What was tested
Saha used three systems, each from a different generation of AMD integrated graphics:
- ASUS ROG Flow Z13 with a Ryzen AI Max 390 and 32 GB of memory.
- First-generation Lenovo Legion Go with a Ryzen Z1 Extreme and 16 GB of memory.
- A desktop Ryzen 5 5600G, using its Vega 7 integrated graphics.
These are shared-memory machines. None has dedicated VRAM. On the 5600G, Saha manually assigned 16 GB to the iGPU in the BIOS. That was the maximum the board allowed.
The benchmark scripts were written with help from Anthropic's Claude. Saha describes the aim as using every resource each machine had.
The reported numbers
These are the author's figures, not independently verified ones:
| Configuration | Generation (tokens/s) | Prompt processing (tokens/s) | Time per run (s) | Avg power (W) | Tokens/s per W | Peak temp (°C) |
|---|---|---|---|---|---|---|
| Ryzen 5 5600G, Vega 7 | 17.5 | 201 | 16.8 | 41 | 0.43 | 57 |
| Z1 Extreme, CPU only | 22.4 | 150 | 13.7 | 18 | 1.25 | 64 |
| Z1 Extreme, 780M iGPU | 44.1 | 783 | 6.3 | 33 | 1.34 | 81 |
| Ryzen AI Max 390, 8050S | 85.1 | 1,609 | 3.2 | 64 | 1.33 | 95 |
Some points stand out:
- The AI Max 390 leads. It generates 85.1 tokens/s, about twice the Z1 Extreme's 780M result and roughly five times the Vega 7.
- The 780M beats the CPU alone. Generation roughly doubles, from 22.4 to 44.1 tokens/s. Prompt processing is about five times faster, at 783 against 150.
- The old Vega 7 is the odd one out. Its generation speed of 17.5 tokens/s is lower than the Z1 Extreme's CPU-only run. Its prompt processing is a little higher than that CPU run, but far below the 780M. It also draws 41 W for the worst efficiency in the table.
- Efficiency is a near tie at the top. The 780M and the 8050S land at 1.34 and 1.33 tokens/s per watt. The Max 390 is much faster, but it needs about twice the power to get there.
- The Max 390 runs hot. It peaked at 95°C.
The 5600G setup
Getting the 5600G running took effort. Saha says the Vega 7 (gfx90c) has no official ROCm support, so he used the Vulkan backend for better compatibility. He also says Ollama doesn't use iGPUs by default, and he had to edit its configuration file to change that. He adds that ROCm can technically be forced onto unsupported GPUs, but he says it wouldn't help.
These are one person's notes on one setup. They are not general install instructions, and the article doesn't give the exact config changes.
A hardware detail the article gets wrong
The article describes the Flow Z13's graphics as a 40-compute-unit iGPU. That doesn't match the Ryzen AI Max 390. AMD's own spec listing gives the Ryzen AI Max 390 Radeon 8050S Graphics with a graphics core count of 32. Phoronix's testing agrees: the 8050S has 32 graphics cores at 2.8GHz, while the Radeon 8060S has 40 cores at 2.9GHz. The 40-CU 8060S is the flagship Ryzen AI Max+ 395 part. It does not ship in the 390.
The article's own table correctly labels the chip as 8050S. The 40 CU figure was probably borrowed from the Flow Z13 marketing copy for the higher-end configuration. The 85.1 tokens/s result therefore reflects a 32-CU part, not the flagship.
What the article leaves out
Readers shouldn't treat the table as a controlled benchmark. The article doesn't say:
- Which model was run, or at what quantization.
- How long the prompts and outputs were.
- Which inference software and version, operating system and drivers were used.
- Whether "time per run" covers identical work on all machines.
- Whether the power figures cover the whole system or only the chip, or how they were measured.
- How many times each test was repeated.
These gaps matter because the machines weren't tested the same way. The 5600G used Vulkan with a manually set 16 GB allocation. The other two machines' backends aren't described. Differences in backend, model size and memory bandwidth can swing results a lot. Read the table as "here's roughly what happened on one person's hardware," not as a ranking.
Memory limits
Memory capacity is the real constraint for local models. The Flow Z13 in this test had 32 GB, and Saha says that limits which models it can run. The chip family supports far more. AMD's own listing says the Ryzen AI Max line supports up to 128 GB of LPDDR5x. Laptop listings show Max 390 machines, such as HP's ZBook Ultra G1a, configured with 128 GB. The Z13 result shouldn't be read as the ceiling for the platform's model capacity.
Saha calls unified memory the future of local AI. That is a fair opinion, but it is an opinion. He also admits the main drawback: Ryzen AI Max machines are expensive, and their memory is soldered, so you can't upgrade it later. The article lists the Flow Z13 at $1,750 at major retailers. We haven't confirmed that price independently.
ROCm support is narrower than people assume
Saha's ROCm comment points to a common trap. AMD's ROCm 7.0.2 compatibility matrix says its accelerator and GPU table covers compute workloads only. For Radeon GPUs or Ryzen APUs used for graphics workloads, it sends readers to a separate "Use ROCm on Radeon and Ryzen" page. The matrix also targets Linux, and its listed GPU targets are datacenter and discrete Radeon parts. Vega-based gfx90c and the APUs in this test aren't in it.
If you plan to buy for local AI, check support for your exact chip, OS and inference app before you spend. A generic "ROCm supported" claim doesn't answer that.
Practical takeaways
- Try it before you buy anything. A recent Ryzen laptop or handheld can run small and mid-size quantized models on its iGPU. In this test, the 780M ran about twice as fast as the CPU alone.
- Don't expect GPU-card speeds from older chips. The Vega 7 worked but was slow and power-hungry, so it's a curiosity rather than a plan.
- Budget for memory first. Model size against available shared memory decides what you can run. Raw compute matters less.
- Expect some setup work. Backend choice matters. Vulkan was the working route on the 5600G. Tools like Ollama may need manual configuration to use an iGPU.
- Watch thermals. The fastest result also hit 95°C. Check cooling and power limits on thin laptops.
Bottom line
The headline holds in a narrow sense. A discrete GPU is no longer mandatory, and recent AMD integrated graphics can make local inference practical. It doesn't mean any iGPU will do. The best result came from a pricey, memory-rich Strix Halo platform, and the numbers come from an undocumented single-author test. The practical lesson is that iGPUs are now a legitimate option for local models, provided you verify your exact hardware and software path.