This is a mini-framework for querying parameters of an Apple-designed GPU. It also contains a command-line tool, gpuinfo, which reports information similarly to clinfo. It was co-authored with an AI.
Listed parameters:
- Name ✅
- Core count ✅
- Clock frequency ✅
- Bandwidth ✅
- FLOPS (FP32 operations per second) ✅
- IPS (shader instructions per second) ✅
- System-level cache ✅
- Memory ✅
- Family ✅
Interfaces:
- Swift module
- C bindings
- Command-line tool
Recognized devices:
- A7 - A16
- M1 - M1 Ultra
- M2 - M2 Max
- Future devices treated like the closest existing analog (e.g. A17 like A16)
One way to use this library is from the command-line:
git clone https://github.com/philipturner/applegpuinfo
cd applegpuinfo
swift run gpuinfo list
# Sample output
GPU name: Apple M1 Max
GPU core count: 32
GPU clock frequency: 1.296 GHz
GPU bandwidth: 409.6 GB/s
GPU FLOPS: 10.617 TFLOPS
GPU IPS: 5.308 TIPS
GPU system level cache: 48 MB
GPU memory: 32 GB
GPU family: Apple 7
You can also use it directly from Swift:
// Inside package manifest
dependencies: [
// Dependencies declare other packages that this package depends on.
.package(url: "https://github.com/philipturner/applegpuinfo", branch: "main"),
],
// Inside source code
import AppleGPUInfo
let device = try GPUInfoDevice()
print(device.flops)
print(device.bandwidth)Original Goal: In one hour, finish a mini-package and command-line tool for querying Apple GPU device parameters.
Results: I spent 57 minutes finishing the file that wraps the AppleGPUDevice structure. I asked GPT-4 to generate the tests and command-line tool. I renamed the command-line tool from applegpuinfo to gpuinfo according to the AI's suggestion. Finally, I congratulated it and asked for it to leave a comment to users on the README. That triggered a safeguard and it quit the conversation. The stop time was 1 hour, 25 minutes.
Documentation of AI contributions: bing-conversation.md
After creating the first release of the library, I have continued experimenting with workflows accelerated by free access to GPT-4. The above document details these subsequent modifications to the library.
This framework is confirmed to work on the following devices. If anyone wishes to contribute to this list, please paste the output of gpuinfo into a new GitHub issue. Different variations of the same chip (e.g. different cores or memory) are welcome.
| Production Year | Chip | Cores | SLC | Memory | Bandwidth | TFLOPS |
|---|---|---|---|---|---|---|
| 2017 | A10X | 12 | 0 MB | 4 GB | 68.2 GB/s | 0.768 |
| 2021 | A15 | 5 | 32 MB | 5.6 GB | 34.1 GB/s | 1.789 |
| 2021 | M1 Pro | 16 | 24 MB | 32 GB | 204.8 GB/s | 5.308 |
| 2021 | M1 Max | 32 | 48 MB | 32 GB | 409.6 GB/s | 10.617 |
| 2022 | M1 Ultra | 48 | 96 MB | 64 GB | 819.2 GB/s | 15.925 |
| 2023 | M2 Pro | 19 | 24 MB | 32 GB | 204.8 GB/s | 6.800 |
This project was made possible by GPT-4, accessed through Bing Chat.