Nvidia, the graphics processing unit specialist, is looking to attract startups working with machine learning technology to opt for their hardware over cloud providers like Microsoft and AWS, in a land grab for the lucrative niche of AI startups.
Startups building products on top of cutting edge artificial intelligence (AI) technology require a huge amount of computing power in order to process large amounts of data, which is expensive.
Cloud giants like Microsoft, Google and Amazon Web Services (AWS) have traditionally given away thousands of dollars in credits for startups to court them onto their platforms. This can pay huge dividends if you manage to get a Snapchat or Airbnb onto your services early on, as shown by Snap's $1 billion commitment to AWS over the next five years.
For AI-focused companies these credits could be put towards running graphics processing units (GPUs) - which are better than traditional CPUs at processing large amounts of data - on their virtual machines in the cloud at a discount. This is extremely useful for the kind of compute-intensive modelling that is required in the background for companies using machine and deep learning techniques.
Now, the GPU manufacturer Nvidia is going head to head with the cloud giants with its Inception Program, in a drive to court the best AI startups directly onto their hardware.
Nvidia Inception Program
Jack Watts, startup business manager at Nvidia explained to Techworld that the aim of the programme, which was launched last summer, is "nurturing those startups and helping them get access to hardware and marketing lift with us, but really the whole thing is about driving this community of startups which are doing weird and wonderful things with AI and Nvidia GPUs."
Members can apply for GPU hardware grants and will be invited to try the latest software and hardware, such as the Nvidia Deep Learning SDK, Digits deep learning GPU training System and the GPU Inference Engine.
Nvidia also provides startups with access to expertise, so they can speak to solution architects to ask questions about which hardware would be best for a specific project (the answer probably being Nvidia), or for advice on certain frameworks or cloud providers. Members also get access to the Nvidia Deep Learning Institute (DLI), with credits going towards training courses.
Startups will get the chance to catch the eye of Nvidia's ventures programme, called GPU Ventures, which got good returns on an investment in mobile game developer NaturalMotion, a company that was acquired by Zynga for $527 million in 2014.
The part that will really grab startup founder's attention though is the discounted hardware. The Nvidia DGX-1 supercomputer, for example, can provide 2-3 times performance improvements over other GPU options in the market and is available to Inception members at rates comparable to that given to academic institutions.
The problem is that a DGX-1 retails for £178,000, so balancing that sort of capital expenditure (capex) against the operational expenditure (opex) of spinning up virtual machines in the cloud is a tricky proposition for startups.
UK AI startup Onfido
This is exactly the issue that UK AI startup Onfido has been grappling with since joining the 80 or so UK startups on the Inception programme.
Speaking to Techworld, Rahul Amin, CTO and cofounder of the identify verification startup said that it is "hard" to decide when to buy hardware against spinning up instances in the cloud, and that they currently do a combination of the two. "We probably haven't found the right balance yet," he said.
Onfido uses machine learning to verify government issued documents for employers and sharing economy platforms that are increasingly keen to verify users and cut down on fraud. Which means Amin understands the importance of GPUs to his core business as well as anyone. "GPUs have been a fundamental part of this process for doing machine learning", he said.
"We started off with a singular machine with a gaming graphics card, as a startup it was all about being efficient. We started to train models for extracting text from a document and as the company grew we have introduced various servers and each development machine has GPUs in them."
Cost is the main issue for a startup though, even one as well funded (£25 million to date) as Onfido. "It has always been a concern how much we spend on GPUs," Amin said.
"If you spin up a machine in the cloud it is going to cost you if you run it for a long time, so balancing what is the best way to train models has been an interesting decision and a conversation we continually have in terms of optimising our hardware and providing engineers with the right level of hardware."