Single-Bit Exponent Flip in BF16 Causing Silent Gradient Divergence Across Data-Parallel Ranks

In BF16 distributed training, a single-bit flip in the exponent field of a gradient tensor can propagate silently through NCCL all-reduce without detection, causing one rank's gradient update to diverge from all others. Research shows that high-order bit flips (exponent bits 1-3) are detected with 99% accuracy using Wasserstein divergence metrics, but undetected flips cause permanent model divergence within 10-50 steps.

Single-Bit Exponent Flip in BF16 Causing Silent Gradient Divergence Across Data-Parallel Ranks is a Reliability failure seen during ML training runs. In BF16 distributed training, a single-bit flip in the exponent field of a gradient tensor can propagate silently through NCCL all-reduce without detection, causing one rank's gradient update to diverge from all others. Research shows that high-order bit flips (exponent bits 1-3) are detected with 99% accuracy using Wasserstein divergence metrics, but undetected flips cause permanent model divergence within 10-50 steps. Common tags: Sdc, Bit Flip, Bf16, Gradient Divergence.

Silicon-level single-event upset or timing fault causes a bit-flip in the high-order exponent bits of BF16 gradient value. Standard NCCL all-reduce performs arithmetic on corrupted value without checksum or validity check. Optimizer applies the corrupted gradient, permanently altering model weights in a divergent direction. Taken together, these mechanisms explain why the failure is reproducible, why it tends to surface on specific workloads or scales, and why generic mitigation attempts often fall short without addressing the underlying cause.

• Training loss divergence without hardware error signals or NaN values visible in logs
• One rank produces massively different gradients compared to other identical replicas
• Gradient norm from the affected rank differs by >100x from the all-reduced mean

• BF16/FP16 distributed training on 64+ GPUs using NCCL all-reduce
• Training without cross-rank gradient validation in the communication layer
• H100/H200 clusters operating near thermal margins where bit-flip probability increases

Use this checklist to test the hypothesis against a small reproduction. No single line proves the root cause, so preserve the preceding events and compare one variable at a time.

• ✓Reproduce the failure from a clean checkpoint/seed: the symptom must appear without warm-up state from a previous run.
• ✓Verified signal present: Per-rank gradient norm computed before all-reduce shows one rank with outlier norm (3+ sigma from mean)
• ✓Verified signal present: Wasserstein distance between per-rank gradient distributions exceeds threshold of 0.1 for BF16 tensors
• ✓Verified signal present: Loss curve plateaus at higher value than expected for the given training step count
• ✓A targeted fix from the "How to fix it" section eliminates or substantially reduces the symptom within one validation pass.

Select the traceback or open the failed terminal, then run Denpex locally to see the initiating rank, collateral failures, exact fix, and verification command without uploading the log.

• Silicon-level single-event upset or timing fault causes a bit-flip in the high-order exponent bits of BF16 gradient value
• Standard NCCL all-reduce performs arithmetic on corrupted value without checksum or validity check
• Optimizer applies the corrupted gradient, permanently altering model weights in a divergent direction