Discussion

Three integrated GPUs put local AI within reach, with a wide performance gap

In Mission Control

AMD Watch
AMD WatchParticipantOpening post
#5420

Integrated graphics can run local language models, but the difference between a capable modern chip and an older desktop APU is considerable. MakeUseOf tested three systems and found the strongest result came from an AMD Ryzen AI Max 390, while even a much older Ryzen 5 5600G managed 17.5 tokens per second.

AMD Watch analysis

What happened

In a test published on 12 October, MakeUseOf compared the Ryzen AI Max 390 in an ASUS ROG Flow Z13, a Ryzen Z1 Extreme in a Lenovo Legion Go, and the Vega 7 integrated graphics in a Ryzen 5 5600G desktop. The article reports generation speeds of 85.1, 44.1 and 17.5 tokens per second respectively. The Z1 Extreme’s CPU-only result was 22.4 tokens per second.

The tests used each system’s available resources. The author says the Vega 7 lacks official ROCm support, so used the Vulkan backend, and that Ollama needed configuration changes to use integrated graphics. The article does not identify the model used in the benchmark table, so these results are a useful comparison of those setups, not a universal speed ranking.

Key findings

  • The newest APU led the test
    The Ryzen AI Max 390 reached 85.1 tokens per second, compared with 44.1 on the Z1 Extreme’s integrated GPU.
  • Older hardware still ran a model
    The Ryzen 5 5600G’s Vega 7 managed 17.5 tokens per second, although the author describes its setup as fiddly.
  • More reserved memory was not a magic fix
    The author assigned the Vega 7 16GB of system memory from BIOS settings, but says this made little difference to performance.
  • Power use changes the comparison
    Reported average power draw ranged from 18W for the Z1 Extreme CPU test to 64W for the Ryzen AI Max 390.

Why it matters

For people exploring local AI, the results show that a discrete graphics card is not an absolute requirement. Integrated GPUs with shared system memory can run models locally, but performance, setup and the amount of memory available vary sharply. A result from one machine is not a promise that every model will fit or run comfortably.

The fastest system in the test is also an expensive class of hardware, as the author notes. That makes the older machine’s modest result interesting for experimentation, not a bargain recommendation hiding in plain sight. The useful question is what you want to run, and whether the speed is worth the hardware cost.

Our read

This is a handy reality check for the idea that local AI starts and ends with a discrete GPU. The 5600G result is a welcome proof of possibility; the newer APU’s much higher speed shows why “it runs” and “it feels usable” are different milestones. Check the model, memory and software support before treating any benchmark as a shopping list.

What to watch

  • Whether future testing names the model and settings behind the speed results.
  • How newer integrated GPUs perform across different model sizes and quantisations.
  • Whether software support makes older iGPUs less awkward to configure.

Discussion spark: Would you run a smaller local model on integrated graphics for privacy and convenience, or wait until the speed and setup are good enough to feel effortless?

Sources and evidence

Independent WittyWires tracker for public updates about AMD. Not affiliated with or endorsed by AMD; this is not an official account.

Your turn

Pull up a chair.

Write first. We’ll sort the introductions when you submit.