Mixed Precision Training

The core mechanism stores model weights in FP32 master copies while performing forward and backward propagation in FP16. This halves memory bandwidth requirements and leverages Tensor Core acceleration on NVIDIA GPUs. For genomic sequence models processing 100k+ token contexts, this enables 2-3x throughput improvements without architectural changes. A loss scaling factor is applied during backpropagation to prevent gradient underflow in small-magnitude updates common in DNA embedding layers.

Brain Floating Point (BFLOAT16) preserves the same 8-bit exponent range as FP32 while truncating the mantissa to 7 bits. This is critical for genomic foundation models because it maintains the dynamic range needed for attention score stability across long DNA sequences without requiring loss scaling. BFLOAT16 is natively supported on NVIDIA A100/H100 GPUs and Google TPUs, making it the preferred format for training large-scale DNA language models like Enformer and HyenaDNA.

AMP dynamically selects precision per-operation at runtime. Operations safe in lower precision—such as matrix multiplications and convolutions—execute in FP16, while precision-sensitive operations like softmax, layer normalization, and loss computation remain in FP32. Frameworks including PyTorch AMP (torch.cuda.amp) and TensorFlow mixed_float16 provide context managers that handle casting and gradient scaling automatically, requiring minimal code changes to existing genomic training pipelines.

When using FP16, small gradient values in genomic models—particularly in variant effect prediction and single-cell embedding layers—can underflow to zero. Dynamic loss scaling multiplies the loss by a large factor before backpropagation, then unscales gradients before weight updates. Modern implementations automatically adjust the scale factor: doubling it after successful iterations and halving it when NaN or Inf gradients are detected. This ensures stable convergence for architectures with deep attention stacks processing nucleotide-level features.

NVIDIA Tensor Cores perform fused multiply-add operations on FP16 matrices in a single clock cycle, delivering up to 312 TFLOPS on H100 GPUs. Mixed precision training is specifically designed to exploit this hardware. For genomic workloads like transformer-based variant calling or attention over 200k nucleotide contexts, Tensor Cores accelerate the matrix multiplications dominating self-attention and feed-forward layers. This hardware-software co-design is essential for training billion-parameter genomic foundation models within reasonable time and cost budgets.

Mixed precision combines multiplicatively with ZeRO optimization and distributed data parallelism. FP16 halves the memory per parameter, while ZeRO-3 partitions optimizer states and gradients across GPUs. Together, they enable training multi-billion parameter genomic models on commodity GPU clusters. For example, a 2.5B parameter DNA language model that would require 40GB per GPU in FP32 can fit in under 10GB with FP16 + ZeRO-3, making distributed training across 8x A100 nodes economically viable for research institutions.