MXFP8
An 8-bit variant of the Microscaling floating-point format defined by the OCP Microscaling Formats specification, jointly promoted by AMD, Intel, NVIDIA, Microsoft, and others. It achieves fine-grained quantization through block-level shared scaling factors and comes in two variants: E4M3 and E5M2.
Core Facts
The MXFP4 format has a group of elements (typically 32) sharing a single 8-bit scaling factor, with each element represented by a 4-bit micro floating-point number (1 sign bit + 2 exponent bits + 1 mantissa bit).
MXFP4 (Microscaling Float Point 4-bit) is part of the OCP MX standard jointly promoted by Microsoft, NVIDIA, AMD, and other companies.
Timeline (last 90 days)
In this test, when reading a 2048-token prompt, FP8 prefill speed was 10,118 tok/s, ahead of BF16 (7,003 tok/s), NVFP4 (6,978 tok/s), MXFP8 (6,061 tok/s) and GGUF Q5_K_M (1,710 tok/s).
In the token-by-token fidelity test, NVFP4 had an 89%–90% probability of choosing the same next token as BF16, while FP8 and MXFP8 reached 94%–96%.
In the GSM8K test (1319 questions, 8192-token limit), BF16 scored 1249, FP8 scored 1241, MXFP8 scored 1250, and NVFP4 scored 1238, with the three quantized formats matching BF16 within a ±3 point tolerance.
NVIDIA released the BF16, NVFP4, MXFP8 and GGUF versions of Nemotron Labs Dory 1.0 4B on October 16.
MXFP8 is a microscaling 8-bit floating-point format proposed by the OCP, in which a group of several elements shares a single scaling factor.
SkyRL's Day-0 support unifies LoRA, full BF16 fine-tuning, and MXFP8 reinforcement learning into vLLM inference and the Megatron-Bridge training framework.
NVFP4 is a block-scaled 4-bit floating-point format defined by NVIDIA, where every 16 numbers share one FP8 scale; MXFP8 is an 8-bit block floating point defined by OCP, where every 32 numbers share one E8M0 scale.
MXFP8 is an eight-bit format with shared block scaling (one scaling factor shared per 32 elements), which in Olmo-core 3 reduces GEMM time, saved activations, and dispatch payload, while retaining an authoritative copy of the FP32 master weights.
In the 1.2-trillion-parameter configuration, combined with MXFP8 and per-layer recomputation, Olmo-core 3 reached 858 useful-model TFLOP/s per GPU.
MXFP8 is a microscaling FP8 format defined by OCP, requiring a compatible stack such as torchao; on Blackwell it can be used to handle aspect buckets.
8 more timeline events
All Facts (19)
The MXFP4 format has a group of elements (typically 32) sharing a single 8-bit scaling factor, with each element represented by a 4-bit micro floating-point number (1 sign bit + 2 exponent bits + 1 mantissa bit).
65%VerifiedMXFP4 (Microscaling Float Point 4-bit) is part of the OCP MX standard jointly promoted by Microsoft, NVIDIA, AMD, and other companies.
65%UnverifiedIn the token-by-token fidelity test, NVFP4 had an 89%–90% probability of choosing the same next token as BF16, while FP8 and MXFP8 reached 94%–96%.
50%UnverifiedIn this test, when reading a 2048-token prompt, FP8 prefill speed was 10,118 tok/s, ahead of BF16 (7,003 tok/s), NVFP4 (6,978 tok/s), MXFP8 (6,061 tok/s) and GGUF Q5_K_M (1,710 tok/s).
50%UnverifiedIn the GSM8K test (1319 questions, 8192-token limit), BF16 scored 1249, FP8 scored 1241, MXFP8 scored 1250, and NVFP4 scored 1238, with the three quantized formats matching BF16 within a ±3 point tolerance.
50%UnverifiedMXFP8 is a microscaling 8-bit floating-point format proposed by the OCP, in which a group of several elements shares a single scaling factor.
50%UnverifiedSkyRL's Day-0 support unifies LoRA, full BF16 fine-tuning, and MXFP8 reinforcement learning into vLLM inference and the Megatron-Bridge training framework.
50%UnverifiedNVFP4 is a block-scaled 4-bit floating-point format defined by NVIDIA, where every 16 numbers share one FP8 scale; MXFP8 is an 8-bit block floating point defined by OCP, where every 32 numbers share one E8M0 scale.
50%UnverifiedIn the 1.2-trillion-parameter configuration, combined with MXFP8 and per-layer recomputation, Olmo-core 3 reached 858 useful-model TFLOP/s per GPU.
50%UnverifiedMXFP8 is an eight-bit format with shared block scaling (one scaling factor shared per 32 elements), which in Olmo-core 3 reduces GEMM time, saved activations, and dispatch payload, while retaining an authoritative copy of the FP32 master weights.
50%UnverifiedMXFP8 is a microscaling FP8 format defined by OCP, requiring a compatible stack such as torchao; on Blackwell it can be used to handle aspect buckets.
50%Unverifiedv0.30.0 simplifies the execution flow related to KDA, AttnRes, and MLA, and adds MXFP8 KV cache and asynchronous Engram mechanisms for some models.
50%UnverifiedLow-precision FlashAttention-4 (FA4 MX8) builds on FlashAttention-4 by fully migrating both forward and backward propagation to the MXFP8 microscaling low-precision format.
50%UnverifiedMXFP8 has two variants: E4M3 (4-bit exponent, 3-bit mantissa) and E5M2 (5-bit exponent, 2-bit mantissa).
50%UnverifiedCompared to tensor-level scaling, block-level scaling can more granularly adapt to differences in numerical distribution within a tensor, and offers better numerical stability in the gradient path.
50%UnverifiedMXFP8 belongs to the microscaling floating-point format family, whose core concept is sharing a single scaling factor within a small block of data.
50%UnverifiedThe NVFP4 version of Qwen3-27B uses sensitivity-aware mixed precision: feed-forward layers run in FP4 (E2M1) with FP8 block scaling, while attention projections, DeltaNet recurrent layers, the vision encoder, and multi-token prediction heads remain in FP8 or BF16.
50%UnverifiedMXFP4已获得NVIDIA Blackwell架构GPU的硬件原生支持
50%UnverifiedNVIDIA released the BF16, NVFP4, MXFP8 and GGUF versions of Nemotron Labs Dory 1.0 4B on October 16.
50%