A local LLM user weighs a 16GB AMD GPU against a second RTX 3090
kobbalt · reddit · 2026-07-25
A Reddit user asks whether adding a second GPU is a good idea for local LLM work.
Current setup:
- i5-12400
- 64 GB RAM
- RTX 3090
- Fedora KDE
- local serving through LM Studio
- models like Qwen 3.5 27B and Gemma 4 26B
The problem is that the context window is filling up too quickly on larger projects. The user is considering a 16 GB AMD card for extra VRAM because it is Linux-friendly, cheaper than another 3090, and easier to cool and power.
They ask whether mixed AMD + Nvidia support is now good enough in 2026, whether the 3090 should hold the model while the second GPU handles context, and whether adding another Nvidia card would still be the better option.
More from Infra
- Bittensor TAO ecosystem map lays out dozens of AI subnets across compute, agents, and security — 0xSammy · 2026-07-25
- Mooncake v0.3.12 adds SSD-backed KV cache pooling and smarter routing — BanghuaZ · 2026-07-25
- Enterprise user asks which models can cut frontier-token spend on a $1M compute budget — mimic751 · 2026-07-25
- The AI chip startup map has swelled to about $58B in private-paper value — deedydas · 2026-07-25
- SpaceX says Starship V3 is improving reusable orbital heat-shield tiles flight by flight — XFreeze · 2026-07-25
- A $8.5B conversational frontier is exposing the real cost of the AI boom — Some-Technology4413 · 2026-07-25