Raspberry Pi 5 vs. Pi 4 AI performance CPU Benchmark: How much leap forward?
We are thrilled about Raspberry Pi 5 finally being released! Hands on this new device with one question, how’s AI performance?
💡Answer: only $5 increases, 3x increases CPU, 4 x inference performance leap forward
Be the first one to get Raspberry Pi 5! >> Buy now
Welcome to the latest generation of Raspberry Pi: the everything computer. Featuring a 64-bit quad-core Arm Cortex-A76 processor running at 2.4GHz, Raspberry Pi 5 delivers a 2–3× increase in CPU performance relative to Raspberry Pi 4. Alongside a substantial uplift in graphics performance from an 800MHz VideoCore VII GPU; dual 4Kp60 display output over HDMI; and state-of-the-art camera support from a rearchitected Raspberry Pi Image Signal Processor, it provides a smooth desktop experience for consumers, and opens the door to new applications for industrial customers.
YOLOv8n model‘s performance on Raspberry Pi 5 8GB
We tested YOLOv8n model on Raspberry Pi 5 8GB with a video input of 640×640. Using Tencent’s ncnn inference framework, the inference rate is around 12 fps.
Benchmark on Raspberry Pi 4 8GB and Raspberry Pi 5 8GB
We conducted benchmark tests using the ncnn framework on both the Raspberry Pi 4 8GB and Raspberry Pi 5 8GB to evaluate inference performance. ncnn is an efficient and user-friendly deep learning inference framework that supports various neural network models (such as PyTorch, TensorFlow, ONNX, etc.) and a range of hardware (including x86, ARM, RISC-V, MIPS, Vulkan, and more). Ncnn was designed with mobile deployment in mind, and it offers GPU acceleration via the low-overhead Vulkan API. When leveraging the GPU capabilities of the Raspberry Pi 5, the results surpass those obtained with CPU-only inference. Here’s a comparison of model performance between the two Raspberry Pi model.
| Model | Raspberry Pi 4 @1thread avg inference rate, ms | Raspberry Pi 4 @4thread avg inference rate, ms | Raspberry Pi 5 @1thread avg inference rate, ms | Raspberry Pi 5 @4thread avg inference rate, ms |
| mobilenet_v2 | 93.48 | 61.88 | 19.24 | 13.63 |
| mobilenet_v3 | 72.81 | 47.53 | 13.25 | 9.48 |
| efficientnet_b0 | 129.73 | 74.19 | 26.24 | 14.87 |
| efficientnetv2_b0 | 155.76 | 79.34 | 39.97 | 19.99 |
| resnet18 | 191.1 | 123.12 | 32 | 24.12 |
| resnet18_int8 | 158.5 | 83.92 | 57.01 | 32.68 |
| resnet50 | 534.11 | 262.83 | 96.28 | 56.29 |
| resnet50_int8 | 423.96 | 196.65 | 117.57 | 73.59 |
| mobilenet_ssd | 251 | 121.46 | 53.07 | 31.48 |
| mobilenet_ssd_int8 | 170.56 | 71.14 | 46.35 | 31.27 |
| yolov4-tiny | 372.99 | 207.22 | 74.1 | 56.19 |
| vision_transformer | 6605.73 | 1909.21 | 1268.82 | 629.51 |
| FastestDet | 52.61 | 38.69 | 9.03 | 6.59 |
Tencent ncnn
ncnn is a versatile, cross-platform tool that performs exceptionally well on mobile phone CPUs and has no third-party dependencies. It empowers developers to deploy deep learning models seamlessly on mobile platforms, enabling the creation of intelligent apps and putting artificial intelligence at your fingertips. Notably, ncnn is currently employed in several Tencent applications, including QQ, Qzone, WeChat, Pitu, and more.
Experience the future of computing with the Raspberry Pi 5 and unlock new horizons for AI applications!
Resources
About Author
You wrote:
> When leveraging the GPU capabilities of the Raspberry Pi 5, the results surpass those obtained with CPU-only inference.
Did you do tests with the GPU? Currently i only see the CPU results, right?
Hi there, yes the results are only leverage CPU only. We will post about GPU performance soon.
Did you label the Pi 4 and Pi 5 the wrong way around? These results seem strange.
Shorter times mean quicker calculation