Read this article at https://www.pugetsystems.com/guides/832
NVIDIA Titan GPUs (3 generations) - CUDA 8 rc performance on Ubuntu 16.04
Written on August 12, 2016 by Dr Donald KinghornWe've got a quick test of CUDA performance on three generations of NVIDIA's Titan X GPU's for you. NVIDA released the Pascal GeForce Titan X much earlier than people were expecting (including NVIDIA :-). Just like the older Titan cards the new Pascal based card did not disappoint! For comparison we have a GTX 1080 in the mix too.
This is just a brief look at performance using the nbody code from the CUDA samples and a Molecular Dynamics simulation (stmv)using NAMD.
The short story is -- the Titan X (Pascal) is an amazing video card and CUDA compute performance is stunning!
The setup for this testing is Ubuntu 16.04 with CUDA 8.0 RC ( as of the date of this writing you need to be an NVIDIA registered developer to access the release candidate. ) The NVIDIA display driver is 367.35 from the "Graphics-Drivers ppa"
Base systems
- The Peak Tower Single
- CPU: Intel Core-i7 6900K 8-core @ 3.2GHz (3.5GHz All-Core-Turbo)
- Memory: 64 GB DDR4 2133MHz Reg ECC
- PCIe: (4) X16-X16 v3
- The Peak Tower Quad
- CPU: (4) Intel Xeon E7 8867v3 16-core @ 2.5GHz (2.7GHz All-Core-Turbo)
- Memory: 512 GB DDR4 2133GHz Reg ECC
- PCIe: (4) X16-X16 v3
Note: The Quad was used for NAMD for two reasons: 1) I had it on the bench and 2) The GPU's are so fast that I wanted lots of CPU performance so we could see the difference between the GPU's! (I'm also using some older v3 CPU's we use E5/E7 v4 in our current Quad builds)
Video cards used for testing. ( data from nvidia-smi )
| Card | CUDA cores | GPU clock MHz | Memory clock MHz* | Application clock MHz** | FB Memory MiB |
|---|---|---|---|---|---|
| Titan X (Pascal) | 3584 | 1911 | 5005 | 1417 | 12186 |
| TITAN X (Maxwell) | 3072 | 1392 | 3505 | 1000 | 12206 |
| Titan Black | 2880 | 1202 | 3500 | 888 | 6082 |
| GTX 1080 | 2560 | 1911 | 5005 | 1607 | 8113 |
Results
| Card | nbody single precision GFLOP/s | NAMD run time (sec) | NAMD day/ns |
|---|---|---|---|
| Titan X (Pascal) | 7507 | 41 | 0.570 |
| TITAN X (Maxwell) | 4292 | 55 | 0.889 |
| Titan Black | 2302 | 81 | 1.460 |
| GTX 1080 | 5429 | 48 | 0.709 |
Notes:
nbody -benchmark -numbodies=256000 -device= {one of (0,1,2,3)}
namd2 +p 64 +setcpuaffinity stmv.namd
This is more CPU cores than needed for balance with the GPU but I wanted the GPU to be performance limiting.
I'm looking forward to setting up NVIDIA DIGITS 4 (just released to developers) and seeing what kind performance we see with training sets on Caffe using the new Titan X.
Extra data ... if you like that sort of thing
kinghorn@u16ps:~/testing/samples-8.0/bin/x86_64/linux/release$ ./bandwidthTest 0 [CUDA Bandwidth Test] - Starting... Running on... Device 0: TITAN X Quick Mode Host to Device Bandwidth, 1 Device(s) PINNED Memory Transfers Transfer Size (Bytes) Bandwidth(MB/s) 33554432 11853.2 Device to Host Bandwidth, 1 Device(s) PINNED Memory Transfers Transfer Size (Bytes) Bandwidth(MB/s) 33554432 12854.1 Device to Device Bandwidth, 1 Device(s) PINNED Memory Transfers Transfer Size (Bytes) Bandwidth(MB/s) 33554432 342979.6 Result = PASS NOTE: The CUDA Samples are not meant for performance measurements. Results may vary when GPU Boost is enabled.
kinghorn@u16ps:~/testing/samples-8.0/bin/x86_64/linux/release$ ./deviceQuery
./deviceQuery Starting...
