Hardware-Portable Fused GPU Kernels

Hardware-algorithm co-design is becoming critical for computational molecular biology, because modern workloads stress memory bandwidth, parallel execution and heterogeneous hardware. Despite fast progress in protein and genomic foundation models, practical performance is held back by GPU execution: 60–80% of runtime goes to memory-bound operations, fragmented kernel launches and repeated device transfers, and utilization worsens as sequences and batches grow.

Each model is optimized independently with custom fused Triton kernels tailored to its architecture: several PyTorch operations become one kernel, improving memory reuse and cutting launch overhead. The kernels are per-model drop-in replacements, not one monolithic framework, and run unmodified on NVIDIA and AMD with numerical correctness preserved to machine precision.

1Foundation models

Protein and genomic models from Meta, Google DeepMind, RostLab, the Baker lab and NVIDIA, each accelerated with no loss in fidelity.

2Classical bioinformatics

The same per-kernel approach accelerates classical algorithms against their established native baselines, showing the method generalizes well past neural networks.

3Portable by construction

Every kernel compiles and runs unmodified on NVIDIA (A100, H100, RTX) and AMD (MI250, MI300X). The gains come from removing memory round-trips, not from vendor-specific intrinsics, so performance portability holds without source changes, and correctness is preserved to machine precision on both backends.

Domain-specific kernel design, applied per model and per algorithm, accelerates both modern foundation models and classical bioinformatics on heterogeneous hardware, up to 2,986×, portably across NVIDIA and AMD.

More research

Part of a broader line of work at Embedded LLM on portable, fused GPU kernels for biological and clinical foundation models, measured on both NVIDIA and AMD.

• NeurIPS 2026Change the Metric, Change the Winner: Auditing How We Decide Which Clinical Foundation Model to Deploy in Cancer Pathology
• MOML 2026From 13 Hours to 4.6: Profiling-Driven Kernel Fusion for TensorNet in Molecular Simulation
• COLM 2026Deploying Clinical Language and Vision-Language Models Where the Data Lives
• ICML 2026From 805 ms to 23 ms: Accelerating State-Space Models for Real-Time ICU Monitoring
• ICPP 2026Error-Bounded Fused Attention Compression for Long-Context Genomic Foundation Models Across Heterogeneous GPUs
• ISCA 2026When the LLM-Tuned Stack Misses: An Infrastructure View of Biological Foundation Model Inference Across NVIDIA and AMD
• MLSys 2026BioTriton: Portable Cross-Vendor GPU Kernels for High-Throughput Bioinformatics via OpenAI Triton
• ISC HPC 2026Portable GPU Kernel Acceleration for Biological Foundation Models & Algorithms using OpenAI Triton

All research →