Portable GPU Kernel Acceleration for Biological Foundation Models & Algorithms

Biological foundation models span DNA alignment, protein language models, genomics and molecular prediction. Their performance is limited not by model design but by GPU execution: fragmented launches, repeated memory transfers and memory-bound operations dominate runtime, and the problem grows with sequence length and batch size.

This work presents one portable Triton acceleration framework across all of these domains. Profiling shows 60–80% of runtime in memory-bound and index-heavy operations, addressed by fusing 3 to 8 operations into single kernels. That cuts memory traffic by 67–88%, lowers launch overhead and raises utilization, with 100% accuracy (or near-identical outputs) on every workload. The same source runs on NVIDIA and AMD without modification.

1Results across six workloads

Rather than optimizing one model, the framework shows that a single portable approach transfers across very different computational patterns, from dynamic-programming alignment to attention-based genomic models.

2One framework, two vendors

The unifying method is portable kernel fusion: the same Triton source compiles natively for NVIDIA (CUDA) and AMD (ROCm) with no source changes. Because the speedups come from eliminating memory round-trips rather than vendor-specific tricks, they hold across both ecosystems and across workloads as different as binding-affinity prediction, genomic attention and dynamic-programming alignment.

Biological foundation models are bottlenecked by memory traffic, not model design, and one portable Triton fusion framework accelerates DNA, protein, genomic and molecular workloads up to 720× across NVIDIA and AMD with no accuracy loss.

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
• RECOMB 2026Hardware-Portable Fused GPU Kernels for High-Throughput Biological Foundation Models

All research →