GPU acceleration and local AI
Jetson Orin Nano vs Raspberry Pi 5.
What changes when you need an NVIDIA GPU.
For Raspberry Pi 5 users who want an NVIDIA GPU for local AI. This page covers how the hardware compares, what carries over from a Pi and what changes. Every Raspberry Pi figure links to Raspberry Pi Ltd and every Jetson figure links to NVIDIA.
Figures checked on 20 October 2026
01 - Which Jetson Nano?
The 2019 Jetson Nano is a different module.
Two generations of NVIDIA modules carry the Nano name, and older comparisons with the Raspberry Pi are often about the first one. The Jetson Nano, which NVIDIA announced in March 2019, has a 128-core Maxwell GPU, a quad-core Cortex-A57 CPU and 4 GB of LPDDR4, according to NVIDIA's Jetson Nano page. The Jetson Orin Nano has an NVIDIA Ampere GPU with 512 or 1024 CUDA cores and a 6-core Cortex-A78AE CPU, according to NVIDIA's Jetson Orin page.
Both modules measure 69.6 mm x 45 mm and plug in with a 260-pin connector, but NVIDIA's FAQ lists the Orin Nano and Orin NX as form-factor compatible with the Jetson Nano, not pin compatible. The software is different too. The same FAQ lists the Jetson Nano under JetPack 4, which it marks as end of life, and JetPack 4.6.6 was the final JetPack 4 release. Our JetPack version table lists JetPack 5.1.1 to 7.2.1 for the Orin Nano and Orin NX.
The CB302B does not take the 2019 Jetson Nano. It takes four modules: the Jetson Orin Nano 4GB, Orin Nano 8GB, Orin NX 8GB and Orin NX 16GB. The rest of this page is about those four.
02 - Module and carrier board
A single board computer vs a module on a carrier board.
A Raspberry Pi 5 is one board. The processor, the memory and every connector are on it, and you choose a memory size when you buy it: the Raspberry Pi 5 product brief lists 1GB, 2GB, 4GB, 8GB and 16GB.
A Jetson Orin Nano or Orin NX is a module: a 69.6 mm x 45 mm card with a 260-pin SO-DIMM connector that carries the GPU, the CPU and the memory. The spec table on NVIDIA's Jetson Orin page lists no onboard storage for these four modules and notes that they support external NVMe. The ports, the power input, the storage slot and the headers all come from the carrier board the module plugs into.
The software comes later. NVIDIA's FAQ says each module ships with no software installed: you attach it to a carrier board and flash it with your software image. NVIDIA's own developer kits pair a module with NVIDIA's reference carrier board.
So moving from a Pi means two choices instead of one: the module, which sets the GPU, CPU and memory, and the carrier board, which sets everything you plug in. The CB302B is a carrier board. The bare board ships without a module, for people who already have one, and the Curve Developer Kit can come with a module we source, or you can bring your own.
03 - Side by side
Side by side.
Each figure below is the maker's own specification, linked in the last column. None of them is a benchmark score. On a Jetson system the ports and power input come from the carrier board, so the module table covers the module only. The CB302B's power and display are in section 07.
