LattePanda Mu Ultra Puts 97 TOPS on a Compute Module

What DFRobot Put on the Module

DFRobot has launched the LattePanda Mu Ultra, an x86 compute module built around Intel's Lunar Lake mobile silicon and aimed squarely at on-device AI. The whole thing measures 69.6 x 60 mm — a shade larger than a credit card — and it is the first LattePanda module to lead with an NPU number rather than a CPU one.

• Two CPU options: Core Ultra 5 226V at up to 4.5 GHz with a 7-core Xe GPU, or Core Ultra 7 256V at up to 4.8 GHz with an 8-core Xe GPU
• 40 or 47 TOPS from Intel AI Boost, for a combined 97 or 115 TOPS across NPU, GPU and CPU
• 16GB of LPDDR5X at 8533 MT/s, giving 136.5 GB/s of memory bandwidth, soldered on the module
• $599 for the 226V module and $699 for the 256V, with developer kits at $643 and $733

The module is officially supported on Windows 11 and Ubuntu 24.04 and 26.04, and rated for 0–60°C operation on a 9–20V input.

What Do You Give Up in This Form Factor?

Two things, and both are worth naming up front. The memory is soldered at 16GB with no upgrade path, which is the trade every LPDDR5X design makes for that 136.5 GB/s of bandwidth. And there is no onboard storage beyond a 256Mbit SPI flash for the BIOS — the NVMe drive lives on the carrier board, so budget for an SSD separately.

What you get back is I/O density that would be hard to route yourself: two USB 3.2 Gen2 ports, six USB 2.0, up to four PCIe 4.0 lanes, three UARTs, two I2C buses and general-purpose GPIO, plus an eDP 1.4 connector and support for three independent displays over HDMI or DisplayPort.

DFRobot's companion carrier, the DFR1293 Mini, breaks that out into M.2 slots, 2.5Gbps Ethernet, dual USB-C with DisplayPort, an audio codec and the PCIe 4.0 expansion. That is the configuration in the developer kit, which also bundles a cooler.

Where a Lunar Lake Compute Module Fits

The interesting case for a single board computer of this class is the workload that needs x86 compatibility and a real NPU in the same box: machine vision on a production line, a kiosk running a local vision model, a robotics controller that has to talk to Windows-only vendor software. An ARM SBC will usually be cheaper; it will not always run the software the customer already owns.

At $599 the Mu Ultra is not competing with a Raspberry Pi. It is competing with the low end of the mini PC market — and losing on price there while winning on integration, because a compute module lets a product team design their own carrier and ship a custom device rather than a beige box with a fan in it. For a sense of the neighbouring price band, our Minisforum MS-S1 MAX workstation piece covers what that money buys in a finished chassis.

If the plan is running language models locally on it, the 16GB ceiling is the real constraint rather than the TOPS figure — our guide to how much RAM a local LLM needs has the sizing. More compact hardware in our mini PC coverage.

Sources: CNX Software — September 10, 2026; DFRobot product page — September 2026.