Number of GPU Machines Globally - Historical
In terms of overall GPU sales, numbers have largely been decreasing from 2014-2017. In 2014, 460 million units were sold, followed by 397 million units in 2015, 420 million units in 2016, and 408.4 million units in 2017.
The estimated number of GPUs sold for use in AI-related functions in 2014 (market share estimated at 14%) were 6.44 million units, in 2015 (market share estimated at 15%) were 5.955 million units, in 2016 (market share, estimated at 16.5%) were 6.31 million units, and in 2017 (exact market share value was available) were 7.7628 million units.
To answer your question, I report data on the number of GPUs sold between 2014 and 2017 below. I could not find specific information on what volume of sales were destined for, or used in AI-related functions. However, I have provided global figures and estimations based on research and news articles on the subject.
OVERVIEW OF GPUS IN AI COMPUTING
Pioneered in 2007 by NVIDIA, graphics processing units (GPUs) were originally used for gaming and doing complex simulations, but now play an important role together with computer processing units (CPUs) in accelerating deep learning, analytics, and various other engineering applications. GPUs facilitate parallel computations, which can vastly improve the time and cost-efficiency of training neural nets. Researchers at Stanford built a GPU-accelerated system to train deep nets for US$33 000, compared to Google’s US$5 billion system CPU cluster system built before the deep learning boom.
GPUs can be categorized into 2 categories: discrete or integrated. Integrated GPUs are already onboard the PC, while discreet GPUs are on separate or stand-alone cards. They are also developed and sold as either server GPUs or gaming GPUs. Server GPUs tend to be more compact and are not permitted to run at higher speeds to prevent overheating.
More people are buying gaming GPUs in their data centers to run and develop artificial intelligence software. GPUs marketed for gaming are sometimes used as they are cheaper and can perform at higher speeds than server-grade chips. However, server GPUs serve machine learning purposes better as gaming GPUs are more limited in terms of memory. The largest data center operators including Google, Microsoft and Facebook are also increasing their purchases of server GPUs.
Research and development on GPU hardware and its applications is active, and industry players can expect progress in terms of GPU acceleration capabilities and flexibility in the near future.
OVERALL GPU SHIPMENT VOLUME
Information was publicly available on the number of GPUs shipped across both PC and notebook, discrete and integrated. The information in this section is derived from data collected by Jon Peddie Research (JPR), a research and consulting firm for graphics and multimedia.
Three companies dominate the global GPU market: Intel, AMD and Nvidia. Combining the figures for both desktop and notebook GPUs, overall GPU shipments from 2014 to 2017 were as follows:
2014 - 460 million units
(230 million discrete notebook + 130 million integrated desktop + 50 million discrete notebook + 50 million discrete desktop)
2015—397 million units
(200 million discrete notebook + 110 million integrated desktop + 42 million discrete notebook + 45 million discrete desktop)
2016—420 million units
(200 million discrete notebook + 120 million integrated desktop + 50 million discrete notebook + 50 million discrete desktop)
2017—408.4 million units
For 2017, figures were not readily available. I determined the sales volume for 2017 using absolute figures and growth figures using the following calculation:
GPU sales for notebooks shipped in 2017 = 200 million + 50 million = 250 million * 6% = 15 million - 250 million = 235 million units
GPU sales for PCs shipped in 2017 = 120 million + 50 million = 170 million x 102% = 173.4 million units
Therefore, the combined GPU sales from PCs and notebooks shipped in 2017 was 408.4 million units.
According to Jon Peddie Research, 2017 was an amazing year for GPU development driven by games, eSports, AI, cryptocurrency mining, and simulations.
WHAT PROPORTION OF GPUS ARE PURCHASED USED FOR AI-RELATED FUNCTIONS?
Nvidia was said to have a monopoly over the market for discrete GPUs used for machine learning purposes. All the major clouds with GPU support (Amazon Web Services, Azure, and Google Cloud) are overwhelmingly Nvidia-powered. In a September 2017 report, Susquehanna Financial Group analyst Christopher Rolland estimated that as much as 10% of Nvidia's reported gaming business is being used for data center infrastructure. Using this information together with JPR’s sales volume and market share data, I estimated the volume of GPU sales used in data-center machine learning (and, with less certainty, this can be extrapolated to AI in general).
The estimation can be made according to the following formula:
Estimated number of GPUs sold for use in AI-related functions in a given year = total GPU sales volume (as calculated above) * Nvidia Discrete Desktop GPU market share in that year (from JPR's market share report; used as a proxy for GPU market share for AI-related purposes, as integrated GPUs are not commonly used in AI-related computing) * 0.1 (speculated proportion of Nvidia’s gaming business used for AI-related purposes)
The market share in Q3 was used in the calculation as those are periods that the market typically experiences the strongest sales.
Using this formula, the breakdown of estimated number of GPUs sold for use in AI-related functions by year is as follows:
2014 (market share estimated at 14%)—6.44 million units
2015 (market share estimated at 15%)—5.955 million units
2016 (market share, estimated at 16.5%)—6.31 million units
2017 (exact market share value was available)—7.7628 million units
More detailed calculations can be made with access to figures in JPR’s full Market Watch report (which has a selling price of US$2500).
Little information on future sales projections was found, though most of the articles online expect GPU sales to experience a steady increase.
AMD has plans to compete with Nvidia in the market for GPUs for use in artificial intelligence and cryptocurrency mining, however, the article emphasises the strong foothold Nvidia has in the market.
There is no rigorous pre-compiled information on the volume of GPUs sold and used globally for AI-related processing purposes. Estimations of the number of machines sold for this purpose seem to hover around 6 million units per year, but this is expected to increase, based on GPU sales trends and reports in the media.