Raspberry Pi 5
| Spec | Raspberry Pi 5 | Source |
|---|---|---|
| Processor | Broadcom BCM2712, quad-core 64-bit Arm Cortex-A76 at 2.4GHz | Raspberry Pi product brief |
| GPU | VideoCore VII at 800MHz, OpenGL ES 3.1, Vulkan 1.3 | Raspberry Pi product briefRaspberry Pi 5 product page |
| Memory | LPDDR4X-4267: 1GB, 2GB, 4GB, 8GB or 16GB | Raspberry Pi product brief |
| Storage | microSD card slot, with support for SDR104 high-speed mode | Raspberry Pi product brief |
| PCIe | PCIe 2.0 x1, needs a separate M.2 HAT or other adapter | Raspberry Pi product brief |
| Camera and display | 2 × 4-lane MIPI camera / display transceivers | Raspberry Pi product brief |
| Display output | Dual 4Kp60 HDMI, on two micro-HDMI ports | Raspberry Pi product briefRaspberry Pi getting started guide |
| Power input | 5V/5A DC over USB-C, with Power Delivery | Raspberry Pi product brief |
| GPIO | Raspberry Pi standard 40-pin header | Raspberry Pi product brief |
Jetson Orin Nano and Orin NX modules
| Spec | Orin Nano 4GB | Orin Nano 8GB | Orin NX 8GB | Orin NX 16GB | Source |
|---|---|---|---|---|---|
| GPU (NVIDIA Ampere) | 512 CUDA cores, 16 Tensor Cores | 1024 CUDA cores, 32 Tensor Cores | 1024 CUDA cores, 32 Tensor Cores | 1024 CUDA cores, 32 Tensor Cores | NVIDIA Super Mode table |
| GPU clock, original / Super Mode | 625 / 1020 MHz | 625 / 1020 MHz | 765 / 1173 MHz | 918 / 1173 MHz | NVIDIA Super Mode table |
| CPU | 6-core Arm Cortex-A78AE | 6-core Arm Cortex-A78AE | 6-core Arm Cortex-A78AE | 8-core Arm Cortex-A78AE | NVIDIA Jetson Orin |
| CPU clock, original / Super Mode | 1.5 / 1.7 GHz | 1.5 / 1.7 GHz | 2.0 GHz | 2.0 GHz | NVIDIA Super Mode table |
| Memory | 4GB 64-bit LPDDR5 | 8GB 128-bit LPDDR5 | 8GB 128-bit LPDDR5 | 16GB 128-bit LPDDR5 | NVIDIA Jetson Orin |
| Memory bandwidth, original / Super Mode | 34 / 51 GB/s | 68 / 102 GB/s | 102 GB/s | 102 GB/s | NVIDIA Super Mode table |
| Peak AI performance, sparse INT8, original / Super Mode | 20 / 34 TOPS | 40 / 67 TOPS | 70 / 117 TOPS | 100 / 157 TOPS | NVIDIA Super Mode table |
| Module | All four: 69.6 mm x 45 mm, 260-pin SO-DIMM connector | NVIDIA Jetson Orin | |||
| Onboard storage | All four: no onboard storage listed; NVIDIA notes external NVMe support | NVIDIA Jetson Orin | |||
NVIDIA added Super Mode, a set of higher power modes, for these four modules with JetPack 6.2 in January 2025. It raises the GPU clock on all four, and the CPU clock and memory bandwidth on the Orin Nano. Where a cell shows one number, NVIDIA's Super Mode table gives the same figure in both modes.
The TOPS figures are NVIDIA's peak INT8 figures with sparsity. NVIDIA's table lists the GPU and the deep learning accelerator (DLA) separately, and on the Orin NX the peak figure is the two added together: for the Orin NX 16GB in Super Mode, 77 TOPS from the GPU and 80 from the DLA. The Orin Nano has no DLA, so its figure is the GPU alone.
The same table gives dense INT8 figures at about half (78 TOPS for the Orin NX 16GB in Super Mode). Some of NVIDIA's pages still quote the original figures, so check which mode a TOPS number refers to before you compare it.
These are NVIDIA's figures for the modules, not measurements of a CB302B.
04 - GPU acceleration and local AI
GPU acceleration and local AI.
Both platforms can run AI models on the device, with no cloud service involved. They get there differently.
Raspberry Pi 5
Raspberry Pi's product page lists a VideoCore VII GPU supporting OpenGL ES 3.1 and Vulkan 1.3, and the product brief gives its clock as 800MHz (the brief still lists Vulkan 1.2).
For AI work, Raspberry Pi makes the AI HAT+ and AI HAT+ 2: add-on boards with a Hailo neural processing unit (NPU) that connect to the Pi 5's PCIe port. The AI HAT+ comes with a Hailo-8L or Hailo-8 and uses the Pi 5's own memory. The AI HAT+ 2 has a Hailo-10H and 8 GB of memory on the HAT, which Raspberry Pi says lets it run local large language models. With either one fitted, Raspberry Pi OS detects the accelerator, and Raspberry Pi's camera frameworks can run vision models on it.
The earlier AI Kit is no longer in production, and Raspberry Pi recommends the AI HATs instead.
Jetson Orin Nano and Orin NX
The main accelerator is the GPU on the module: an NVIDIA Ampere GPU with 512 or 1024 CUDA cores and 16 or 32 Tensor Cores, per NVIDIA's Super Mode table. The two Orin NX modules also have NVIDIA deep learning accelerators (DLA), which the same table lists separately. The Orin Nano has none.
The software comes with NVIDIA JetPack: CUDA, cuDNN and TensorRT on top of Linux for Tegra (L4T). NVIDIA's JetPack 6.2.2 page lists CUDA 12.6, TensorRT 10.3 and cuDNN 9.3, and Vulkan 1.3 for graphics.
