PROPOSAL: Latte Panda Mu

[OrsoEric]

I took a challenge to design a board for the latte panda mu compute module with N305 X86 processor and 16GB of LPDDR5 RAM.

Reason is Intel makes drivers for their iGPU, and this module is dog compatible in size and power.

Reason is I tried to find accelerators for the Pi, but it's too early for that. none of the half a dozen startup I queried answered, and Hailo is just 2GB.

Cost is likely going to be 300€ for 8GB modules when they eventually come out. Meanwhile intel mobile processors have 16GB ram and drivers for their iGPU acceleration and cost less. None of the upcoming accelerators even bother using GDDR6 or GDDR7 memory, which is baffling to me. If they stick to LPDDR5 they don't have much of an edge over APUs, if any. They can do smart L3 cache thingies but that is no sobstitute for having just faster ram. I guess it makes the PCB more expensive and there aren't good memory controller IP to choose from so they stick to LPDDR5.

Reason is I want to make a pupper with local AI intelligence to do useful applications with it

My plan is to make a robot dog board, but with the power to run locally the AI

Is anyone interested in contributing on the software side? In theory, as long as you target X86 with UHD graphics on linux, it should work on latte panda mu.

Features I'm planning:

• 18650 batteries
• Pupper motors
• Multiple ESP32s for motors and AT commands for WiFi radio
• CSI for raspicam
• Ubuntu 22
• LLM brain with audio, text and image models local
• OpenVINO
• MCP
• ROS2
• Audio array
• IMU
• USB dongle for LTE connectivity

Component	Content	Specs.
Processor	CPU	Intel®Processor N100 / Intel® Core™ i3-N305
	Cores / Threads	4-Core, 4-Thread / 8-Core, 8-Thread
	Max Turbo Frequency	3.4GHz / 3.8GHz
	Cache	6MB
	TDP	6 35W / 9 35W configurable
	Graphics	Intel® UHD Graphics
	Max Dynamic Frequency	750MHz / 1.25GHz
	Execution Units	24 / 32
Memory	Capacity	8GB LPDDR5 4800MT/s / 16GB LPDDR5 4800MT/s
	ECC Memory Support	With In-band ECC
Storage	eMMC	64GB eMMC 5.1 Speed: HS400
Expansion	Expansion Capability	Up to 9x PCle 3.0 Lane Up to 4x USB 3.2 10Gbps Up to 2x SATA 6Gbps 8x USB 2.0 4x UART 4x I2C Expandable to 64x GPIO
Display	eDP	1x eDP 1.4(onboard), up to 4K 60Hz
	HDMI	3x HDMI 2.0/DisplayPort 1.4, up to 4K 60Hz
	Max Outputs	3 Ouputs; Max Resolution 4096 x 2160@60Hz
Operating System	Microsoft Windows	Windows 10, Windows 11
	Linux	Ubuntu 22.04 & 24.04
BIOS	BIOS Info	AMI UEFI 128Mbit SPI BIOS
Power	Power Supply	DC 9-20V
	RTC Battery	3V
Environment	Operating Temperature	0~60℃
	Relative Humidity	0%~80%RH
Dimension	Form Factor	60mm x 69.6mm
Model(SKU)	DFR1146	LattePanda Mu (N100, 8GB RAM, 64GB eMMC)
	DFR1146-ENT	LattePanda Mu (N100, 8GB RAM, 64GB eMMC, with Win11 Enterprise License)
	DFR1147	LattePanda Mu (N100, 16GB RAM, 64GB eMMC)
	DFR1149	LattePanda Mu (N305, 16GB RAM, 64GB eMMC)
Compliance	Certification	FCC, CE, UKCA, RoHS, KC, Safety

[OrsoEric]

It's an idea that I have been working on for a long while

This version I made for the thesis

An upgraded version I did

Then I met MangDang and helped him with the second pupper board

Acceleration on the Pi is sad, memory bandwidth is poo poo, and I haven't even found binaries for ONNX NEON, let alone that sad excuse for an iGPU the VideoCore. It has a great x264 transcoder for the raspicam, but that's about it. There are competitor for the Pi like Rock 5B that are vastly superior in performance per dollar, and performance, but still cost 200 $ and are ARM with proprietary blob, and I see reports of mixed success accelerating things on it.

