AI/rocm-systems
2
0 Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

fix max BF16 flop rate on CDNA2

Browse Source 
Signed-off-by: Nicholas Curtis <nicurtis@amd.com>
This commit is contained in:
Nicholas Curtis committed 2023-07-28 12:38:11 -04:00
1 parent 6487ba3853
commit 42248d5391
2 files changed
+3 -3
Download Patch File Download Diff File Expand all files Collapse all files

No files matched your search

src/omniperf_analyze/configs/gfx90a/0200_system-speed-of-light.yaml
+2 -2
View File
@@ -45,9 +45,9 @@ Panel Config:
MFMA FLOPs (BF16):
value: AVG(((SQ_INSTS_VALU_MFMA_MOPS_BF16 * 512) / (EndNs - BeginNs)))
unit: GFLOP
peak: ((($sclk * $numCU) * 512) / 1000)
peak: ((($sclk * $numCU) * 1024) / 1000)
pop: ((100 * AVG(((SQ_INSTS_VALU_MFMA_MOPS_BF16 * 512) / (EndNs - BeginNs))))
/ ((($sclk * $numCU) * 512) / 1000))
/ ((($sclk * $numCU) * 1024) / 1000))
tips:
MFMA FLOPs (F16):
value: AVG(((SQ_INSTS_VALU_MFMA_MOPS_F16 * 512) / (EndNs - BeginNs)))