mruby-gpu The Joy of Taking to Hardware in Ruby RubyKaigi 2026 | Lightning Talk Yuji Teshima 

Self-introduction Yuji Teshima @yujiteshima I work at Stadium Inc. This is our first time serving as a Silver Sponsor. I’m building FANTS in Ruby. 

The Spark "Getting Started with GPU & NPU Programming on Raspberry Pi" "I want to write this in mruby." 

Yes, Raspberry Pi has a GPU too. 

Architecture → mruby mrbgem → C → Vulkan API → GPU Just write GPU.add(a, b) — under the hood, C dispatches commands to the GPU via Vulkan. 

mrbgem — the bridge between C and Ruby // src/gpu_ops.c void mrb_mruby_gpu_gem_init(mrb_state *mrb) { struct RClass *gpu = mrb_define_module(mrb, "GPU"); mrb_define_module_function(mrb, gpu, "add", mrb_gpu_add, MRB_ARGS_REQ(2)); mrb_define_module_function(mrb, gpu, "matmul", mrb_gpu_matmul, MRB_ARGS_REQ(5)); mrb_define_module_function(mrb, gpu, "relu", mrb_gpu_relu, MRB_ARGS_REQ(1)); // ... } 

# gpu_add.rb GPU.init("shader") a = GPU.array([1.0, 2.0, 3.0]) b = GPU.array([4.0, 5.0, 6.0]) c = GPU.add(a, b) puts c.head(3).inspect #=> [5.0, 7.0, 9.0] GPU.add Adding two vectors on the GPU. The very first step. 

It worked. I'm stoked. TERMINAL 

Stacking methods, one by one → MNIST # 2-layer MLP: 784 → 128 (ReLU) → 10 # Forward z1 = GPU.matmul(w1, x, 128, 784, 1) h = GPU.relu(GPU.add(z1, b1)) o = GPU.add(GPU.matmul(w2, h, 10, 128, 1), b2) # Backward grad_w2 = GPU.matmul_nt(grad_o, h, 10, 1, 128) grad_h = GPU.matmul_tn(w2, grad_o, 128, 10, 1) grad_h_pre = GPU.mul(grad_h, mask) # SGD update w1 = GPU.sub(w1, GPU.scale(grad_w1, LR)) w2 = GPU.sub(w2, GPU.scale(grad_w2, LR)) 

The wall: painfully slow. It works. But it's slow. Where's the bottleneck? # Profile one MNIST forward step — just wrap with Time.now t0 = Time.now x = GPU.load("data/train_images.bin", i * 784, 784) # CPU → GPU t1 = Time.now z1 = GPU.matmul(w1, x, 128, 784, 1) h = GPU.relu(GPU.add(z1, b1)) # compute t2 = Time.now scores = h.head(128) # GPU → CPU t3 = Time.now puts "Transfer: #{((t1 - t0) * 1000).round(1)} ms" puts "Compute: #{((t2 - t1) * 1000).round(1)} ms" puts "Readback: #{((t3 - t2) * 1000).round(1)} ms" It's mruby — just wrap it with Time.now. 

Breakthrough: Packing Send data to the GPU in batches, not one sample at a time. Make each GPU dispatch bigger — feed batched matmuls. Read back from the GPU once per batch, not per sample. TERMINAL 

Camera face detection — swap it in one line. # GPU mode detector = FaceDetector.new("models/ultraface-slim", use_gpu: true) # CPU mode — change 1 keyword detector = FaceDetector.new("models/ultraface-slim", use_gpu: false) W, H = 640, 480 cam = Camera.open("/dev/video0", W, H) disp = Display.open(W, H, "face demo") loop do break if Display.poll_quit rgb = Camera.yuyv_to_rgb(cam.capture, W, H) detector.detect_rgb(rgb, W, H, threshold: 0.6).each do |f| disp.draw_rect(rgb, W, H, f[:x], f[:y], f[:w], f[:h], 0, 255, 0) end disp.show(rgb, W, H) end No rebuild. Try it now. Compare it now. 

Result: GPU < CPU GPU mode 6 FPS Choppy. CPU mode 30 FPS Smooth. TERMINAL ## GPU: [FPS] 6.2 [FPS] 5.8 [FPS] 6.1 ## CPU: [FPS] 30.1 [FPS] 29.8 [FPS] 30.3 * Actual inference time: GPU 165ms vs CPU 12ms CPU is 14× faster, but display is capped at 30 FPS. Each layer's data was small — not the pattern GPUs are built for. 

Demo 1. Adding 1M-element arrays — GPU vs CPU 2. Camera face detection — switching between GPU and CPU 

The ideas just keep coming. Fuse two inputs: infrared + regular camera There are even thermal cameras out there, right? Run inference only where motion happens — via frame differencing Microphones, LiDAR, accelerometers… The list of input devices is endless. Too much fun! 

Let's talk to the hardware — with mruby. Come play with me. github.com/yujiteshima/mruby-gpu Thank you! Peel. See. Picture.