LEVI - KERNEL RITUS

Traditional GPU libraries like cuBLAS are optimized for massive workloads - think training giant neural networks with thousands of parameters. But what happens when you need to multiply small matrices? You get:

• Setup overhead that takes longer than the actual computation

• Memory allocation optimized for gigabytes, not kilobytes

• Complex dispatch logic that assumes large-scale parallelism

• Enterprise features you don't need for edge computing

• ​

Result: A Ferrari stuck in city traffic.

​

LEVI GPU Library fills the gap between "basic code" and "industrial-strength cuBLAS" with adaptive kernel selection that picks the right tool for each job.

​

The Sweet Spot: LEVI dominates in the 64-256 range where edge computing lives.

​

LEVI uses proprietary Kernel Ritus technology to automatically select optimal algorithms:

​

Two-Kernel Architecture:

// Simple Kernel (small matrices)

• Minimal setup overhead

• Cache-optimized access patterns

• Loop unrolling for better IPC

• Perfect for edge devices
   

// Tiled Kernel (medium+ matrices)

• Shared memory utilization

• Bank conflict avoidance

• Optimized for throughput

• Competitive with cuBLAS
  ​

Perfect For Edge Computing

Why Edge Needs Different Optimization:

Traditional Data Centers:

• Batch size: 1024+ samples

• Matrix size: 2048×2048+

• Memory: Abundant

• Power: Unlimited
   

Edge Computing:

• Batch size: 1-32 samples

• Matrix size: 64-512×64-512

• Memory: Limited

• Power: Battery-constrained
   

LEVI targets exactly this gap.

​

When to Use LEVI vs cuBLAS:

Use LEVI when:

✅ Matrix size < 512×512

✅ Edge/mobile deployment

✅ Power/memory constraints

✅ Batch processing many small problems

Use cuBLAS when:

✅ Matrix size > 512×512

✅ Data center deployment

✅ Maximum absolute throughput needed

✅ Deep learning training

​

More Information on our GitHub Repository!

https://github.com/forgottenforge/levi-gpu