CUDA

	TECHNICAL SPECIFICATIONS
	Tesla A30 на THEOR4 ( Micro-architecture Ampere GA100 ) Peak FP64: 5.2 TF Peak FP64: Tensor Core 10.3 TF Peak FP32: 10.3 TF TF32 Tensor Core: 82 TF \| 165 TF* BFLOAT16 Tensor Core: 165 TF \| 330 TF* Peak FP16 Tensor Core: 165 TF \| 330 TF* Peak INT8 Tensor Core: 330 TOPS \| 661 TOPS* Peak INT4 Tensor Core: 661 TOPS \| 1321 TOPS* Media engines: 1 optical flow accelerator (OFA) 1 JPEG decoder (NVJPEG) 4 Video decoders (NVDEC) GPU Memory 24GB HBM2 GPU Memory Bandwidth 933GB/s Interconnect PCIe Gen4: 64GB/s Third-gen NVIDIA® NVLINK® 200GB/s** Form Factor 2-slot, full height, full length (FHFL) Max thermal design power (TDP) 165W Multi-Instance GPU (MIG): 4 MIGs @ 6GB each 2 MIGs @ 12GB each 1 MIGs @ 24GB Virtual GPU (vGPU) software support: NVIDIA AI Enterprise NVIDIA Virtual Compute Server
	Tesla P100 на THEOR3 ( Micro-architecture Pascal GP100 )
	RTX A2000 на i9a ( Micro-architecture Ampere GA106 ) Form factor → PCIe x16 form factor # OF CUDA CORES → 3328 # OF Tensor cores → 104 CUDA compute capability → 7.5 Frequency of cuda cores → up to 1.2 GHz Double precision floating point performance (peak) → 249 Gflops Single precision floating point performance (peak) → 7.987 Tflops Total dedicated memory → 12GB GDDR6* Memory speed → 1.5 GHz Memory interface → 192-bit Memory bandwidth → 288 GB/sec Power consumption → 70W TDP System interface → PCIe x16
	Tesla C2075 на Theor2 ( Micro-architecture Fermi GF100 ) Form factor → 9.75. PCIe x16 form factor Number of CUDA cores → 448 Frequency of CUDA cores → 1.15 GHz Double precision floating point performance (peak) → 515 Gflops Single precision floating point performance (peak) → 1.03 Tflops Total dedicated memory → 6GB GDDR5* Memory speed → 1.5 GHz Memory interface → 384-bit Memory bandwidth → 144 GB/sec Power consumption → 225W TDP System interface → PCIe x16 Gen2 Thermal solution → Active Fansink Display support → Dual-Link DVI-I: 1 → Maximum Display Resolution 1600x1200

TECHNICAL SPECIFICATIONS

Tesla A30 на THEOR4 ( Micro-architecture Ampere GA100 )

Peak FP64: 5.2 TF
Peak FP64: Tensor Core 10.3 TF
Peak FP32: 10.3 TF
TF32 Tensor Core: 82 TF | 165 TF*
BFLOAT16 Tensor Core: 165 TF | 330 TF*
Peak FP16 Tensor Core: 165 TF | 330 TF*
Peak INT8 Tensor Core: 330 TOPS | 661 TOPS*
Peak INT4 Tensor Core: 661 TOPS | 1321 TOPS*
Media engines:
1 optical flow accelerator (OFA)
1 JPEG decoder (NVJPEG)
4 Video decoders (NVDEC)
GPU Memory 24GB HBM2
GPU Memory Bandwidth 933GB/s
Interconnect PCIe Gen4: 64GB/s
Third-gen NVIDIA® NVLINK® 200GB/s**
Form Factor 2-slot, full height, full length (FHFL)
Max thermal design power (TDP) 165W Multi-Instance GPU (MIG): 4 MIGs @ 6GB each 2 MIGs @ 12GB each 1 MIGs @ 24GB
Virtual GPU (vGPU) software support: NVIDIA AI Enterprise NVIDIA Virtual Compute Server

Tesla P100 на THEOR3 ( Micro-architecture Pascal GP100 )

RTX A2000 на i9a ( Micro-architecture Ampere GA106 )