JetPack 6.2.2 is also the release behind the one CB302B software configuration we publish, for the curve-extras package. To see what a board is running, check its JetPack and L4T version.
TOPS figures are not directly comparable
Vendors measure TOPS differently, at different precisions and with or without sparsity, so one maker's TOPS figure does not line up with another's. A TOPS figure also does not tell you which board runs a given model faster, and none of the figures on this page is a benchmark of a real model.
05 - What carries over
What carries over.
Python and Linux
Both are 64-bit Arm Linux computers: the Pi 5 product brief lists a Cortex-A76 CPU, and NVIDIA's Jetson Orin page lists Cortex-A78AE cores for the Orin Nano and Orin NX. The Linux command line works as you know it. Python code that does not depend on Pi-specific libraries usually carries over, though the Python version and system packages are the ones in L4T's root file system rather than Raspberry Pi OS. That file system is based on Ubuntu: 22.04 for JetPack 6.2.2, per NVIDIA's JetPack 6.2.2 page. Code that uses the GPU on a Jetson goes through CUDA and the libraries in JetPack.
GPIO code through Jetson.GPIO
If your GPIO code uses the RPi.GPIO API, NVIDIA's Jetson.GPIO library has the same API. Its README says the module can also be imported under the name RPi.GPIO for existing code, and it supports the same BOARD and BCM pin numbering modes.
One difference to plan for: Jetson.GPIO does not emulate PWM in software. PWM works only on pins connected to a hardware PWM controller, with the pinmux set up for it.
We have not tested Jetson.GPIO on the CB302B yet. The library works out which board it is on from the device tree, so whether it recognises the CB302B by itself depends on the board support package.
The 40-pin header
The CB302B uses the same 40-pin GPIO format and 22-pin camera connector format found on the NVIDIA Jetson Orin Nano Developer Kit. We have not tested Raspberry Pi HATs on the CB302B, so we make no claim that they fit or work.
06 - What changes
What changes.
The operating system
In place of Raspberry Pi OS, a Jetson runs Linux for Tegra (L4T), NVIDIA's board support package: the bootloader, the kernel, an Ubuntu-based root file system and NVIDIA's drivers. JetPack adds the GPU software on top. The CB302B runs Linux for Tegra with JetPack, and native Android 17 / LineageOS 24 (beta) is a separate software environment.
Flashing instead of imaging an SD card
Raspberry Pi recommends writing the OS with Raspberry Pi Imager on another computer, and by default a Pi looks for an OS on its SD card (see Raspberry Pi's getting started guide).
A Jetson module arrives blank and is flashed from a host computer. NVIDIA's L4T flashing guide runs its scripts, flash.sh and l4t_initrd_flash.sh, on the host, with the Jetson in Force Recovery Mode and its recovery port connected to the host by a USB cable. NVIDIA SDK Manager does the same job from a host: it flashes the OS and installs the SDK over USB or a network connection.
NVIDIA's SD card image is for its own Jetson Orin Nano Developer Kit (see the JetPack 6.2.2 page), and JetPack 7.2.1 replaces it with an ISO image written to a USB stick.
On the CB302B, the Type-C 2 port carries USB debug and flashing support, and a button header brings out force recovery and reset. We have not published CB302B flashing steps yet, so don't assume a developer kit image or procedure applies to it.
Storage
A Pi 5 boots from a microSD card by default. NVIDIA lists no onboard storage for the four Orin modules and notes that they support external NVMe. The CB302B has an M.2 Key M slot for 2280, 2260 and 2242 NVMe SSDs, tested with drives from 256GB to 4TB.
Choosing a module
A Pi 5 comes in memory sizes. A Jetson module sets the GPU, the CPU and the memory together, and the four modules differ on all three, as the module table shows. We cover how to choose a Jetson module in its own guide.
07 - Power and display
Power and display away from a desk.
Raspberry Pi 5
The product brief lists 5V/5A DC power over USB-C, with Power Delivery, and Raspberry Pi's getting started guide recommends its 27 W USB-C power supply.
Displays connect to two micro-HDMI ports, and the brief lists dual 4Kp60 HDMI output.
CB302B
Two ways to power it: USB-C PD up to 45 W (15 V / 3 A) on the Type-C 2 port, or M12 battery power from a genuine Milwaukee M12 pack (3Ah, 6Ah or 9Ah) that you supply. Curve does not sell battery packs. An internal bridge keeps the board powered while you change packs.
