02.11.2026 at 00:58

LattePanda Sigma 2 unveiled with Core Ultra X7 and 64GB of RAM for local AI

LattePanda has unveiled the Sigma 2, a single-board server for running neural networks locally, featuring Intel Arc B390 graphics and 64GB of RAM. The 146 × 102 mm board packs two 10GbE network ports, four M.2 slots, and a connector for an external graphics card.

Performance

The Sigma 2 is built around the Intel Core Ultra X7 358H from the Panther Lake family. The processor has 16 cores: four performance cores, eight efficiency cores, and four low-power cores. It handles 16 threads, runs at up to 4.8GHz, and has 18MB of cache.

The integrated Intel Arc B390 graphics feature 12 Xe cores clocked at up to 2.5GHz. According to LattePanda, peak INT8 compute performance reaches 122 TOPS, while the dedicated neural processing unit (NPU) delivers up to 50 TOPS. Including the CPU, the total is up to 180 TOPS. This is the sum of the peak capabilities of the three compute units; the speed of a particular neural network depends on the model and software.

The board has 64GB of dual-channel LPDDR5X-8533 memory, with a theoretical bandwidth of 136.5GB/s. Up to 59GB of the total RAM can be allocated to graphics. According to the developers, this is enough for quantized MoE models with around 35 billion parameters, along with a context cache.

Storage and expansion

A 1TB WD SN7100 NVMe SSD is included. Three M.2 M-Key slots with different interfaces are available for storage and compatible AI accelerators:

PCIe 5.0×2 for 2280-format devices;
PCIe 4.0×4 for 2280-format devices;
PCIe 4.0×1 for 2230- or 2280-format devices.

The fourth slot, an M.2 E-Key for 2230-format devices, is intended for a wireless adapter and supports PCIe 4.0×1, USB 2.0, and CNVio3. The adapter and antennas are sold separately. All four slots share a 3.3V power rail with a combined current limit of 8A.

An external graphics card or other compatible PCIe device can be connected via OCuLink with a PCIe 4.0×2 interface. Two Thunderbolt 4 ports provide additional expansion options.

Software and connectivity

The SSD comes preloaded with Ubuntu 26.04, GPU and NPU drivers, the OpenVINO Model Server, the OpenClaw AI agent, and the Qwen3.6-35B-A3B model in OpenVINO format with INT4 quantization. The prepared environment lets users run a local assistant without downloading a model or connecting to a cloud AI service. An OpenAI Chat Completions-compatible API is available for connecting custom applications.

The board also supports Windows 11 and Ubuntu 24.04 LTS or later. Security is provided by a TPM 2.0 built into the processor.

External ports include:

two RJ45 ports with Realtek RTL8127 controllers, each supporting speeds of up to 10Gb/s;
two Thunderbolt 4 ports with USB Type-C connectors and speeds of up to 40Gb/s;
two USB 3.2 Gen 2 Type-A ports supporting up to 10Gb/s and two USB 2.0 Type-A ports supporting up to 480Mb/s;
HDMI 2.1 and a combined 3.5mm audio jack.

HDMI supports output at up to 4096 × 2304 pixels at 60Hz, while DisplayPort 2.1 over USB Type-C supports up to 7680 × 4320 pixels at 60Hz. The internal eDP 1.5 interface supports panels up to 3840 × 2400 pixels at 120Hz. A separate I2C connector is provided for a touch panel.

The RP2350 microcontroller, preloaded with MicroPython, works with sensors and actuators independently of the main processor. The board provides a 34-pin GPIO header, I2C, SPI, UART, and analog inputs. It also has internal connectors for USB 2.0, RS-232/485, the case front panel, an audio panel, and two fans.

Power and package contents

The Sigma 2 comes with active cooling installed, a metal protective plate on the underside, a real-time clock battery, and three thermal pads for the SSD. The package includes a 180W power supply with a 20V output and power cables for European, American, and British outlets.

Power can be supplied via the 5,5 × 2,5 mm barrel connector, the internal eight-pin DC-IN connector, or USB Type-C with USB PD. For DC input, 15-20V and an adapter rated at 120W or higher are recommended; for USB PD, 20V and at least 100W are recommended. The board also works with 45W or 65W USB PD adapters, but automatically reduces the processor’s power limits. The operating temperature range is 0-40 °C.