Imagine installing a local AI tool on your gaming PC. You have a Radeon RX 7900 XT in there, a card with 20 GB of memory built for exactly the kind of parallel math AI needs. You run the tool. It thinks for a moment, then quietly routes everything through your CPU instead.
No error message. No warning. Just slow.
This happened to a lot of people. It was not a hardware failure or a misconfiguration. It was a matchmaking problem, a gap in a small configuration table. Understanding it makes the whole stack from Part 1 click into place.
The Name on the Box vs. the Name in the Code
When you buy a graphics card, you know it by its retail name. Radeon RX 7900 XT. RX 6700 XT. These names are designed to be memorable and to signal a product tier. They live on the box and in marketing copy.
Inside the chip, a different name matters: the architecture code. Every GPU belongs to a family of chips built on the same underlying design. AMD names these families with codes like gfx1100, gfx1102, gfx1151. The code describes which generation of AMD's GPU design you have, specifically what instruction sets the chip supports, what memory layout it uses, and what interfaces it exposes to drivers.
Software that wants to do serious work with your GPU needs the architecture code, not the retail name. The retail name is for the shelf. The architecture code is for the driver stack.
| Retail name | Architecture | Code pattern |
|---|---|---|
| Radeon RX 5000 series | RDNA 1 | gfx101X |
| Radeon RX 6000 series | RDNA 2 | gfx103X |
| Radeon RX 7000 series | RDNA 3 | gfx110X |
| Radeon RX 9000 series | RDNA 4 | gfx120X |
The X at the end is a wildcard covering the variants within that generation. gfx1100, gfx1101, and gfx1102 are all RX 7000-class chips with minor differences; they all belong to the gfx110X family and share the same ROCm package.
What ROCm Needs to Work
In Part 1, we covered the compute layer, the software bridge between AI math and GPU hardware. ROCm is AMD's version of that bridge.
ROCm does not ship as one universal package that runs on any AMD chip. It ships as a collection of prebuilt packages, each compiled specifically for a given architecture family. The compilation matters: AI operations translated for gfx110X will not run correctly on gfx103X hardware because the instruction sets differ.
So to install ROCm support for your card, a tool needs to do three things in sequence:
- Detect your GPU and find its architecture code
- Look up the matching package in a configuration table
- Download and install it
Step 2 is where things broke down for consumer Radeon cards.
The lookup was a table, a mapping from architecture code to package name. For datacenter hardware, those entries existed. For the RX 5000 through RX 9000 families that gamers actually own, the table had gaps. The tool would detect gfx1100 (an RX 7000 card), search the table, find nothing, and fall back silently to Vulkan or CPU.
The GPU was fully capable. The installation just did not know where to send it. This has since been fixed in the Lemonade project, and the cards that were falling through now get the ROCm path automatically.
Find Your Card
The Fallback Ladder
When the best path is unavailable, local AI tools do not crash. They step down quietly through a priority list:
- ROCm is AMD's full GPU acceleration path. Fastest for AMD hardware, and now the default for RX 5000 through RX 9000 cards on Lemonade.
- Vulkan is a general-purpose graphics standard that runs on almost any GPU made in the last decade. It was designed for games rather than AI math, so it is noticeably slower for inference, but it works broadly and requires no special installation.
- CPU uses no GPU at all. It works on any machine. For larger models it can be 10 to 50 times slower than GPU inference.
The fallback is silent by default. There is no error, no warning. The tool just takes the slower path. If you have ever run a local AI tool and wondered why it felt sluggish on hardware that seemed capable, this is the most common explanation.
Why Honest Fallback Beats Forcing It
One thing worth understanding about how this kind of problem gets fixed in open-source software: when a chip has no published ROCm binary, the right answer is to fall back cleanly rather than try to map it to the nearest available package.
Forcing a mapping to an incompatible package might work by accident on some machines. On others, it produces driver failures that are hard to diagnose because nothing obviously went wrong. The person who gets Vulkan and knows why is better off than the person who gets a broken ROCm install and has no idea what happened.
Honest fallback is better software, even when it looks like it is doing less.
What This Means for the Hands-On Part
With the mapping table complete for current Radeon gaming hardware, the path from "install Lemonade" to "ROCm running on your RX card" is now automatic. There is nothing to configure. The detection happens, the package installs, and the fast path is used.
Part 3 is where we test it: install Lemonade on a 16 GB-class Radeon card, pick a model sized for the hardware, and verify the ROCm path is actually in use. Then we look at what the hardware realistically does with it in terms of speed and model capability.