Month: September 2025

Hey hey, howdy y’all. Please know that I need to update the below numbers as I had to reconfigure my benchmark script. Any video released prior to Oct 1, 2025 is now dated. I will be updating this page entirely very soon.

This is an ongoing experiment to determine the performance of dual gpu’s as it relates to the mighty RTX 5090 running Ollama, both on Windows and on Linux, as required by non-supported gpu’s.

This page will be updated regularly as I continue to test different gpu’s. As of now I’ve been utilizing the AsRock Taichi Lite X870E motherboard, as it’s the most affordable AM5 option that has TWO pcie 16 slots – both directly connected to the CPU. The typical arrangement is to have the 2nd (or 3rd) pcie16 slot passed off to the chipset… thus severely limiting its overall bandwidth. There are a growing number of skus available, ASUS CROSSHAIR, MSI MAG, etc, that offer a similar arrangement, but they are considerably more expensive.

The graph below will show the gpu sku, llm model, results in 1 card vs 2 card setup, and wattage of overall system.

If there is a particular gpu setup that you’d like to see, please reach out to me at david@countryboycomputersbg.com or leave a comment on YouTube.

Most of these tests are on YouTube, check ’em out.

(rotate your phone for better viewing)

nVidia gpu’s

GPU	# of gpu’s	model	token / s	wattage	watt / token
5090	1	Qwen3:32b	42.3	730	17.3
5080	1	tbd
	2	Qwen3:32b	24.9	505	20.3
5070 ti	1	tbd
	2	Qwen3:32b	23.8	505	21.2
5070	1	tbd
	2	Qwen3:14b	17.9	445	24.9
RTX 3060	1	Qwen3:14b	21.9
	2	Qwen3:32b	9.9	355	35.9

AMD gpu’s

GPU	# of gpu’s	model	token / s	wattage	watt / token
RX 570-8gb	1
	2	Qwen2.5:14b	8.34	365	43.5

Here’s the two RTX 5070 ti’s going at it, with Qwen3:32b

And here’s the two RTX 5080’s & single 5090 – Qwen3:32b

Here’s a link to watch them all:

Country Boy Computers BG – YouTube