The GPU in the clouds


the Need to build more GPU

the Deep Learning is one of the most intensively developing directions in the field of machine learning. Research achievements in the field of deep (deep) learning call for an increase of ML/DLframeworks (including Google, Microsoft, Facebook), implementing these algorithms. For increasing computational complexity of DL algorithms, and, as a consequence, the increasing complexity of DL-frameworks have not hijacked hardware power nor table, nor even the server CPUs.

The solution was found, and it is simple (I think so) to use for this type of compute-intensive tasks calculations on GPU/FPGA. But here's the problem: you can, of course, to use the graphics card beloved laptop, but a Russian data scientist does not like quick drive NVidia Tesla?

Approaches to owning high-performance GPU at least two: buy (on-premises) and rent (on-demand). How to save up and buy the topic of this article. In this, we will consider what proposals have been made for renting VM instances c high performance GPU cloud providers Amazon Web Service and Azure.

the

1. The GPU in Azure


In early August of 2016, it was announced the beginning of closed testing (private preview) instances of virtual machines equipped with NVidia Tesla [1]. This option appears in the context of the service Azure VMIaaS-service provides virtual machines on demand (similar to Amazon EC2).


from the point of view of application access to a graphics processor architecture of the service looks like this:


Azure VM Instances GPU Architecture

Calculations on the GPU available for virtual machines N series, which, in turn, are divided into 2 categories:


the
    the
  • NC Series (computer-focused): GPU focus on calculation;
  • the
  • NV Series (visualization-focused): GPU aimed at the graphics calculations.

the

1.1. NC Series VMs


Graphics processors are designed for compute-intensive load with the use of CUDA/OpenCL. Graphics is NVidia Tesla K80: 4992 CUDA cores, >2.91/8.93 c Tflops double/single precision). Access cards are made using the technology of DDA (discrete device assignment), which brings the GPU performance when using via VM to bare-metal performance of the card.


As you can guess, the VM NC series is intended for ML/DL-task.


In Azure VM the following configuration, equipped Tesla K80.

the the the the the
NC6
NC12
NC24
Cores
6 (E5-2690v3)
12 (E5-2690v3)
24 (E5-2690v3)
GPU
1 x K80 GPUs (1/2 of the Physical Card)
2 x K80 GPU (1 Physical Card)
4 x K80 GPU (2 Physical Cards)
Memory
56 GB
112 GB
224 GB
Disk
380 GB SSD
680 GB SSD
1.44 TB SSD

1.2. NV Series VMs


Virtual machines NV series is designed for visualization. On the data VM are GPU Tesla M60 (4086 CUDA cores, 36 threads for 1080p H. 264). These cards are suitable for the task (de)coding, rendering, 3D modeling.


Announced the availability of the VM instances with the following configurations:

the the the the the
NV6
NV12
NV24
Cores
6 (E5-2690v3)
12 (E5-2690v3)
24 (E5-2690v3)
GPU
1 x M60 GPU (1/2 Physical Card)
M60 2 x GPU (1 Physical Card)
M60 4 x GPU (2 Physical Cards)
Memory
56 GB
112 GB
224 GB
Disk
380 GB SSD
680 GB SSD
1.44 TB SSD

1.3. Prices


Prices for N-Series Azure VM as follows (October 2016) [5]:
azurevm gpu instances.

But let Your curiosity, these 4-ehsani numbers do not decrease: as always, in the cloud we pay for the use of resources. For IaaS services, which service is the Azure VM, it is necessary to understand how hourly billing. In addition, Microsoft Azure there are many ways to gold computing resources absolutely for free.


This applies to new accounts in Azure at students for startups, if you looking for a cure for cancer Explorer, or if You/the company You work for, the owner of the MSDN subscription.


the

2. Amazon EC2 GPU Instances (dangerous+comparison)


Cloud provider Amazon Web Services (AWS) started providing VM instances with GPUs back in 2010.


at the beginning of September (2016) GPU instances on AWS was represented only by the family of G2.


