GPU Powered * : Accelerated computing for the future

GPUs are making a big buzz in the market. I am seeing the word GPU Powered as a norm. You can also see that with NVIDIA’s share price rising at a faster pace. Couple of months back, I had to look into the GPU world since there were multiple requests around doing data science using GPUs and multiple people has started asking me about it.

We all know that it is used in mobile devices for things like games. My curiosity started rising and some of the questions that came into my mind was how does a Graphic Processing unit relate to the application/database/analytics world? How is different from CPUs? Why everyone is suddenly talking about it?

Following are some of links that has helped me understand this world. Sharing my experience so far. You may want to start watching these videos and posts
What is a GPU and how does it work?
https://www.youtube.com/watch?v=0_TN845dxUU&t=1s>
What is GPU Accelerated computing?
http://www.nvidia.com/object/what-is-gpu-computing.html
CPUs and GPUs – There’s enough room for everyone
http://sqream.com/cpu-and-gpus-theres-room-for-everyone/

First things first. You will realize that there are two types of GPUs. Server Grade and Consumer Grade GPUs

Server-grade GPUs are designed so that they can be installed next to each other with no space separation, and fill all the available PCI slots available on the server, thus optimizing space usage and maximizing the amount of compute power per rack space unit. You can fit four Tesla K80 boards in a 1U server; that’s 8 GPUs total (K80 boards have 2 GPUs each), and that’s an impressive amount of compute throughput. The same applies to Tesla Pascal P100 models, with the due differences (one GPU per board). If you are building a supercomputer or a GPU-based server farm and buying hundreds or thousands of GPUs, these details matter a lot.

Consumer-grade GPUs typically have active cooling with a fan that ingests airflow orthogonally to the longitudinal axis of the board. That requires space clearance to accommodate the cooling airflow. That leads to less dense configurations than in servers. Typical consumers do not care because they have computer cases with more vertical development, less need for density, and most users only have one GPU card per host.

Via : Daniele Paolo Scarpazza, https://www.quora.com/What-is-the-different-between-gaming-GPU-vs-professional-graphics-programming-GPU

My interests are around Server grade GPUs.Where can I find a Server grade GPU to explore?
Use AWS – AWS has G2, P2 and F instances supporting. P2 generally fits in my ecosystem very well. Refer to this presentation Deep Dive on Amazon EC2 Instances (Slide 42 onwards). Google and Microsoft also provides similar instances. NVIDIA has a GPU Cloud, which is in Private Beta I believe.

How do I start Programming in GPUs?
To program in a GPU, NVIDIA has created a platform called CUDA 10 years back. One can also use OpenCL to program. But these are low level languages like “C”.
May not be for everyone. Are there higher level abstractions available? Yes. Let us start looking from Databases side.

Welcome to the world of GPU Powered Databases. MapD and Kinetica are very promising in this space and there are other databases like BlazingDB. Some of the benchmarks these guys are quoting are at the scale of 30X to 100X difference with a typical MPP database in the market.
How do you setup one of these databases? AMIs are available in AWS Marketplace. Use the Opensource one to play with (will work out cheaper). Try MapDs New York City taxi rides demo in your environment.  When I saw the demo first time, I was speechless for some time.

Watch these videos by Todd Mostak, the founder of MapD talking. Very insightful.
The Rise of the GPU: How GPUs Will Change the Way You Look at Big Data
https://www.youtube.com/watch?v=mwpd13urFog&t=25s
The Promise of GPU Analytics
https://www.youtube.com/watch?v=qxWSSz8x6NI
On a side note, it is also important why column stores work better in an in-memory world. I understand now that this can be used for powering databases (In-memory powered by CPUs).

What else can we use it for? How about Visualization? If the Visualization doesn’t support these kind of model, then you still will not get the interactivity what you are looking for.
I liked the way MAPD has done their visualization. Taking the power of GPUs at the consumer end and rendering it. Look for OPENGL, VEGA and D3.Js. You will be able to see how to use GPUs for Visualization.
http://developer.download.nvidia.com/presentations/2008/NVISION/NVISION08_MGPU.pdf
You may want to take a look at these JS libraries
https://stardustjs.github.io/
http://gpu.rocks/

What else can we do with GPUs? The real advantage of using GPUs will be when we leverage its power for Parallel Processing. Which means, this would be extremely useful for any complex Mathematical calculations etc. or scenarios like deep learning.
Today all the deep learning frameworks (Tensorflow or Deeplearning4J or H20) support GPU natively. All you need to do is install it on GPUs and establish the device mapping. Tensorflow would then take care of it automatically. Setting this up, is straight forward. This enables support not only for CPUs or GPUs but also for TPUs (Google latest one) or anything that comes in future.

To understand more watch this video: Effective TensorFlow for Non-Experts
https://www.youtube.com/watch?v=5DknTFbcGVM

BTW, I run Keras with Tensorflow as a backend and use Python (PyCharm) in my machine -:)

Conclusion
I believe that this is a great technology and that this is going to stay . In my opinion it is not CPU Vs GPU. It is GPGPU and it will be the way forward. MapD has raised 25M and Kinetica has raised 50M recently. This definitely shows the potential in this space. I also believe this model would bring back the Scale-up model from our current Scale-out model. Power consumption/Energy conservation is one definite plus, with reduced maintenance efforts. Though the cost of the server (P2 Instance vs T2 Instance) is high right-now, it is not apple to apple comparison. Also, the prices will come down over a period of time.

One thing which I liked the most in the GPU Powered database world is that, since you can run these queries and get sub-second performance, this will help you move out of all the Pre-computations, aggregations we used to in our analytics world. And this also doesn’t have that 32 concurrent user restriction a typical MPP database may have. You will see lot more traction in this area with all major companies moving in this direction and acquisitions.

As always, it’s great to play in the bleeding edge area. Get to learn new things on a regular basis.

Happy Learning!