DPC without GeForce and Titan: NVIDIA changed license agreement
NVIDIA has changed the license agreement for the driver, and now the use of GeForce and Titan GPUs in data centers is prohibited. Why it happened, who will be affected by the changes and what are the alternatives, read under the cut.
/ photo Fritzchens Fritz PD
NVIDIA added a new provision to article 2.1.3 of the GeForce Software User Agreement. It reads: “No Datacenter Deployment. The SOFTWARE is not licensed for datacenter deployment, except that the blockchain processing is a datacenter is permitted ”and concerns NVIDIA GeForce and Titan products. Now it is forbidden to use these cards for any tasks, except for work with blockchain technologies. To work with machine learning and simulation in the data center now you can only use Tesla V100. Why the company made this decision will tell further.
')
NVIDIA Company began its way as a developer of graphics processors for the gaming industry. Later, these accelerators were used for research and business purposes, including in data centers (we already wrote about why the GPU is better suited for high-performance computing).
However, as stated by representatives of NVIDIA, graphic accelerators GeForce and Titan are not intended for deployment in data centers. The work of data centers is associated with high requirements for hardware and software solutions in 24x7 mode. At the same time, it is impossible to guarantee the performance of these cards in a room with a high density of equipment in racks and, accordingly, a high ambient temperature.
Instead, NVIDIA proposes to use the Tesla V100 - a graphics accelerator designed specifically for use in machine room conditions and with greater performance. If we talk about 64-bit operations with floating-point numbers, Tesla is much more powerful than GeForce: 7 teraflops versus 0.355 teraflops (GTX 1080 Ti). Tesla cards also outperform GeForce on floating-point and half-precision single-precision floating-point operations.
NVIDIA also considers Tesla solution more stable, therefore more suitable for business. For example, stability is provided by NVIDIA NVLink technology - a computer bus that serves to connect CPU and GPU and uses cache coherence protocols.
NVIDIA's CUDA parallel computing architecture is widely supported by cuDNN machine learning libraries, so researchers and developers of artificial intelligence systems switched to NVIDIA products, which caused the company's growth by 85% in 2017.
For example, universities in Florida and North Carolina using NVIDIA are developing a neural network engine for modeling in the field of quantum mechanics.
Many organizations work with GeForce and Titan cards because of the price. GeForce GTX 1080 Ti cost $ 699. For comparison, the last Tesla V100 card, sharpened for data centers, costs about 8 thousand dollars. The added agreement clause may become an obstacle to research and development of new products.
But for all that, NVIDIA notes that the changes will not affect those researchers and developers who adapt the company's products for non-commercial purposes and use accelerators outside the data centers.
Reddit users indicate that the ban applies only to software, not hardware. Therefore, you can write your own drivers and then use NVIDIA graphics cards legally. However, this is difficult to implement in practice, because the IT giant has not provided the appropriate specifications for hardware (some users are convinced that this is impossible at all, since the controllers on all NVIDIA cards do not support anything other than proprietary drivers).
Another option is to use old drivers that are not subject to changes in the new license agreement. But in this case you will have to accept a limited level of support.
Also, residents of Reddit note that you should pay attention to the project ROCm (Radeon Open Compute) - an open-source-platform for high-performance computing on graphics processors, which does not depend on the programming language. The platform allows you to work with almost any video card, including NVIDIA products.
This is possible thanks to the HIP (Heterogeneous-Computing Interface for Portability) - a C ++ dialect that simplifies the conversion of CUDA applications into portable C ++ code. The Hipify tool automates the conversion process, which allows you to run HIP code on AMD hardware (using the HCC compiler) and NVIDIA (using the NVCC compiler).
PS Materials on the topic from our blog on Habré:
PPS Materials from the First Corporate IaaS Blog:
/ photo Fritzchens Fritz PD
What changed
NVIDIA added a new provision to article 2.1.3 of the GeForce Software User Agreement. It reads: “No Datacenter Deployment. The SOFTWARE is not licensed for datacenter deployment, except that the blockchain processing is a datacenter is permitted ”and concerns NVIDIA GeForce and Titan products. Now it is forbidden to use these cards for any tasks, except for work with blockchain technologies. To work with machine learning and simulation in the data center now you can only use Tesla V100. Why the company made this decision will tell further.
')
What is the reason
NVIDIA Company began its way as a developer of graphics processors for the gaming industry. Later, these accelerators were used for research and business purposes, including in data centers (we already wrote about why the GPU is better suited for high-performance computing).
However, as stated by representatives of NVIDIA, graphic accelerators GeForce and Titan are not intended for deployment in data centers. The work of data centers is associated with high requirements for hardware and software solutions in 24x7 mode. At the same time, it is impossible to guarantee the performance of these cards in a room with a high density of equipment in racks and, accordingly, a high ambient temperature.
Instead, NVIDIA proposes to use the Tesla V100 - a graphics accelerator designed specifically for use in machine room conditions and with greater performance. If we talk about 64-bit operations with floating-point numbers, Tesla is much more powerful than GeForce: 7 teraflops versus 0.355 teraflops (GTX 1080 Ti). Tesla cards also outperform GeForce on floating-point and half-precision single-precision floating-point operations.
NVIDIA also considers Tesla solution more stable, therefore more suitable for business. For example, stability is provided by NVIDIA NVLink technology - a computer bus that serves to connect CPU and GPU and uses cache coherence protocols.
Who will be affected by the change
NVIDIA's CUDA parallel computing architecture is widely supported by cuDNN machine learning libraries, so researchers and developers of artificial intelligence systems switched to NVIDIA products, which caused the company's growth by 85% in 2017.
For example, universities in Florida and North Carolina using NVIDIA are developing a neural network engine for modeling in the field of quantum mechanics.
Many organizations work with GeForce and Titan cards because of the price. GeForce GTX 1080 Ti cost $ 699. For comparison, the last Tesla V100 card, sharpened for data centers, costs about 8 thousand dollars. The added agreement clause may become an obstacle to research and development of new products.
But for all that, NVIDIA notes that the changes will not affect those researchers and developers who adapt the company's products for non-commercial purposes and use accelerators outside the data centers.
Alternative solutions
Reddit users indicate that the ban applies only to software, not hardware. Therefore, you can write your own drivers and then use NVIDIA graphics cards legally. However, this is difficult to implement in practice, because the IT giant has not provided the appropriate specifications for hardware (some users are convinced that this is impossible at all, since the controllers on all NVIDIA cards do not support anything other than proprietary drivers).
Another option is to use old drivers that are not subject to changes in the new license agreement. But in this case you will have to accept a limited level of support.
Also, residents of Reddit note that you should pay attention to the project ROCm (Radeon Open Compute) - an open-source-platform for high-performance computing on graphics processors, which does not depend on the programming language. The platform allows you to work with almost any video card, including NVIDIA products.
This is possible thanks to the HIP (Heterogeneous-Computing Interface for Portability) - a C ++ dialect that simplifies the conversion of CUDA applications into portable C ++ code. The Hipify tool automates the conversion process, which allows you to run HIP code on AMD hardware (using the HCC compiler) and NVIDIA (using the NVCC compiler).
PS Materials on the topic from our blog on Habré:
- Sparse matrices: how scientists accelerated machine learning on the GPU
- The first supercomputer DGX-1 based on Tesla V100 will be used in medicine
PPS Materials from the First Corporate IaaS Blog:
Source: https://habr.com/ru/post/346072/
All Articles