Form factor → PCIe x16 form factor
# OF CUDA CORES → 3328
# OF Tensor cores → 104
CUDA compute capability → 7.5
Frequency of cuda cores → up to 1.2 GHz
Double precision floating point performance (peak) → 249 Gflops
Single precision floating point performance (peak) → 7.987 Tflops
Total dedicated memory → 12GB GDDR6*
Memory speed → 1.5 GHz
Memory interface → 192-bit
Memory bandwidth → 288 GB/sec
Power consumption → 70W TDP
System interface → PCIe x16

Tesla C2075 на Theor2 ( Micro-architecture Fermi GF100 )

Form factor → 9.75. PCIe x16 form factor
Number of CUDA cores → 448
Frequency of CUDA cores → 1.15 GHz
Double precision floating point performance (peak) → 515 Gflops
Single precision floating point performance (peak) → 1.03 Tflops
Total dedicated memory → 6GB GDDR5*
Memory speed → 1.5 GHz
Memory interface → 384-bit
Memory bandwidth → 144 GB/sec
Power consumption → 225W TDP
System interface → PCIe x16 Gen2
Thermal solution → Active Fansink
Display support → Dual-Link DVI-I: 1 → Maximum Display Resolution 1600x1200

theor2:> maple test_cuda.mpl |\^/| Maple 16 (X86 64 LINUX) ._|\| |/|_. Copyright (c) Maplesoft, a division of Waterloo Maple Inc. 2012 \ MAPLE / All rights reserved. Maple is a trademark of <____ ____> Waterloo Maple Inc. | Type ? for help. > CUDA:-IsEnabled(); false > CUDA:-Enable(true); false > CUDA:-IsEnabled(); true > > CUDA:-HasDoubleSupport(); table([0 = true]) > > with(LinearAlgebra): > M:=RandomMatrix(4000,outputoptions=[datatype=float[4]]); [ 4000 x 4000 Matrix ] M := [ Data Type: float[4] ] [ Storage: rectangular ] [ Order: Fortran_order ] > N:=RandomMatrix(4000,outputoptions=[datatype=float[4]]); memory used=124.1MB, alloc=126.0MB, time=0.88 [ 4000 x 4000 Matrix ] N := [ Data Type: float[4] ] [ Storage: rectangular ] [ Order: Fortran_order ] > > time[real](MatrixMatrixMultiply(M,N)); memory used=185.2MB, alloc=187.1MB, time=0.92 0.617 > CUDA:-Enable(false); true > time[real](MatrixMatrixMultiply(M,N)); 5.623 > > > CUDA:-Enable(true); false > M:=RandomMatrix(4000,outputoptions=[datatype=float[8]]); memory used=368.4MB, alloc=248.1MB, time=7.48 [ 4000 x 4000 Matrix ] M := [ Data Type: float[8] ] [ Storage: rectangular ] [ Order: Fortran_order ] > N:=RandomMatrix(4000,outputoptions=[datatype=float[8]]); memory used=490.6MB, alloc=370.2MB, time=7.88 [ 4000 x 4000 Matrix ] N := [ Data Type: float[8] ] [ Storage: rectangular ] [ Order: Fortran_order ] > > time[real](MatrixMatrixMultiply(M,N)); 1.640 > > CUDA:-Enable(false); true > > time[real](MatrixMatrixMultiply(M,N)); 10.614 > > CUDA:-Properties(); [table(["Max Threads Dimensions" = [1024, 1024, 64], "Clock Rate" = 1147000, "Max Grid Size" = [65535, 65535, 65535], "Memory Pitch" = 2147483647, "Max Threads Per Block" = 1024, "Warp Size" = 32, "Kernel Exec Timeout Enabled" = false, "Resisters Per Block" = 32768, "ID" = 0, "Texture Alignment" = 512, "Minor" = 0, "MultiProcessor Count" = 14, "Shared Memory Per Block" = 49152, "Total Global Memory" = 4294967295, "Major" = 2, "Name" = "Tesla C2075", "Total Constant Memory" = 65536, "Device Overlap" = 1 ])] > quit memory used=734.8MB, alloc=614.3MB, time=20.10

CUDA

Технология вычислений на графических процессорах Nvidia

Введение

Аппаратное обеспечение

Программное обеспечение

Производительность

Пример для Maple

Источники информации