Work done at Embedded LLM on GPU kernel acceleration for biological and medical foundation models, benchmarked across vendors on NVIDIA H100 and AMD MI300X. Every entry has its own findings page.
NeurIPS2026 · ASCI Workshop
A five-axis deployment card (latency, throughput, energy and cost, portability, diagnostic effect) reported in the units a pathology department budgets in. Filled in on a real clinical slide on both H100 and MI300X: one card goes from 222 to 364 slides per shift at half the energy, |ΔAUC| ≤ 0.001 on four cancer cohorts, and an MSK-scale day drops from ten accelerators to six.
NeurIPS2026 · TAE Workshop
An audit of the protocol, not the model. On a pathology metastasis-detection grid, twelve defensible ways to pick the "best" fine-tuning strategy name five different winners; accuracy and calibration disagree at every label budget, and the scarce-label accuracy winner survives a seed redraw only 53% of the time. We propose a claims ledger in place of a leaderboard row.
MOML @ MIT2026 · Molecular ML Conference
Profiling shows 60.8% of TensorNet inference is memory-bound gather, scatter and element-wise work. Fusing it into single-launch Triton kernels gives 2.82× end-to-end on one A100, cutting a 1 ns trajectory from 13 h to 4.6 h (1.8 → 5.2 ns/day), with energies matching the PyTorch reference to 10⁻⁶ kcal/mol.
ICML2026 · SD4H Workshop
A fused GPU kernel folds irregular-sampling interpolation and SSM inference into a single launch, cutting end-to-end latency 35.7× and clearing the sub-50ms bedside target while improving AUROC over GRU-D.
ISCA2026 · HotInfra Workshop · Talk
A measurement study across six bio models and three GPUs: 60–80% of runtime sits outside dense compute, vendor-stack asymmetry turns a trivial bug into a 38.8× swing, and one Triton source closes the gap on both NVIDIA and AMD.
COLM2026 · DAIH Workshop
Accuracy gets a model published; speed and cost get it deployed. One Triton source, two GPU vendors and a fidelity gate make clinical models faster with the same outputs: 805 → 23 ms at the ICU bedside, half the energy per pathology slide with |ΔAUC| ≤ 0.001, 1.51× to 43.1× across six protein and genomic models, and a warning that speed measured on one GPU vendor may not hold on another.
ICPP2026 · DC4AI Workshop
Long-context genomic attention is bandwidth-bound on both H100 and MI300X. A single-source Triton kernel fuses low-rank projection and error-bounded quantization of KV state inside attention, cutting data movement up to 4.1× for 1.9×/2.2× end-to-end speedups within a stated error bound.
EurIPS2025 · SimBioChem Workshop
Profiling-driven kernel fusion folds 3–8 TensorNet operations into single GPU launches, cutting kernel launches by 67–88% for a 2.82× end-to-end speedup that turns a 13-hour MD run into 4.6 hours, with physical accuracy preserved exactly.
ACM ICS2026 · Arch4Health · TalkAn accepted talk on the fused-kernel ICU work, presenting the systems story behind sub-50ms bedside inference: the 85% preprocessing bottleneck, the single-launch fix, and the latency, throughput, and accuracy gains across NVIDIA and AMD.
MLSys2026 · Young Professionals Symposium
A library of 20+ write-once Triton kernels delivering 10–19,000× speedups for sequence alignment, k-mer indexing, and quality control, compiling through MLIR to run identically on NVIDIA and AMD with no source changes.
RECOMB2026 · ARCH
Per-model drop-in Triton kernels accelerate protein and genomic foundation models plus classical algorithms (Smith-Waterman, Needleman-Wunsch, k-mer indexing, BWT) up to 2,986× across NVIDIA and AMD, with correctness preserved to machine precision.
ISC HPC2026 · Hamburg
One portable Triton fusion framework accelerates six biological models and algorithms (DualBind, AlphaGenome, Enformer, ESM-2, ProtBERT, Needleman-Wunsch) up to 720× across NVIDIA and AMD GPUs, with zero accuracy loss.