{
  "id": 20465414,
  "title": "I paired two cheap 16GB GPUs instead of buying one expensive 24GB card, and the math completely changed how I run local models",
  "url": "https://urgent.news/2026/11/17/i-paired-two-cheap-16gb-gpus-instead-of-buying-one-expensive-24gb",
  "topic": "tech",
  "section": "Tech",
  "published": "2026-11-17T17:00:24.000Z",
  "source": {
    "name": "XDA Developers",
    "slug": "xda-developers",
    "url": "https://www.xda-developers.com/two-cheap-16gb-graphics-cards-run-bigger-local-models-than-one-expensive-24gb-card/"
  },
  "original_language": "en",
  "account": "For those who want to run local LLMs, buying a single expensive graphics card is not always the most cost-effective solution. Two relatively cheap 16GB GPUs offer a cheaper alternative to a single 32GB card. The writer built a test bench using two Asus Prime GeForce RTX 5060 Ti 16GB cards for $1,680, and compared the performance against an RTX 5090 32GB card. The results showed that the pair of GPUs provided a good compromise between cost and performance for running local AI models.\n\nThe writer tested three models in llama.cpp: Qwen3-14B and Qwen3-32B, and Llama 3.3 70B Instruct. All models were tested using the same prompts, settings, and llama.cpp build. The results showed that a single RTX 5060 Ti could not handle the 14B, 32B, or 70B models due to memory limitations. However, the pair of GPUs could handle all three models, albeit with varying performance. The 14B ran at 46.37 tokens per second on the pair, compared to 46.31 tokens per second on a single RTX 5060 Ti. The 32B and 70B models showed similar improvements in performance when running on the pair of GPUs.\n\nOne challenge the writer encountered was tensor split, which divides each layer between both cards so they work at the same time. This method proved to be 1.7 to 1.8 times faster than layer split, and used about the same memory. However, the writer initially encountered a CUDA error when using tensor split, which was resolved by switching to a CUDA 13 image for the CUDA libraries used by llama.cpp.\n\nThe writer also noted that the performance of the pair of GPUs was not as great as the flagship RTX 5090 card. The 5090 had 1,792GB/s of memory bandwidth compared to 448GB/s for each RTX 5060 Ti. This meant that the 5090 could read the model's weights faster, leading to improved performance. However, the cost of the 5090 was significantly higher than the pair of RTX 5060 Ti cards, making the pair a more cost-effective solution.",
  "summary": "Sometimes cheaper wins",
  "key_points": [],
  "editors_take": null,
  "illustration": null,
  "coverage": {
    "outlets": 1,
    "also_reported_by": []
  },
  "ai_generated": true,
  "disclaimer": "Summaries, key points and the editor’s take are written by software from other outlets’ reporting and may contain errors — always check the linked original."
}