CUDA Device Query (Runtime API) version (CUDART static linking)
Detected 4 CUDA Capable device(s)
Device 0: "TITAN X"
CUDA Driver Version / Runtime Version 8.0 / 8.0
CUDA Capability Major/Minor version number: 6.1
Total amount of global memory: 12186 MBytes (12778274816 bytes)
(28) Multiprocessors, (128) CUDA Cores/MP: 3584 CUDA Cores
GPU Max Clock rate: 1531 MHz (1.53 GHz)
Memory Clock rate: 5005 Mhz
Memory Bus Width: 384-bit
L2 Cache Size: 3145728 bytes
Maximum Texture Dimension Size (x,y,z) 1D=(131072), 2D=(131072, 65536), 3D=(16384, 16384, 16384)
Maximum Layered 1D Texture Size, (num) layers 1D=(32768), 2048 layers
Maximum Layered 2D Texture Size, (num) layers 2D=(32768, 32768), 2048 layers
Total amount of constant memory: 65536 bytes
Total amount of shared memory per block: 49152 bytes
Total number of registers available per block: 65536
Warp size: 32
Maximum number of threads per multiprocessor: 2048
Maximum number of threads per block: 1024
Max dimension size of a thread block (x,y,z): (1024, 1024, 64)
Max dimension size of a grid size (x,y,z): (2147483647, 65535, 65535)
Maximum memory pitch: 2147483647 bytes
Texture alignment: 512 bytes
Concurrent copy and kernel execution: Yes with 2 copy engine(s)
Run time limit on kernels: Yes
Integrated GPU sharing Host Memory: No
Support host page-locked memory mapping: Yes
Alignment requirement for Surfaces: Yes
Device has ECC support: Disabled
Device supports Unified Addressing (UVA): Yes
Device PCI Domain ID / Bus ID / location ID: 0 / 5 / 0
Compute Mode:
- Default (multiple host threads can use ::cudaSetDevice() with device simultaneously) -
Device 1: "GeForce GTX TITAN Black"
CUDA Driver Version / Runtime Version 8.0 / 8.0
CUDA Capability Major/Minor version number: 3.5
Total amount of global memory: 6082 MBytes (6377439232 bytes)
(15) Multiprocessors, (192) CUDA Cores/MP: 2880 CUDA Cores
GPU Max Clock rate: 980 MHz (0.98 GHz)
Memory Clock rate: 3500 Mhz
Memory Bus Width: 384-bit
L2 Cache Size: 1572864 bytes
Maximum Texture Dimension Size (x,y,z) 1D=(65536), 2D=(65536, 65536), 3D=(4096, 4096, 4096)
Maximum Layered 1D Texture Size, (num) layers 1D=(16384), 2048 layers
Maximum Layered 2D Texture Size, (num) layers 2D=(16384, 16384), 2048 layers
Total amount of constant memory: 65536 bytes
Total amount of shared memory per block: 49152 bytes
Total number of registers available per block: 65536
Warp size: 32
Maximum number of threads per multiprocessor: 2048
Maximum number of threads per block: 1024
Max dimension size of a thread block (x,y,z): (1024, 1024, 64)
Max dimension size of a grid size (x,y,z): (2147483647, 65535, 65535)
Maximum memory pitch: 2147483647 bytes
Texture alignment: 512 bytes
Concurrent copy and kernel execution: Yes with 1 copy engine(s)
Run time limit on kernels: No
Integrated GPU sharing Host Memory: No
Support host page-locked memory mapping: Yes
Alignment requirement for Surfaces: Yes
Device has ECC support: Disabled
Device supports Unified Addressing (UVA): Yes
Device PCI Domain ID / Bus ID / location ID: 0 / 9 / 0
Compute Mode:
- Default (multiple host threads can use ::cudaSetDevice() with device simultaneously) -
Device 2: "GeForce GTX TITAN X"
CUDA Driver Version / Runtime Version 8.0 / 8.0
CUDA Capability Major/Minor version number: 5.2
Total amount of global memory: 12207 MBytes (12799574016 bytes)
(24) Multiprocessors, (128) CUDA Cores/MP: 3072 CUDA Cores
GPU Max Clock rate: 1076 MHz (1.08 GHz)
Memory Clock rate: 3505 Mhz
Memory Bus Width: 384-bit
L2 Cache Size: 3145728 bytes
Maximum Texture Dimension Size (x,y,z) 1D=(65536), 2D=(65536, 65536), 3D=(4096, 4096, 4096)
Maximum Layered 1D Texture Size, (num) layers 1D=(16384), 2048 layers
Maximum Layered 2D Texture Size, (num) layers 2D=(16384, 16384), 2048 layers
Total amount of constant memory: 65536 bytes
Total amount of shared memory per block: 49152 bytes
Total number of registers available per block: 65536
Warp size: 32
Maximum number of threads per multiprocessor: 2048
Maximum number of threads per block: 1024
Max dimension size of a thread block (x,y,z): (1024, 1024, 64)
Max dimension size of a grid size (x,y,z): (2147483647, 65535, 65535)
