FP16 vs BF16 vs FP8 Runtime Behavior
FP8 introduces dual 8-bit data formats (E4M3 and E5M2) to navigate extreme tradeoffs between mantissa precision and exponent dynamic range.
Source: mortalapps.com- FP8 introduces dual 8-bit data formats (E4M3 and E5M2) to navigate extreme tradeoffs between mantissa precision and exponent dynamic range.
- The core purpose of FP8 is to double arithmetic throughput and halve memory bandwidth pressure compared to 16-bit systems, avoiding the catastrophic quantization noise of INT8.
- The primary optimization strategy is the hybrid application of these formats: E4M3 for forward-pass activations/weights and E5M2 for backward-pass gradients.
- The most critical engineering insight is that neural network gradients span up to ten orders of magnitude and will fatally underflow to zero if represented in E4M3, necessitating dynamic format switching at runtime.
Why This Matters
As generative AI parameter counts scale into the trillions, cluster-level memory bandwidth and compute density become hard constraints. The transition from 16-bit (FP16/BF16) to 8-bit (FP8) floating-point arithmetic is a mandatory optimization that effectively doubles the peak teraflops (TFLOPS) of Hopper and Blackwell Tensor Cores while halving the high-bandwidth memory (HBM) footprint. Mastering FP8 runtime behavior is critical for infrastructure engineers because nave casting directly to FP8 causes immediate training instability and NaN losses, risking the destruction of multi-million-dollar distributed training runs.
Core Intuition
To understand FP8, one must visualize the strict economy of bits. Floating-point formats allocate bits to the sign, exponent (determining the dynamic range), and mantissa (determining the precision). BF16 historically solved FP16's overflow issues by borrowing bits from the mantissa to expand the exponent. FP8 pushes this tradeoff to the physical limit by offering only 8 bits total. Because forward-pass activations typically span a relatively narrow range of approximately three orders of magnitude, they require higher mantissa precision to maintain signal fidelity. Conversely, gradients propagated during the backward pass span up to ten orders of magnitude. For gradients, retaining fine precision is less important than simply capturing their massive dynamic range to prevent the values from underflowing to absolute zero.
Technical Deep Dive
The FP8 standard formalizes two distinct encoding schemes to address the forward/backward dichotomy. The E4M3 format consists of 1 sign bit, 4 exponent bits, and 3 mantissa bits. Using a bias of, its maximum representable value is clamped at 448 (reserving specific encodings for NaN and Inf), and its minimum subnormal value is 
. The E5M2 format consists of 1 sign bit, 5 exponent bits, and 2 mantissa bits. With a bias of, it doubles the exponent range at the cost of half the mantissa precision, supporting a maximum value of,344 and a minimum subnormal value of 
.
| Format | Sign Bits | Exponent Bits |
|---|---|---|
| Mantissa Bits | Max Value | Min Subnormal |
| Primary Role | FP32 |  |
 | Master Weights, Accumulators | BF16 |
|  | High-Precision Baselines |
| FP8 (E5M2) | 57,344 | |
| Backward Gradients | FP8 (E4M3) | 448 |
| Forward Activations/Weights | NVIDIA's 4th and 5th generation Tensor Cores process these formats natively, accepting E4M3 or E5M2 inputs and performing the internal dot-product accumulation in high-precision FP32 to prevent intermediary overflow. |