In this article
Short answer
For AI inference, the NVIDIA Jetson Orin Nano Super is the board built for the job. NVIDIA rates it at 67 sparse INT8 TOPS, and Ultralytics lists a small vision model at 15.60 ms per image on it with PyTorch against 299.09 ms on the Raspberry Pi 5 CPU. The Pi 5 has no AI engine on the board.
- The Jetson has a GPU with 32 tensor cores; the Pi 5 lists no AI engine on the board.
- The Jetson's 67 TOPS is a sparse INT8 rating; NVIDIA lists 33 dense INT8 TOPS.
- Ultralytics times come from two different pages and versions, so they show scale and not a strict ranking.
- Raspberry Pi's AI HAT+ adds 13 TOPS or 26 TOPS over PCIe as a separate purchase.
- Pick 1NVIDIA Jetson Orin Nano Super Developer Kit 8GB
- Pick 2Raspberry Pi 5 8GB
What each maker lists#
The table sets out what NVIDIA and Raspberry Pi publish for the 8 GB boards. Where a maker does not list a figure, the cell says so.
| Spec | Jetson Orin Nano Super 8GB | Raspberry Pi 5 8GB |
|---|---|---|
| AI performance | 67 sparse INT8 TOPS (33 dense INT8 TOPS) | No AI engine on the board |
| Processor | 6-core Arm Cortex-A78AE, 1.7 GHz | Quad-core Arm Cortex-A76, 2.4 GHz |
| Graphics | Ampere GPU, 1024 CUDA cores, 32 tensor cores | VideoCore VII |
| Memory | 8 GB LPDDR5, 102 GB/s | 8 GB LPDDR4X-4267 |
| Power | 7 W to 25 W | 5 V/5 A USB-C supply |
| Storage and expansion | SD card slot, external NVMe | PCIe 2.0 x1 interface, 40-pin GPIO |
| In the box | 19 V power supply | Not stated |
NVIDIA's Jetson Orin table gives the Orin Nano Super as 67 sparse INT8 TOPS and 33 dense INT8 TOPS, and says an update lifted the kit from 40 TOPS and from 68 GB/s of memory bandwidth. Raspberry Pi lists the Pi 5 with a VideoCore VII graphics unit and offers AI add-ons separately.
Where the AI work happens#
The Jetson runs AI on its GPU. The GPU has 32 tensor cores, and NVIDIA says the kit supports vision transformers, language models and vision-language models. Software reaches the GPU through NVIDIA's JetPack tools, with CUDA and TensorRT. NVIDIA also names its Isaac robotics tools and its Metropolis vision platform as part of the ecosystem.
The Raspberry Pi 5 has no AI engine of its own. Without extra hardware, models run on the Arm CPU. Raspberry Pi's route to faster AI is the AI HAT+, a board that stacks on the Pi 5 and conforms to the HAT+ specification. It ships with a 16 mm stacking header, spacers and screws, and Raspberry Pi lists it for the Pi 5 with the Active Cooler in place. The 13 TOPS version uses a Hailo-8L chip. The 26 TOPS version uses a Hailo-8.
Raspberry Pi says the add-on is integrated with its camera software stack and names TensorFlow and PyTorch. It lists an operating temperature of 0 to 50 degrees Celsius ambient. A newer AI HAT+ 2 uses a Hailo-10H, is listed at 40 TOPS at INT4 and carries 8 GB of on-board RAM. Raspberry Pi says its vision performance is comparable to the 26 TOPS board, with added support for generative AI models.
The practical effect is that the two platforms divide the work differently. On the Jetson, the AI engine and the computer are one board. On the Pi, the computer and the accelerator are two purchases, and the accelerator has its own limits, set by the model formats its chip supports.
Memory and power
Both boards offer 8 GB in the versions compared here. On the Jetson the 8 GB is LPDDR5 at 102 GB/s, and NVIDIA says the same update that raised the AI rating raised bandwidth from 68 GB/s. Raspberry Pi lists the Pi 5's memory as LPDDR4X-4267 and gives no bandwidth figure, so the comparison here stays on capacity and type.
On both boards the memory is shared with the processor, so the model and the operating system must fit together in 8 GB. Puget Systems' llama.cpp testing links generation speed to bandwidth, which is why bandwidth is worth watching for language models. Our Jetson Orin Nano Super review covers what that means for the Jetson.
On power, NVIDIA lists a 7 W to 25 W range for the kit, which ships with a 19 V adapter. Raspberry Pi specifies a 5 V/5 A supply via USB-C with Power Delivery for the Pi 5, and says the board performs best with active cooling, such as the official case with a fan or the Active Cooler. Because the two makers describe power in different ways, the figures cannot be ranked against each other.
What the Ultralytics figures show#
Ultralytics, which makes the YOLO models, publishes a Jetson guide and a Raspberry Pi guide. Both list inference times for its YOLO26n model at 640 pixels, excluding pre-processing and post-processing. Lower is better. The Jetson figures come from the Jetson guide, run with Ultralytics 8.4.33 and JetPack 6.1.
