# Can an RX 5700 XT run AI? Measured: 63–72 tokens a second

> Real test: Cube 4B (Gemma 4 E4B) wrote 63–72 tokens a second on a desktop with an AMD RX 5700 XT (8 GB) and 48 GB of RAM. What it means for chat, study and pictures.

Canonical: https://blog.greencube.app/blog/ai-on-rx-5700-xt  
Published: 2026-09-22  
Author: Hector Gras

**Yes.** On a desktop with an AMD RX 5700 XT (8 GB) and 48 GB of RAM, Cube 4B (Gemma 4 E4B) wrote 63–72 tokens a second. You read about 5, so answers come out faster than you can read them.

- **63–72** tokens a second with Cube 4B (Gemma 4 E4B)
- **14×** faster on the card than on the processor
- **20–36** with Cube 12B (Gemma 4 12B) (short chats only)

## What that means in practice

| Task | Time on this PC |
| --- | --- |
| Write a normal answer (400 tokens, about 300 words) | **6 s** |
| Read a long document you give it (5,000 tokens, about 7 pages) | **8 s** |

Worked out from the speeds measured on this PC: time is length divided by speed, and a real answer also spends a moment getting started. For scale, most people read about 4 words a second ([Brysbaert, 2019](https://doi.org/10.1016/j.jml.2019.104047)), so a 400-token answer takes you over a minute just to read.

## What we measured

**Writing** is how fast the AI types its answer, in tokens a second; a token is about three quarters of an English word. **Reading** is how fast it takes in what you give it: your question, a pasted text or a PDF. Reading is always much faster than writing, because the AI can read many tokens at once but has to write them one by one.

| What | Ran on | Result | Measured |
| --- | --- | --- | --- |
| Cube 4B (Gemma 4 E4B) | graphics card | **63–72** tokens a second writing · 594–609 reading | 22 September 2026 |
| Cube 4B (Gemma 4 E4B) | processor only | **5** tokens a second writing · 357 reading | 22 September 2026 |
| Cube 12B (Gemma 4 12B) | graphics card | **20–36** tokens a second writing · 108–113 reading | 22 September 2026 |

How each number was measured: [More on how we measure](https://blog.greencube.app/blog/how-we-measure-ai-speed)

1. llama.cpp's own speed test (llama-bench), the engine GreenCube runs on, with the same AI files: 128 tokens written, a 512-token text read.

Where we have several runs we show the range: speeds move a little between runs on the same PC.

## What it means

### The graphics card does the work

Same PC, same AI (Cube 4B (Gemma 4 E4B)): 5 tokens a second on the processor alone, 63–72 on the graphics card, about **14 times faster**. A graphics card is built to do thousands of small sums at once for games, and running an AI is mostly that same kind of maths. GreenCube times both when it starts and uses the faster one.

### The whole AI fits on the card

Cube 4B (Gemma 4 E4B) is a 4.2 GB file and the card has 8 GB of its own memory (VRAM). All of it fits, with room left for the conversation, so the card never waits for the slower memory in the rest of the PC. That is the biggest reason it runs this fast.

### The bigger Cube 12B (Gemma 4 12B)

Only for short chats. The bigger Cube 12B (Gemma 4 12B) (6.4 GB) wrote 20–36 tokens a second on a short question, but once the chat reached 8,192 tokens (a long conversation) it crashed, with the card almost full. That is why GreenCube only suggests the 12B on cards with 10 GB or more.

### Pictures

Very likely, but we haven't timed a picture on this card yet. When we do, the number goes here. A picture uses its own AI of about 6 GB and far more maths than a chat answer, so the graphics matter even more here.

## How it compares

Of the 8 PCs we measured with Cube 4B (Gemma 4 E4B), this one is **number 2** for writing speed. The [RTX 2080 Ti](https://blog.greencube.app/blog/ai-on-rtx-2080-ti) writes 65–71, about the same. The [RTX 5060](https://blog.greencube.app/blog/ai-on-rtx-5060) writes 56 (startup test), so this one is about 20% faster.