Technical details about the family of G2 instances

the virtual machine Configuration collection G2:

the the the
Model GPUs vCPU Mem (GiB) SSD Storage (GB) Price per hour/month
g2.2xlarge 1 8 15 1 x 60 0.65/468
g2.8xlarge 4 32 60 2 x 120 2.6/1872

the G2 Instances are equipped with graphics processors NVidia GRID K520 with 1556 CUDA cores, support for 4 streams of 1080p H. 264. Announced support for CUDA/OpenCL. There is also a technology support HVM (hardware virtual machine), which is similar to the DDA in Azure VM, minimizes the overhead associated with virtualization by allowing the guest VM to GPU performance close to bare-metal performance.



While I wrote an article a month ago (end of September 2016) AWS announced a P2-instances with more modern graphics card.


Instances of the family of P2 can include up to 8 cards NVIDIA Tesla K80. Entered 7.5 CUDA, and OpenCL 1.2. Instances of p2.8xlarge and p2.16xlarge support high-speed GPU-to-GPU connection, and LAN connection is available up to 20 Gbps technology ENA (Elastic Network Adapter – high-speed network interface for Amazon EC2).

the the the the
Instance Name GPU Cores vCPU Cores Memory, Gb CUDA Cores GPU Memory Network Gbps
p2.xlarge 1 4 61 2496 12 High
p2.8xlarge 8 32 488 19968 96 10
p2.16xlarge 16 64 732 39936 192 20

For comparison* we take the most productive (NC24) and the budget (NC6) instances in Azure VM and closest in performance to Jurowski instances in Amazon EC2.

the the the the the
Instance Family GPU Model GPU Cores vCPU Core RAM, Gb Network Gbps CUDA/OpenCL Status Price, $/mo Price, $ per GPU/mo
Amazon p2.xlarge K80 1 4 61 High 7.5/1.2 GA 648 648
Azure NC6 K80 1 6 56 10 (?) +/+ Private preview 461 461
Amazon p2.8xlarge K80 8 32 488 10 7.5/1.2 GA 5184 648
Azure NC24 K80 8 24 224 10 (?) +/+ Private preview 1882 235
* UDP: the Price and configuration current at 25 October 2016.

the

Conclusion


AWS long "tortured" the data-science community is pretty weak and however expensive GPU instances G2 family. But the competition in the market of cloud providers has done its job – a month ago there was a GPU instances of the family of P2, and they look very decent.

Microsoft Azure also long plagued the community in General the lack of GPU instances (this was one of the most anticipated features of the Azure platform). At the moment GPU-instances in Azure looks very good, though lacking in technical details. Рreview status of this capability is a big minus normal stage in the life cycle of most cloud services.
Actually Microsoft just a year or two has seriously acquired a variety of AI technologies / frameworks/ tools, including (maybe primarily) for developers and data scientist s. How serious this is and whether it is possible to evaluate yourself, looking at records from the past at the end of September Microsoft ML & DS Summit [6].

the in addition, a week – 1 November – will be the conference Microsoft DevCon School, one of the tracks which is completely dedicated to machine learning. And tell them there will be not only about proprietary MS technology, and about the usual and "loose" Python, R, Apache Spark.

the

sources List


    the
  1. NVIDIA GPUs in Azure: check in the preview program.
  2. the
  3. Leveraging NVIDIA GPUs in Azure. Webcast on Channel 9.
  4. the
  5. Linux GPU Instances documentation.
  6. the
  7. the announcement of the P2-instance on AWS, September 29, 2016.
  8. the
  9. Prices on Azure Virtual Machines (including the Azure VM GPU).
  10. the
  11. Conference Microsoft Machine Learning &Data Science Summit.
Article based on information from habrahabr.ru

Комментарии

Популярные сообщения из этого блога

Templates ESKD and GOST 7.32 for Lyx 1.6.x

Monitoring PostgreSQL + php-fpm + nginx + disk using Zabbix

Custom table in MODx Revolution