Retired Engineer, a Twitter user with some credible leaks, offers his perspective on a recent DigiTimes report about NVIDIA's plans.
We'll cover the main points in bullet points:
As expected, gaming will not be the main focus of next week's GTC 2020 keynote (May 14). The company will reveal its new Ampere architecture at this digital event, but AI, HPC, self-driving car, and healthcare applications will be the main talking points.
Ampere was expected in Q3 2020 but the COVID-19 pandemic may cause further delays.
TSMC 7nm EUV to be the main working horse. Lower volume mid-to-low-end GPUs from NVIDIA may use the cheaper 7nm EUV and 8nm nodes from Samsung.
NVIDIA Hopper (next architecture after Ampere) will use TSMC 5nm in 2021. Samsung may be used for smaller volume applications.
NVIDIA reportedly underestimated the impact of AMD's switch to TSMC. NVIDIA's strategy to switch some products to Samsung didn't pay off and resulted in NVIDIA being unable to secure enough supply on TSMC's fully-booked nodes as AMD is now also fighting for capacity on TSMC's advanced nodes for its processors. In an effort to leapfrog AMD, NVIDIA accelerated its plan to adopt TSMC's 7nm and 5nm EUV nodes.
Samsung wants to build up its foundry business but is still having a hard time to find clients. As such, the South Korean giant has to turn to low pricing to lure clients. But NVIDIA and others primarily use Samsung in an effort to pressure TSMC to lower its prices, as TSMC offers better yields and performance, as well as a proven history of being able to deliver what it promises. This echos rumors and reports that circulated last year, even with low pricing Samsung is unable to lure major clients away from TSMC. As TSMC is the undisputed leader in the foundry market, the Taiwanese foundry sees no reason to lower its prices as demand for its advanced nodes is higher than the available capacity.