🚨🚨🚨 Did AMD and/or Gigabyte intentionally or unintentionally just make local "AI" models available to a lot more people?
I have an old AM4 Gigabyte GA-AX370-Gaming 5 motherboard with a 6 core 5600G processor in it. Gigabyte released a new BIOS for that motherboard on September 3rd 2026. Changelog story is that version F54d is a "Fix AMD TPM Reference Code Errata (CVE-2026-6726, CVE-2026-6727)"
https://www.gigabyte.com/Motherboard/GA-AX370-Gaming-5-rev-10/support#Support-Bios
I updated the motherboard to the shiny new F54d BIOS and it re-set the BIOS settings. This is annoying but not unusual when upgrading a BIOS. Thus, I had to poke around in the BIOS in order to restore my previous preferences and that's when I noticed something new that's not indicated in the changelog:
The area of the BIOS where you can choose how much RAM should be assigned to the integrated graphics part of the CPU (if you have one with integrated graphics) goes all the way up to 16 GB in BIOS F54d. The max used to be 2 GB, I've never seen any BIOS where you could assign more than 2 GB to the iGPU/APU.
I tried assigning 8GB to the 5600G in the BIOS and then I tried running the Q3_K_M version of the Kreamagine Krea 2 Turbo checkpoint on it. Krea 2 is a image generation model, so horsepower matters more than it does if you run some LLM. I didn't expect it to run very fast, but I was curious if it would run at all since that was previously not possible when the max assigned GPU RAM was limited to 2 GB.
Kreamagine Q3_K_M did run and I got a awfully slow 29.35 seconds per iteration. For context, the RX 6600 runs that model at around 4.93s/it. Half a minute times 10 iterations (what you ideally want for Krea 2 Turbo models) works out to 5 minutes compared to 50 seconds using the RX 6600. That's quite the execution time just to get an image resembling "1girl, confused, using computer running ai app, glasses, blond hair, tight shirt". You're not going to get lightning fast execution if you assign more RAM to GPU integrated in a CPU, but the more important detail is that it actually works.
WHY I THINK THIS IS SO COOL: If you have a now nearly 10 year old computer with an AMD CPU with integrated graphics and 16 or 32 GB DDR4 bought back when it was less than half of today's price then you're now able to assign 8 or 16 GB RAM to the graphics part of the CPU and use that to do a whole range of useful things like as offloading an image models text encoder if you have some older graphics card like a RX 6600 with only 8 GB VRAM.
It's not like I bought anything new or expected to get this new feature. I just updated the BIOS and unexpectedly got the ability to assign up to 16 GB VRAM to the integrated graphics chip. If you have some old AM4 board with a G series CPU and you play around with local models then you should see if there is a new BIOS for your board that may or may not give you the ability to bump the iGPUs memory from 2 to 16 GB.
#ai #localai #amd #apu