7+ Main precision formats used in AI: Precision is very important in AI as it shapes how accurate and efficient models are. It controls how finely numbers are represented, approximating real-world values with formats like fixed-point and floating-point. A recent BF16 → FP16 study renewed attention to precision impact. Here are the main precision types used in AI, from full precision for training to ultra-low precision for inference: 1. FP32 (Float32): Standard full-precision float used in most training: 1 sign bit, 8 exponent bits, 23 mantissa bits. Default for backward-compatible training and baseline numerical stability 2. FP16 (Float16) → https://arxiv.org/abs/2305.10947v6 Half-precision float. It balances accuracy and efficiency. 1 sign bit, 5 exponent bits, 10 mantissa bits. Common on NVIDIA Tensor Cores and mixed-precision setups. There’s now a new wave of using it in reinforcement learning: https://www.turingpost.com/p/fp16 3. BF16 (BFloat16) → https://cloud.google.com/blog/products/ai-machine-learning/bfloat16-the-secret-to-high-performance-on-cloud-tpus Same dynamic range as FP32 but fewer mantissa bits: 1 sign bit, 8 exponent bits (same as FP32), 7 mantissa bits. It was developed by the research group Google Brain as part of their AI/ML infrastructure work at Google. Preferred on TPUs and modern GPUs 4. FP8 (E4M3 / E5M2) → https://proceedings.neurips.cc/paper_files/paper/2018/file/335d3d1cd7ef05ec77714a215134914c-Paper.pdf Emerging standard for training and inference on NVIDIA Hopper (H100) and Blackwell (B200) tensor cores and AMD MI300. Also supported in NVIDIA’s Transformer Engine: https://developer.nvidia.com/blog/floating-point-8-an-introduction-to-efficient-lower-precision-ai-training/ E4M3 = 4 exponent, 3 mantissa bits E5M2 = 5 exponent, 2 mantissa bits Read further below ⬇️ If you like this, also subscribe to the Turing post: https://www.turingpost.com/subscribe Kseniase 1 day ago FP4 → https://arxiv.org/abs/2310.16836 (4-bit Transformer); https://arxiv.org/abs/2305.14314 (QLoRA) Experimental format for ultra-compact inference. It's used in research and quantization-aware inference, including 4-Bit Floating-Point Quantized Transformers and 4-bit NormalFloat (NF4) in QLoRA INT8/INT4 → https://arxiv.org/abs/2004.09602 Integer low-precision formats that use 8 or 4 bits. Primary used in inference. The model's weights and activations are converted into integer values that can be processed efficiently on hardware optimized for integer arithmetic 2-bit (ternary or binary quantization) → https://research.ibm.com/blog/low-precision-computing Experimental ultra-low precision for computation in ultra-efficient AI accelerators. Uses values like {-1, 0, 1}. It turns multiplications into additions/subtractions - extremely cheap operations
