DigiTimes just published a massive report (hidden behind a paywall) that sheds some light on NVIDIA's plan for Ampere GPUs as well as their future plans. It had been rumored that NVIDIA was thinking of going with Samsung's 7nm process but those rumors appeared to have been debunked. As it turns out, and if DigiTimes report is to be believed (and they have an excellent track record for what it's worth), there was some credence to them after all and NVIDIA is considering dividing its lineup into TSMC and Samsung's 7nm process.
NVIDIA contemplating dividing Ampere GPUs into TSMC's high-yield 7nm EUV and Samsung's low-cost 7nm/8nm EUV node
So here is what happened as per the Digitimes report: NVIDIA wanted to lower its dependence on TSMC and was courting Samsung to use their 7nm/8nm EUV process. Since NVIDIA is a fabless company, this was an attempt to diversify away the risk. According to the report, the company decided, in the end, to go back to TSMC. Unfortunately for them, however, AMD had taken this opportunity to buy up spare 7nm capacity aggressively and established itself as the name behind the world's first 7nm GPU. The report also states that one of the rationales behind approaching Samsung was for TSMC to lower their pricing. Considering TSMC is now operating in a high demand, low supply ecosystem, I am fairly certain NVIDIA did not get their price cuts.
According to the report, NVIDIA has already placed a "major" order for 7nm EUV node at TSMC. So this node has been confirmed in use for the company's upcoming Ampere GPUs which will be unveiled at GTC 2020. But, the company is also considering tapping Samsugn's 7nm/8nm EUV process to produce some of its lower-end GPUs on in order to match demand. A word of caution though: the report itself admits that this decision has not been set in stone and while using both TSMC and Samsung seems like the best option NVIDIA has right now, they could still change their minds.
The crux of the Digitimes report, via RetiredEngineer @chiakohua (and reproduced with permission):
Focus of GTC'2020 online conference on 5/14 will be around AI, HPC, data science, autonomous machines, healthcare, and recent advances in graphics technology. Jensen Huang will unveil Nvidia's next-gen Ampere's architecture and discuss the future outlook.
Due to uncertainties surrounding Covid-19, Ampere which was originally scheduled for Q3'20 launch may experience slight delays. 3. Due to the high cost of advanced nodes, and the desire to have a second source for risk diversification, Nvidia may use the higher-cost, but better-yielding TSMC 7nm EUV process for its mid-to-high-end products, while part of the lower volume, mid-to-low end products will use Samsung's aggressively priced 7nm EUV and 8nm processes.
Worth noting, Nvidia previously underestimated the effect of AMD embracing TSMC, and made some errors in its own plans for migrating to advanced nodes. Moving to Samsung and wavering in its TSMC strategy, only to go back to TSMC later, Nvidia was unable to secure enough 7nm capacity, resulting in AMD grabbing the limelight, eroding Nvidia's brand value, and increasing its own market share for GPUs.
As the costs for advanced nodes began skyrocketing, there have been rumors that Nvidia and others frequently leaked news about their intention to increase Samsung's share of orders, to pressure TSMC into lowering prices. However, thanks to its absolute leadership position and demand exceeding supply, TSMC did not budge. Customers had little choice but to throw cash at TSMC to fight for limited production capacity while giving some small orders to the aggressive discounter Samsung. For Qualcomm, in addition to aggressive discounts, there's also the tacit understanding that Samsung will use Qualcomm's chipsets in their mobile handsets.
There are also rumors that AMD may shift some orders to Samsung. However, industry observers believe this is very unlikely at the moment. Besides yield and performance, TSMC's brand power is also important, as can be seen from AMD's repeated emphasis of the TSMC partnership. Others like Huawei, Ali Baba, Mediatek, Qualcomm, etc. have done the same. Ever since Samsung announced mass production of its 7nm EUV process, yields and customer acceptance have largely been unknown. Samsung's 5nm and 3nm are also paper launches. This year, the pandemic has resulted in significant delays in installation schedules for EUV-related equipment, and competition for capacity amongst fabless chip companies remains intense. As such, the likes of Nvidia and AMD will not risk shifting orders to Samsung. (DigiTimes)
This is big news for NVIDIA fans. Since the process has been confirmed to be TSMC 7nm EUV for the high-end lineup, we can assume some truly phenomenal performance leaps. The EUV technique is a difficulty reset which results in super clean etches and will improve ASIC quality significantly. According to the story, even the Samsung 7nm/8nm process is an EUV based one but is much much lower cost (but has lower yields than TSMC's).
[speculation] This also means that there might be a big pricing difference in the high and low-end lineups as NVIDIA hands down the cost to the consumers. We have already seen the company split current generation cards into RTX and GTX and this is something that might be amplified further with two - entirely differently priced nodes - in the mix. One thing is for sure though, Ampere GPUs are going to kick some *** [/speculation].