My problem is that nobody can make decent acceleration, except Nvidia and Intel. And Nvidia needs ALL the money, while Intel especially as of late with the Arc tiles has taken second place overtaking AMD in acceleration, so I'm switching SBC and using an intel atom from lattepanda. the reason I got that is because they have a challenge, and gave me a 200 $ voucher to get it. They also gave me a 500$ voucher for the board, component and assembly service.

I have been using laptop with LLM and doing lots of testing. At 40GB/s the N305 should provide quite impressive performance, and with 16GB of unified memory I should be able to run 3B and 7B models without breaking a sweat in the 10 to 40 T/s range and run an ASR model with ram to spare. This should be viable. This week I'll be powering up the latte panda install ubuntu and do preliminary tests to see if Intel does really provide acceleration on Atom.

[gmacario]

Hi @OrsoEric, that sounds like an exciting proposal!

I volunteer to help and port the FREISA-GPT stack to your new robot hardware.
However I would need some additional clarifications:

1. Where can I find a system block diagram of your new robot? I mean, what do you plan to add to the robot beyond the Latte Panda Mu which you mentioned in the title of PROPOSAL: Latte Panda Mu #86?
2. Is the robot mechanics/kinematics already decided and locked down?
3. Would you be able to build and provide me with a dedicated hardware instance for me be able to develop, integrate and test the software stack? Otherwise we should plan to go with a simulated route first
4. Do you have a rough timeplan in mind? I.e. are you looking to participate to some specific challenges, MakerFaires, PiWars, etc? Knowing which one(s) could help better scope the work for the first milestone and prioritize the efforts accordingly

I think that - once those points are clear and agreed upon - other people from the @B-AROL-O/freisa-gpt_team and/or @B-AROL-O/freisa-supporters could consider joining your project and speed up the development.

[OrsoEric]

UPDATE 2025-09-17

Despite many attempts, the Lattepanda mu N305 16GB doesn't power on. The power rails are dead. I believe I got a dead board, bummer.

I ordered an 8GB board to check if I'm doing something wrong, it'll likely arrive today.

My plan is to test LLMs on the N100 8GB and that will tell us the viability of the N100 8GB and N100 16GB and N305 16GB for local AI on a mobile robot.

As suggested by @gmacario a short term goal would be to install the Latte Panda Mu on board on a wheeled robot for the Linux Day.

Once I see that this path is viable, I'll propose a conf call to discuss the project with who'd like to contribute.

The proposal in case I succeed to power on the Mu and test LLM performance, is for you to help me port what you did for FREISA-GPT onto to X86

• Microphone array
• ASR (whisper)
• LLM as poor's man AGI
• MCP as glue
• TTS (vibevoice but I tried half a dozen and I showcased decent ones)
• ROS2 as backbone

I'll take care of making a pupper grade board for the Lattepanda Mu, idea for features is as follow

• integrated 4 mic array to detect speaker and cancel noise
• 18650 battery pack
• IMU
• Pupper motors
• Multiple ESP32
• M.2 slot with PCIe lanes for wifi and bluetooth radio
• M.2 slot with full fat 4X PCIe 3.0 for storage
• Ethernet port
• anemic USB, I plan 2X or 4X USB2.0, with no USB3.0
• NO onboard charging, I plan to charge the 18650 outside the pupper
• Mini screen for onboard stats
• I can put up to three cameras, all three makes sense, one forward, one rear, one up mounted with a fish eye lens for 360° vision

[gmacario]

for the Linux Day

FYI: Our talk to LDTO25 was accepted - please see https://github.com/B-AROL-O/openai-devpost-hackathon/issues/41 for details

[OrsoEric]

Proposal for the MCP challenge

have a repo in BAROLO of a robot

• agentic
• local
• ASR - LLM
• MCP agent connected to robot status and robot motors

Track: multimodal application

sound/image to sound/move

An MVP application doesn't use the ROS2 backbone, and perhaps single LLM to simplify things.

I imagine something like this:

"Puppy, describe me what you see."

"I see ..."

"Puppy, rotate 30° clockwise and advance 50cm"

"VROOOM"

[gmacario]

Reassigning to @OrsoEric as project lead