From a119b88a5687e3bf2b98ed2d28028dd73d1e2348 Mon Sep 17 00:00:00 2001 From: "Meng, Hengyu" Date: Thu, 7 Nov 2024 03:22:36 +0000 Subject: [PATCH] benchdnn: doc: add newly add f8_e4m3 and f8_e5m2 --- tests/benchdnn/doc/knobs_dt.md | 2 ++ 1 file changed, 2 insertions(+) diff --git a/tests/benchdnn/doc/knobs_dt.md b/tests/benchdnn/doc/knobs_dt.md index 0a0efdbdbbf..ec54f6555be 100644 --- a/tests/benchdnn/doc/knobs_dt.md +++ b/tests/benchdnn/doc/knobs_dt.md @@ -10,6 +10,8 @@ The following data types are supported: | s32 | standard int or int32_t | s8 | standard char or int8_t | u8 | standard unsigned char or uint8_t +| f8_e4m3 | 1-byte float (1 sign bit, 4 exp bits, 3 mantissa bits). Only reference support on CPU and GPU +| f8_e5m2 | 1-byte float (1 sign bit, 5 exp bits, 2 mantissa bits). Optimized implementation on Intel(R) Data Center GPU Max Series only. Only reference support on CPU. | f16 | 2-byte float (1 sign bit, 5 exp bits, 10 mantissa bits) | bf16 | 2-byte float (1 sign bit, 8 exp bits, 7 mantissa bits) | f64 | double precision float