TITAN V and R performance

February 20, 2018
R fun

I have been using an AMD/ATI Radeon 7970 3GB GPU since it first came out in 2012 and for a number of reasons I have finally decided it was time for an upgrade! (One of the reasons is that TensorFlow is only in CUDA).

Goodbye ol’faithful 7970

I’m going to miss this card. It’s held up very well over the years, and the Arctic Accelero cooler I put on top of it made the card whisper quiet in addition to cooling it better by about 20C!

7970 w/ Accelero cooler 1

7970 w/ Accelero cooler 1

7970 w/ Accelero cooler 2

7970 w/ Accelero cooler 2

Hello TITAN V

I got the TITAN V as the replacement to the 7970. Here’s some unboxing pics I took:

Pretty white box

Pretty white box

Card in box

Card in box

TITAN V next to 7970

TITAN V next to 7970

The TITAN V was released a few months ago and there are already several excellent reviews out about the card already. I highly recommend Anandtech.

Specification Comparison
Titan V Titan Xp GTX Titan X (Maxwell) GTX Titan
CUDA Cores 5120 3840 3072 2688
Tensor Cores 640 N/A N/A N/A
ROPs 96 96 96 48
Core Clock 1200MHz 1485MHz 1000MHz 837MHz
Boost Clock 1455MHz 1582MHz 1075MHz 876MHz
Memory Clock 1.7Gbps HBM2 11.4Gbps GDDR5X 7Gbps GDDR5 6Gbps GDDR5
Memory Bus Width 3072-bit 384-bit 384-bit 384-bit
Memory Bandwidth 653GB/sec 547GB/sec 336GB/sec 228GB/sec
VRAM 12GB 12GB 12GB 6GB
L2 Cache 4.5MB 3MB 3MB 1.5MB
Single Precision 13.8 TFLOPS 12.1 TFLOPS 6.6 TFLOPS 4.7 TFLOPS
Double Precision 6.9 TFLOPS (1/2 rate) 0.38 TFLOPS (1/32 rate) 0.2 TFLOPS (1/32 rate) 1.5 TFLOPS (1/3 rate)
Half Precision 27.6 TFLOPS (2x rate) 0.19 TFLOPs (1/64 rate) N/A N/A
Tensor Performance (Deep Learning) 110 TFLOPS N/A N/A N/A
GPU GV100 (815mm2) GP102 (471mm2) GM200 (601mm2) GK110 (561mm2)
Transistor Count 21.1B 12B 8B 7.1B
TDP 250W 250W 250W 250W
Manufacturing Process TSMC 12nm FFN TSMC 16nm FinFET TSMC 28nm TSMC 28nm
Architecture Volta Pascal Maxwell 2 Kepler
Launch Date 12/07/2017 04/07/2017 08/02/2016 02/21/13
Price $2999 $1299 $999 $999

Benchmarks

R/Keras - CNN/LSTM Model

This is a run against the IMDb sentiment classification task dataset, as found as an example on the Keras RStudio website.

The model itself consists of a convolutional network layer, followed by a LSTM layer and 1 epoch contains 25,000 samples for training.

Training time with an Intel i7 980x CPU at 3.4GHz:


It’s taking about 25s to complete 1 epoch on the Intel CPU.

Training time with TITAN V:


It’s taking about 4s to complete 1 epoch on the TITAN V.

The first run on the TITAN V seems to be abnormally longer, at 14s, however each successive run after that completed in 4s. As seen in the graph above, the TITAN V absolutely crushes the Intel CPU whilst using the same batch size. Note, that the TITAN V is only using about 30% of its capacity. On average, the Intel CPU was 6.25x slower than the TITAN V.

R/Keras - MLP Model

A multilayer perceptron model trained on Reuters newswire data for a topic classification task, found here.

Here, the TITAN V is again faster, but not by quite as much as the CNN/LSTM model. The Intel CPU is 1.9x slower.

Other - Mining Garlicoins (cryptocurrency)

Garlicoin is a new cryptocurrency and I’ve been mining it for fun on my 7970 GPU. I will also do some mining (or baking rather) on the TITAN V.

Speed comparison on 7970:

7970 hashrate

7970 hashrate

Speed comparison on TITAN V:

TITAN V hashrate

TITAN V hashrate

The 7970 was able to mine with a hashrate of approx. 310Kh/s, while the TITAN V was able to mine with a hashrate of approx. 12,500Kh/s. That’s a 40.32x improvement!

Other - Unigine Superposition

Superposition is a new-generation benchmark tailored for testing reliability and performance of the latest GPUs.

Running at 1080p Extreme - scores:

NVIDIA TITAN V: 8742
AMD 7970: 1887

The TITAN V is 4.63x faster here.

Conclusion

Really expensive… but really fast!

A look at teacher misconduct in Canada...

August 8, 2018
R data

RNN vs. Dense neural networks for time-series

June 27, 2018
R modelling

Auto neural networks vs. Manual Keras neural model

June 11, 2018
R modelling