YOLO26n COCO128 imgsz 640: Inference time (ms/image, lower is better)
4 products. Lowest: Jetson Orin Nano Super, TensorRT INT8, 3.8 ms/image. Highest: Jetson Orin Nano Super, PyTorch, 15.6 ms/image.
| Product | ms/image |
|---|---|
| Jetson Orin Nano Super, TensorRT INT8 | 3.8 |
| Jetson Orin Nano Super, TensorRT FP16 | 4.57 |
| Jetson Orin Nano Super, TensorRT FP32 | 7.53 |
| Jetson Orin Nano Super, PyTorch | 15.6 |
About this data: figures are copied from the named source, not measured by us, and the test system and settings are those the source states above. See all benchmark sources and credits.
The chart shows the Orin Nano Super at 15.60 ms with PyTorch, falling to 3.80 ms with TensorRT INT8. The Pi 5 figures come from a different page, run with a newer Ultralytics version, so they sit in a table and not on the same chart.
| Board | Route | Time per image | Ultralytics page and version |
|---|---|---|---|
| Jetson Orin Nano Super | TensorRT FP16 | 4.57 ms | Jetson guide, 8.4.33 |
| Jetson Orin Nano Super | PyTorch | 15.60 ms | Jetson guide, 8.4.33 |
| Raspberry Pi 5 (CPU only) | NCNN | 67.03 ms | Raspberry Pi guide, 8.4.108 |
| Raspberry Pi 5 (CPU only) | PyTorch | 299.09 ms | Raspberry Pi guide, 8.4.108 |
On the plain PyTorch route the Jetson is roughly 19 times faster. With TensorRT FP16 against NCNN, the gap is about 15 times. Neither Ultralytics page lists a Pi with the AI HAT+ in this table, so these numbers say how the Pi's CPU compares, not how the add-on route performs. Treat the ratios as a guide to scale.
Reading the TOPS figures across both platforms
Putting 67 beside 26 invites the wrong conclusion. The Jetson figure is 67 TOPS under a sparse INT8 rating, and NVIDIA lists 33 dense INT8 TOPS for the same kit. Raspberry Pi's AI HAT+ is listed at 26 TOPS for the Hailo-8 version. The AI HAT+ 2's 40 TOPS is an INT4 figure, which is a different precision again.
So three numbers use three conventions. The honest approach is to compare published results on your kind of model, as in the table above, and to use the TOPS figures only to sort products into tiers. Our what is TOPS explainer goes through precision and sparsity in detail.
A closer look at the Raspberry Pi add-on route#
The Pi's AI story is really two stories: the host board and the add-on. The Pi 5 supplies the computer, with a 2.4 GHz quad-core Cortex-A76 CPU, 8 GB of LPDDR4X, two 4-lane MIPI connectors and the PCIe 2.0 x1 interface that the HAT+ uses. The AI HAT+ supplies the accelerator, with 13 TOPS from a Hailo-8L or 26 TOPS from a Hailo-8.
That split has consequences for a buyer. First, the add-on is a separate listing with its own seller, so you check two titles and two sets of stock. Second, Raspberry Pi lists the add-on for a Pi 5 with the Active Cooler in place, so the cooler is worth adding to the shopping list. Third, the accelerator handles only the models and formats its chip and software support, which Raspberry Pi describes through its camera software stack, TensorFlow and PyTorch.
For the Jetson, the equivalent decisions sit inside one product. The GPU, the memory and the CPU come on the module, NVIDIA's power supply comes in the box, and the same CUDA and TensorRT tools cover every model format. The cost of that integration is flexibility: you do not get the Pi's wide set of hobby connectors out of the box, and you cannot swap the AI engine for a different one.
Neither route is better in every case. A project that needs the Pi's camera connectors, GPIO header and light AI can be served well by a Pi 5 with the AI HAT+ added. A project that needs the largest model a small board can hold, run through TensorRT, is better served by the Jetson. Because there is no published Ultralytics table for the Pi with the add-on, this guide does not rank the add-on route against the Jetson on speed.
Questions to settle before ordering
- What is the model? A small YOLO-class detector runs on both platforms. A larger or newer architecture may need the Jetson's GPU and its memory bandwidth.
- Where does it run? If the board will sit in an enclosure, read each maker's listed power and cooling notes first.
- How many parts are you buying? A Pi with an AI HAT+ means a board, a cooler and an add-on. The Jetson kit includes its power supply.
- Who supports it? Both makers publish documentation and forums. Both boards reach India through third-party sellers, so check return terms.
- What comes next? If the project may grow into larger NVIDIA hardware, the Jetson's software carries over. If it may grow into a general-purpose Linux machine, the Pi is the more natural base.
Software and what each platform runs
The Jetson uses NVIDIA's JetPack with CUDA and TensorRT. For a student, that is the same software model used on larger NVIDIA hardware, and our GPU guides show where those skills lead. For a camera or robotics prototype, NVIDIA names robot development, vision AI for retail and factory inspection, and generative AI at the edge as target uses.
