SpacemiT K3 AI Computing Guide

SpacemiT K3 AI Computing Guide

K3 Overview

• CPU: 8x X100 @ 2.4 GHz (4-issue OoO) — 130K DMIPS
• AI Cores: 8x A100 — 60 TOPS
• Memory: LPDDR5 up to 32GB (51 GB/s bandwidth)
• LLM: Runs 30B parameter models locally (>10 tokens/s)
• Display: Dual 4K@60
• Video: 4K@180fps decode, 4K@90fps encode
• Camera: 4x MIPI-CSI (12 cameras)
• Security: SM2/3/4, AES, SHA, RSA, Secure Boot
• Virtualization: RVH 1.0, AIA, IOMMU
• OS: Ubuntu 26.04 LTS, Bianbu OS
• TDP: 15-25W
• Compatibility: Jetson Orin Nano pin-compatible

Key Innovation: RVA23 Compliance

The K3 is the world's first RVA23-compliant RISC-V processor. RVA23 is the latest application processor profile from RISC-V International, ensuring:

• Full software ecosystem compatibility
• Standard vector (V) and vector crypto extensions
• Hypervisor (H) extension for virtualization
• Standard API for all major OS and compilers

Setting Up Ubuntu 26.04

# Flash Ubuntu image to NVMe
sudo dd if=ubuntu-26.04-k3.img of=/dev/nvme0n1 bs=4M
sync

# First boot configuration
sudo apt update && sudo apt upgrade -y
sudo apt install python3-pip cmake ninja-build

Running LLM Inference

Ollama on K3

# Install Ollama (RISC-V native)
curl -fsSL https://ollama.com/install.sh | sh

# Run a 7B model (very fast on K3)
ollama run llama3.2:7b

# Run a 13B model
ollama run llama3.2:13b

# Run a 30B model (K3 can handle this!)
ollama run qwen2.5:32b
# Expected: >10 tokens/s

Python AI Pipeline

import k3_ai as ai
import numpy as np

# Initialize AI engine
engine = ai.Engine(device='a100')

# Load YOLOv8 model
model = ai.Model('yolov8s.kmodel')

# Real-time detection from camera
cap = ai.Camera(0, resolution=(1920, 1080))
while True:
    frame = cap.read()
    detections = model.infer(frame, threshold=0.5)
    
    for det in detections:
        print(f"{det.label}: {det.confidence:.2f} at {det.bbox}")
        frame = ai.draw_bbox(frame, det)
    
    ai.imshow(frame)

NPU Acceleration (60 TOPS)

import k3_npu

# Allocate NPU buffer
buf = k3_npu.Buffer(shape=(1, 3, 640, 640), dtype='float16')

# Run inference on A100 cores
result = k3_npu.infer(
    model='yolov8s.bin',
    input=buf,
    cores=8,  # Use all 8 A100 cores
    precision='fp16'
)
print(f"Inference time: {result.latency_ms:.1f} ms")
print(f"Throughput: {1000/result.latency_ms:.1f} FPS")

Jetson Orin Nano Compatibility

The K3 is pin-compatible with NVIDIA Jetson Orin Nano:

Migration: Existing Jetson Orin Nano carrier boards work with K3 SoM. Software requires recompilation for RISC-V.

ROS2 Robot Integration

# Install ROS2 on K3 Ubuntu
sudo apt install ros-humble-desktop

# ROS2 node with K3 AI
ros2 run k3_vision yolov8_detector --ros-args -p device:=0

Virtualization (RVH 1.0)

# K3 supports hardware virtualization
sudo apt install qemu-system-riscv64

# Run VM
qemu-system-riscv64 -machine virt -m 4G -smp 4   -drive file=vm.img,format=qcow2   -bios /usr/lib/riscv64-linux-gnu/opensbi/generic/fw_jump.bin

Resources

• K3 Product Brief
• Bianbu OS Downloads
• K3 Pico-ITX Documentation

Buy K3 Development Kit

Request K3 evaluation board

---