YOU ARE VIEWING ONE ITEM FROM THE AICRIER FEED

Dual MI50 rig spews gibberish

AICrier tracks AI developer news across Product Hunt, GitHub, Hacker News, YouTube, X, arXiv, and more. This page keeps the article you opened front and center while giving you a path into the live feed.

// WHAT AICRIER DOES

7+

TRACKED FEEDS

24/7

SCRAPED FEED

Short summaries, external links, screenshots, relevance scoring, tags, and featured picks for AI builders.

Dual MI50 rig spews gibberish
OPEN LINK ↗
// 70d agoINFRASTRUCTURE

Dual MI50 rig spews gibberish

A LocalLLaMA user says a mixed AMD GPU setup runs fine with one MI50, but adding a second MI50 makes inference spit endless “??????” and then crash on stop. The failure shows up on Ubuntu 22.04 with Vulkan, which points more toward a driver/backend compatibility problem than raw compute limits.

// ANALYSIS

Classic mixed-GPU pain: this looks less like a bad model and more like a brittle inference stack tripping over legacy AMD cards, Vulkan fallback, and device enumeration.

  • Mixing a 7900 XT with dual MI50s can expose backend assumptions about GPU order, memory layout, and supported kernels.
  • Infinite garbage tokens usually smell like a broken runtime path, bad backend compilation, or driver mismatch rather than a model spontaneously failing.
  • An 850W PSU is probably too tight for three hungry AMD cards under load, so power instability could be making an already fragile setup worse.
  • The quickest debug path is to test each MI50 alone, then the two MI50s without the 7900 XT, and only then reintroduce the third GPU or fix ROCm properly.
  • Vulkan can be a useful escape hatch, but for older Vega20-era cards it often behaves like “it boots” is not the same as “it is actually stable.”
// TAGS
amd-radeon-instinct-mi50gpuinferencellmself-hostedvulkan

DISCOVERED

70d ago

2026-03-18

PUBLISHED

70d ago

2026-03-18

RELEVANCE

5/ 10

AUTHOR

Savantskie1