We measured 4 cards with 8 GB, and they are far apart:

- **RX 5700 XT**: 63–72 tokens a second
- [RTX 5060](https://blog.greencube.app/blog/ai-on-rtx-5060): 56 tokens a second (startup test)
- [RX 6600](https://blog.greencube.app/blog/ai-on-rx-6600): 42–44 tokens a second
- [RX 580](https://blog.greencube.app/blog/ai-on-rx-580): 10–11 tokens a second

Memory decides which AI fits. How new the card is decides how fast it runs.

## The test PC

A real tester's own PC, used the way they normally use it, not a lab machine:

- Desktop
- **Graphics:** AMD RX 5700 XT
- **Graphics memory (VRAM):** 8 GB
- **RAM:** 48 GB
- **Processor:** Intel Core i7-9700, 8 threads

## Check your own PC

Not sure what your PC has? It takes 30 seconds:

1. Press **Ctrl + Shift + Esc** to open Task Manager.
2. Go to **Performance**.
3. Click **Memory**: the number at the top right is your RAM.
4. Click **GPU**: the name at the top right is your card, and **Dedicated GPU memory** is your VRAM. Only see Intel or AMD graphics with no dedicated memory? Then you have no separate card.

Then find the closest card in [the full table](https://blog.greencube.app/blog/can-my-pc-run-ai). The same card with more VRAM can fit a bigger AI; a newer card of the same size is usually faster.

## Other PCs we measured

- [RTX 2080 Ti](https://blog.greencube.app/blog/ai-on-rtx-2080-ti): 65–71 tokens a second (Cube 4B (Gemma 4 E4B))
- [RTX 5060](https://blog.greencube.app/blog/ai-on-rtx-5060): 56 tokens a second (Cube 4B (Gemma 4 E4B))
- [RTX 3060](https://blog.greencube.app/blog/ai-on-rtx-3060): 53–55 tokens a second (Cube 4B (Gemma 4 E4B))
- [RX 6600](https://blog.greencube.app/blog/ai-on-rx-6600): 42–44 tokens a second (Cube 4B (Gemma 4 E4B))
- [GTX 1660 Super](https://blog.greencube.app/blog/ai-on-gtx-1660-super): 37–38 tokens a second (Cube 4B (Gemma 4 E4B))
- [GTX 1650](https://blog.greencube.app/blog/ai-on-gtx-1650): about 20 tokens a second (Cube 4B (Gemma 4 E4B))
- [Intel Core Ultra 5 laptop](https://blog.greencube.app/blog/ai-on-intel-core-ultra-5-laptop): 14–18 tokens a second (Cube 2B (Gemma 4 E2B))
- [RX 580](https://blog.greencube.app/blog/ai-on-rx-580): 10–11 tokens a second (Cube 4B (Gemma 4 E4B))
- [Intel Iris Xe laptop](https://blog.greencube.app/blog/ai-on-intel-iris-xe-laptop): 5–7 tokens a second (Cube 2B (Gemma 4 E2B))

[Can my PC run AI?](https://blog.greencube.app/blog/can-my-pc-run-ai) All 10 PCs in one table, and what your PC will feel like.

## FAQ

### Can an RX 5700 XT run AI locally?

Yes. On a desktop with an AMD RX 5700 XT (8 GB) and 48 GB of RAM, Cube 4B (Gemma 4 E4B) wrote 63–72 tokens a second. You read about 5, so answers come out faster than you can read them.

### Can an RX 5700 XT make AI pictures?

Very likely, but we haven't timed a picture on this card yet. When we do, the number goes here.

### Is AI faster on the RX 5700 XT or on the processor?

On the card, by far: 63–72 tokens a second against 5 on the processor, about 14 times faster.

### Can an RX 5700 XT run Gemma 4 12B?

Only for short chats. The bigger Cube 12B (Gemma 4 12B) (6.4 GB) wrote 20–36 tokens a second on a short question, but once the chat reached 8,192 tokens (a long conversation) it crashed, with the card almost full. That is why GreenCube only suggests the 12B on cards with 10 GB or more.