Video goes out as DisplayPort over USB-C on the Type-C 1 port. There is no HDMI port.
Battery runtime depends on the module, power mode, workload, peripherals and pack.
08 - Robotics teaching labs
For school and university robotics labs: batteries students don't have to wire.
Putting a board on a robot means getting power to it. With the CB302B, students use genuine Milwaukee M12 packs rather than building or wiring their own battery packs. The power options are in section 07.
One carrier board takes any of the four modules, from the Orin Nano 4GB to the Orin NX 16GB, so groups on different budgets can work on the same board. More on this in robotics teaching labs on our use cases page.
09 - Where the CB302B fits
Where the CB302B fits.
The CB302B Jetson Orin Nano and Orin NX carrier board is for Pi users who want an NVIDIA GPU without designing a carrier board. It takes the Jetson Orin Nano 4GB, Orin Nano 8GB, Orin NX 8GB or Orin NX 16GB, and supports Orin Super Mode. It runs Linux for Tegra (L4T) with JetPack, it is ROS compatible, and native Android 17 / LineageOS 24 is in beta.
If you don't have a module yet, the Curve Developer Kit can come with one we source.
As published on /products/cb302b
10 - Questions
Questions from Pi users.
Will my Raspberry Pi camera work on the CB302B?
We have not tested Raspberry Pi camera modules on the CB302B, so we make no claim either way. If a particular camera matters for your project, ask us before you reserve.
Do Raspberry Pi HATs work on the CB302B?
We have not tested them. The CB302B uses the same 40-pin GPIO format found on the NVIDIA Jetson Orin Nano Developer Kit, but a matching header format on its own does not tell you that a HAT will work.
Can a Raspberry Pi 5 use a GPU?
The Pi 5 has its own VideoCore VII GPU. For add-on hardware, Raspberry Pi's product brief lists a PCIe 2.0 x1 interface for fast peripherals, which needs a separate M.2 HAT or other adapter. We don't cover third-party graphics card setups here.
11 - Sources
Sources.
Every figure on this page comes from one of these pages, checked on 20 October 2026. CB302B details are as published on our own product pages.
Raspberry Pi Ltd
- Raspberry Pi 5 product brief
Edition published October 2026: processor, GPU clock, memory, PCIe, MIPI, power input, GPIO
- Raspberry Pi 5 product page
GPU graphics APIs: OpenGL ES 3.1 and Vulkan 1.3
- Getting started
micro-HDMI ports, 27 W supply, Raspberry Pi Imager, SD card boot
- AI HAT+ and AI HAT+ 2
Hailo NPU, PCIe connection, memory and LLM support
- AI Kit
No longer in production
NVIDIA
- JetPack 6.2 brings Super Mode (Table 2)
Original and Super Mode figures for all four modules, with GPU and DLA TOPS listed separately, January 2025
- Jetson Orin
CPU, memory, module size, connector and storage
- Jetson Nano
The 2019 module: GPU, CPU, memory and size
- Jetson Nano announcement
NVIDIA press release, March 2019
- Jetson FAQ: JetPack versions
JetPack 4 end of life, and the modules each release line supports
- Jetson FAQ: developer kits and modules
Modules ship with no software installed
- Jetson FAQ: module compatibility
Pin and form-factor compatibility between modules
- End of life for JetPack 4
JetPack 4.6.6 is the final JetPack 4 release
- JetPack 6.2.2
CUDA, cuDNN, TensorRT and Vulkan versions, SD card image for the Orin Nano Developer Kit
- JetPack 7.2.1
SD card image replaced by a USB stick ISO
- L4T r36.5 flashing support
Flash scripts, Force Recovery Mode and the USB recovery port
- SDK Manager
Flashing and SDK installation over USB or network
- Jetson.GPIO on GitHub
RPi.GPIO API, PWM and board detection
12 - Next
Bring an NVIDIA GPU to your next project.
Reserve a board for the Orin Nano or Orin NX module you have, or start by choosing one. The CB302B is not shipping yet: a reservation joins the waitlist, and no payment is taken.
Figures checked against Raspberry Pi Ltd and NVIDIA pages on 20 October 2026. Re-check at the next Raspberry Pi 5 product brief edition or JetPack release.
Curve Reality - Rev 2026.04
Designed in New Zealand