Intel Gaudi 3 AI Accelerator Technical Paper Table 5 - peak TFLOP/s by datatype, MME vs TPC (VERBATIM)

VERBATIM ROW, 'MME (Matrix)' column then 'TPC (Vector)' column: FP8 1678 | 57.3, BF16 1678 | 28.7, FP16 (signed) 459 | 28.7, TF32 459 | -, FP32 229 | 14.3. Independently confirms the Gaudi 3 record's two flops rows (BF16 1678, FP8 1678, both MME/matrix path) and confirms the 28.7 BF16 VECTOR figure the record's notes quote. Also supplies four figures NOT in the record (FP16 459, TF32 459, FP32 229 MME; FP8 57.3, FP32 14.3 TPC) - available if anyone wants to extend the record. Intel states no dense/sparse basis for any row here or in Table 1; 'Peak' without qualification. The Gaudi 3 record's dense:true on both rows is therefore a documented inference, not a vendor claim - the record's notes already say so, which is honest.