Maximum memory pitch: 2147483647 bytes
Texture alignment: 512 bytes
Concurrent copy and kernel execution: Yes with 2 copy engine(s)
Run time limit on kernels: No
Integrated GPU sharing Host Memory: No
Support host page-locked memory mapping: Yes
Alignment requirement for Surfaces: Yes
Device has ECC support: Disabled
Device supports Unified Addressing (UVA): Yes
Device PCI Domain ID / Bus ID / location ID: 0 / 6 / 0
Compute Mode:
- Default (multiple host threads can use ::cudaSetDevice() with device simultaneously) -
Device 3: "GeForce GTX 1080"
CUDA Driver Version / Runtime Version 8.0 / 8.0
CUDA Capability Major/Minor version number: 6.1
Total amount of global memory: 8113 MBytes (8507555840 bytes)
(20) Multiprocessors, (128) CUDA Cores/MP: 2560 CUDA Cores
GPU Max Clock rate: 1734 MHz (1.73 GHz)
Memory Clock rate: 5005 Mhz
Memory Bus Width: 256-bit
L2 Cache Size: 2097152 bytes
Maximum Texture Dimension Size (x,y,z) 1D=(131072), 2D=(131072, 65536), 3D=(16384, 16384, 16384)
Maximum Layered 1D Texture Size, (num) layers 1D=(32768), 2048 layers
Maximum Layered 2D Texture Size, (num) layers 2D=(32768, 32768), 2048 layers
Total amount of constant memory: 65536 bytes
Total amount of shared memory per block: 49152 bytes
Total number of registers available per block: 65536
Warp size: 32
Maximum number of threads per multiprocessor: 2048
Maximum number of threads per block: 1024
Max dimension size of a thread block (x,y,z): (1024, 1024, 64)
Max dimension size of a grid size (x,y,z): (2147483647, 65535, 65535)
Maximum memory pitch: 2147483647 bytes
Texture alignment: 512 bytes
Concurrent copy and kernel execution: Yes with 2 copy engine(s)
Run time limit on kernels: No
Integrated GPU sharing Host Memory: No
Support host page-locked memory mapping: Yes
Alignment requirement for Surfaces: Yes
Device has ECC support: Disabled
Device supports Unified Addressing (UVA): Yes
Device PCI Domain ID / Bus ID / location ID: 0 / 10 / 0
Compute Mode:
- Default (multiple host threads can use ::cudaSetDevice() with device simultaneously) -
> Peer access from TITAN X (GPU0) -> GeForce GTX TITAN Black (GPU1) : No
> Peer access from TITAN X (GPU0) -> GeForce GTX TITAN X (GPU2) : No
> Peer access from TITAN X (GPU0) -> GeForce GTX 1080 (GPU3) : No
> Peer access from GeForce GTX TITAN Black (GPU1) -> TITAN X (GPU0) : No
> Peer access from GeForce GTX TITAN Black (GPU1) -> GeForce GTX TITAN X (GPU2) : No
> Peer access from GeForce GTX TITAN Black (GPU1) -> GeForce GTX 1080 (GPU3) : No
> Peer access from GeForce GTX TITAN X (GPU2) -> TITAN X (GPU0) : No
> Peer access from GeForce GTX TITAN X (GPU2) -> GeForce GTX TITAN Black (GPU1) : No
> Peer access from GeForce GTX TITAN X (GPU2) -> GeForce GTX 1080 (GPU3) : No
> Peer access from GeForce GTX 1080 (GPU3) -> TITAN X (GPU0) : No
> Peer access from GeForce GTX 1080 (GPU3) -> GeForce GTX TITAN Black (GPU1) : No
> Peer access from GeForce GTX 1080 (GPU3) -> GeForce GTX TITAN X (GPU2) : No
deviceQuery, CUDA Driver = CUDART, CUDA Driver Version = 8.0, CUDA Runtime Version = 8.0, NumDevs = 4, Device0 = TITAN X, Device1 = GeForce GTX TITAN Black, Device2 = GeForce GTX TITAN X, Device3 = GeForce GTX 1080
Result = PASS
Happy computing --dbk
Tags: NVIDIA, CUDA, Pascal GPU, Linux
A good read as usual Dr. Kinghorn
I wonder is it possible to test an AMD GPU because they have a higher FLOP/s at lower price.
Maybe in the budget range, but they really don't stand a chance against Intel chips even at the i5 range. The Bulldozer FX chips shared FPUs between two cores at a time, and so the "8 cores" became 4 cores when doing numerical computing, and those 4 cores are clocked a lot slower / less efficient than those in something like an i7. In my personal tests an overclocked i7 was doing something like 5-10x better for PDE/SDE solving. Thus I relegated that computer to doing purely machine learning work on a GPU. I had similar experiences with Opertron cores in compute clusters, even 64 core nodes vs 16 core Intel Xeon nodes. The moral of the story is, for serious numerical computing, in any benchmark I've done, Intel processors lead by a large margin.