The Pi runs Raspberry Pi's own operating software. Its AI add-on plugs into the camera tools Raspberry Pi provides. For general electronics, the Pi 5 lists two 4-lane MIPI connectors for cameras or displays, dual 4Kp60 HDMI, two USB 3.0 ports, two USB 2.0 ports, Gigabit Ethernet and a 40-pin GPIO header.
Small vision models run on either board. Large language models do not fit in 8 GB, and neither maker positions its board for them. For those, see the VRAM explainer and the GPU guides.
Common mistakes#
- Comparing headline TOPS across platforms. The Jetson's 67 is a sparse INT8 rating, while the Pi add-ons use their own figures at INT8 and INT4.
- Forgetting that the Pi needs an add-on for AI. On the board alone, the Pi runs models on the CPU, which Ultralytics lists at 299.09 ms with PyTorch.
- Assuming 8 GB is enough for any model. Memory is shared with the operating system on both boards.
- Treating two Ultralytics pages as one test. The Jetson and Pi guides used different versions and setups.
- Ordering accessories before choosing the board. Choose the platform first, then the case, cooler and storage.
What this means when buying#
Choose by the job, not by the headline figure.
- A student learning edge AI: the Jetson. NVIDIA's CUDA and TensorRT tools are part of the kit, and the published times are low.
- A general computer with the option of AI later: the Raspberry Pi 5, with the AI HAT+ as a separate purchase if a project needs it.
- A camera or robotics prototype that must run a vision model on the board: the Jetson.
- A project where the AI is light or runs elsewhere: the Pi.
Both boards reach India through resellers, so check the seller and the exact model name. Prices change, so check current prices before buying.
- NVIDIA Jetson Orin Nano Super Developer Kit 8GB: the AI-first board, with 67 sparse INT8 TOPS and a 19 V adapter in the box.
- Raspberry Pi 5 8GB: the general-purpose host, which takes the AI HAT+ over PCIe.
For wider reading, see the AI accelerators hub.
Before ordering in India, confirm that the seller is an authorised distributor, because boards sold through unofficial channels may carry no local warranty and can arrive without the right power adapter for 230 V mains.
Frequently asked questions#
Does the Raspberry Pi 5 have an NPU?
Raspberry Pi lists none on the board. The AI HAT+ adds one over PCIe, at 13 TOPS or 26 TOPS depending on the version. Without the add-on, AI work on the Pi 5 runs on its CPU.
Is the Jetson harder to set up than a Raspberry Pi?
It involves flashing and configuring a Linux image with NVIDIA's tools. Plan time for setup and read NVIDIA's get-started guide for the kit. The Pi's setup is simpler, but the Jetson's tools suit GPU-based AI work.
Can the Pi 5 run YOLO models?
Yes. Ultralytics lists CPU results for the Pi 5, including 67.03 ms per image with NCNN for its YOLO26n model. Larger YOLO models run slower on the CPU alone, so an add-on accelerator helps for live video.
Do both boards need extra cooling?
Raspberry Pi says the Pi 5 performs best with active cooling. NVIDIA lists a 7 W to 25 W range for the Jetson kit, so airflow is sensible in a closed enclosure.
Which board runs on less power?
NVIDIA's low mode is 7 W. Raspberry Pi specifies a 5 V/5 A supply for the Pi 5 but gives no typical wattage, so the two cannot be ranked on power from maker figures alone.
Are the TOPS figures comparable?
Not directly. Precision and sparsity change what a TOPS number means, and the Jetson's 67 is a sparse figure. Compare boards using published results for your own model type, not the headline rating alone.
Sources & credits#
We have not tested these products ourselves. Specs come from the makers; benchmark figures come from the sources we name.
- NVIDIA India, Jetson Orin Nano Super Developer Kit (opens in a new tab)
- NVIDIA India, Jetson Orin series (opens in a new tab)
- NVIDIA Developer, Jetson Orin Nano Developer Kit get started (opens in a new tab)
- Raspberry Pi, Raspberry Pi 5 product page (opens in a new tab)
- Raspberry Pi, AI HAT+ product page (opens in a new tab)
- Raspberry Pi, AI HAT+ 2 product page (opens in a new tab)
- Ultralytics Docs, NVIDIA Jetson guide (opens in a new tab)
- Ultralytics Docs, Raspberry Pi guide (opens in a new tab)
- Puget Systems, NVIDIA GeForce RTX 5090 & 5080 AI Review (opens in a new tab)
Facts and figures in this article come from these sources. Articles are compiled from the makers' published specifications and the benchmarks named above, and checked by the NextAIGrid Editorial Team.
Disclaimer
This explainer is for information only. Specifications and availability change, so check the maker's page. Found an error? Tell us.
How was this explainer put together?
It draws on 2 named sources: maker specification pages and published third-party benchmarks, listed in the article. We have not tested these products ourselves. Specs come from the makers; benchmark figures come from the sources we name. Specifications and sources were last checked on 17 November 2026. Read how we research and our editorial policy.
Spotted an error? Tell us. Confirmed fixes are listed on the corrections page.