CT × CT matrix multiplication
N = 65,536 · Scale 250· Encrypted matrix × encrypted matrix
ScalingThroughput
Log Y
Workload performance
HoistingOn
Complete workload
NVIDIA RTX PRO 6000 Blackwell Max-Q Workstation EditionNVIDIA GeForce RTX 4060 TiNVIDIA RTX A6000 · 300 WNVIDIA H100 NVL · 400 WAMD Ryzen Threadripper PRO 7975WX 32-Cores
Operations
6 operationsRotate (1 offset)
NVIDIA RTX PRO 6000 Blackwell Max-Q Workstation EditionNVIDIA GeForce RTX 4060 TiNVIDIA RTX A6000 · 300 WNVIDIA H100 NVL · 400 WAMD Ryzen Threadripper PRO 7975WX 32-Cores
Rotate-many (7 offsets)
HoistingOn
NVIDIA RTX PRO 6000 Blackwell Max-Q Workstation EditionNVIDIA GeForce RTX 4060 TiNVIDIA RTX A6000 · 300 WNVIDIA H100 NVL · 400 WAMD Ryzen Threadripper PRO 7975WX 32-Cores
Multiply (CT × CT)
NVIDIA RTX PRO 6000 Blackwell Max-Q Workstation EditionNVIDIA GeForce RTX 4060 TiNVIDIA RTX A6000 · 300 WNVIDIA H100 NVL · 400 WAMD Ryzen Threadripper PRO 7975WX 32-Cores
Add (CT + CT)
NVIDIA RTX PRO 6000 Blackwell Max-Q Workstation EditionNVIDIA GeForce RTX 4060 TiNVIDIA RTX A6000 · 300 WNVIDIA H100 NVL · 400 WAMD Ryzen Threadripper PRO 7975WX 32-Cores
Relinearize
NVIDIA RTX PRO 6000 Blackwell Max-Q Workstation EditionNVIDIA GeForce RTX 4060 TiNVIDIA RTX A6000 · 300 WNVIDIA H100 NVL · 400 WAMD Ryzen Threadripper PRO 7975WX 32-Cores
Rescale (2 components)
NVIDIA RTX PRO 6000 Blackwell Max-Q Workstation EditionNVIDIA GeForce RTX 4060 TiNVIDIA RTX A6000 · 300 WNVIDIA H100 NVL · 400 WAMD Ryzen Threadripper PRO 7975WX 32-Cores
Measurement conditions
ConfigurationN = 65,536 · scale 250
ExecutionEager · prepared inputs and keys
PlatformsNVIDIA RTX PRO 6000 Blackwell Max-Q Workstation Edition / NVIDIA GeForce RTX 4060 Ti / NVIDIA RTX A6000 · 300 W / NVIDIA H100 NVL · 400 W / AMD Ryzen Threadripper PRO 7975WX 32-Cores