Estimate machine learning semis will be a significant market of $35B by 2021
Artificial Intelligence and Machine Learning(ML) gained much visibility in recent years as systems such as IBM Watson and AlphaGo surpassed humans in competitions. The democratization of AI now underway could drive a new phase of growth in Semis that our new proprietary model estimates could represent a $35B revenue TAM by 2021 (ex-memory), or a 41% CAGR from current levels. UBS Evidence Lab did testing between GPU chips and CPUs and found accelerators to be the primary beneficiaries of this growth. This report seeks to address: 1) How big is the ML Semis opportunity, 2) When is the demand inflection, 3) Who are the key beneficiaries due to ML adoption?
Data center the dominant demand driver with GPUs reaching 35% penetration
Our proprietary model forecasts a 41% 5-yr CAGR in ML semis based on our top-down and bottoms-up model. We estimate chips sold into hyperscale data center and auto markets will represent 86% / 14% of the revenue mix by 2021, with GPUs at 54% of the mix followed by CPU at 40% and FPGA at 5%. In terms of penetration rates, we assume GPUs reach 35% in the data center and 50% in auto, as CPUs remain dominant in the cloud and make inroads into autonomous driving.
UBS Evidence Lab GPU vs CPU benchmark tests keep us positive on accelerators
UBS Evidence Lab results of chip performance running machine learning workloads show that a single GPU has a 12-18x performance advantage over server CPUs, and this gap could translate into greater revenue growth for GPU suppliers vs processors given the performance-to-price ratio. Results for dual-/quad-GPU configurations show 35x/70x advantages over CPUs while running image classification training algorithms.
Most/Least Preferred: Favor the current AI semiconductor market leaders
We estimate NVIDIA has a time to market advantage of 1+ year and should be able to sustain this lead. We believe Broadcom is also well positioned in ASICs and expect to see more design wins at cloud customers. We like TSMC given its leadership position in 7nm and CoWoS, plus expanding demand. Samsung is the closer foundry contender, but will benefit more from memory. We expect Intel to also benefit as it integrates the Nervana chips into its AI portfolio in a few years, but execution is key. We believe AMD can also gain some CPU share in the next 2-3 years but GPUs is still a work in progress.