTX</title><link>https://kmarble.dev/posts/qwen36-vs-gemma4-7900xtx-workload-benchmarks/</link><pubDate>Sat, 30 May 2026 00:00:00 +0000</pubDate><guid>https://kmarble.dev/posts/qwen36-vs-gemma4-7900xtx-workload-benchmarks/</guid><description/></item><item><title>Strix Halo at Full Context — Why Your Decode Drops 64% and What Actually Fixes It</title><link>https://kmarble.dev/posts/strix-halo-full-context-decode-drops/</link><pubDate>Sat, 16 May 2026 00:00:00 +0000</pubDate><guid>https://kmarble.dev/posts/strix-halo-full-context-decode-drops/</guid><description/></item></channel></rss>