Nothing stopped nvidia to postpone the RTX 4000 series launch by several months and make it on the far superior N3 process instead.
But nvidia is run by lunatics, while AMD is run by lazy who will lag behind until 2025.
"Several months"? How long, exactly, do you think the processing of a single wafer takes, and how long, exactly, do you think it will take for TSMC's 3nm production capacity to surpass Apple's orders? And, crucially, how long do you think it takes to build up sufficient stock of a new chip to ship to AIB partners that can then make it into a sufficient amount of GPUs for a retail launch?
Let's be
extremely generous and assume the N3 "H2 2022" date means July 1st. And let's be even more absurdly generous and say it took TSMC three months to have significant spare N3 capacity that Apple either didn't want or didn't have an option to buy first. That would place the start of GPU die mass production on October 1st. Processing a single wafer takes ~1 month (a bit less on some processes, more on others). Let's assume TSMC had 10 000 wafer starts/month of free N3 capacity (which is likely closer to their
total N3 capacity that early), and that they all went into producing a 3nm AD102. Assuming perfect scaling, that would be a ~375mm2 die (
using scaling numbers from here). Let's assume the chip is a bit rectangular, something like 15*25mm. Assuming the "N7 standard" 0.09 defects/cm2 - which is again highly unlikely this early in mass production of a new node, and thus a very generous assumption - and plugging those numbers into a wafer yield calculator, that gives us 101 good dice per wafer, and 39 defective dice, the majority of which can likely be used for a cut-down SKU. Let's say they get 130 RTX 4090-capable dice out of each wafer. That gives us 1.3M GPU dice after a month of production. Which sounds like a lot! Except that
every single assumption made in this calculation is
wildly unrealistic.
- Apple most likely has 6 months or more exclusive access to N3 mass production, and past any exclusivity period they will have first purchasing rights to any spare capacity
- Even disregarding that, there is no way there are 10 000 available wafer starts after three months for Nvidia - 10% of that is more likely, if not even less.
- This all assumes that the chip is already fully taped out and has had test production runs so that the ramp to mass production is smooth.
- N3 yields are likely significantly lower than 0.09 defects/cm2 this early in mass production.
And, crucially, the die being done doesn't mean it's ready to go on shelves - the wafers need to be shipped to a packaging plant where the wafer is cut, the dice are tested, binned and packaged. The packaged dice are then sold and shipped on to AIB partners, who - again, assuming they've had sufficient access to ES silicon - will then ramp mass production of GPU PCBs and coolers (some of this might be made ahead of time, but these companies generally don't have the cash flow to keep significant parts stock for long periods of time). Then the finished GPUs need to go through QC, testing and retail packaging, before being shipped to distribution centers and then shipped across the globe. This is, at best, a 4-6 month process from a wafer starts processing till the finished product reaches retail shelves. Assuming everything goes smoothly. So, even with
wildly unrealistic base assumptions, Nvidia would in a best case scenario be able to launch the 4000 series on N3 somewhere in the early February - late March 2023 timeframe. With any kind of realistic assumptions, we'd be talking H2 2023, if not later. That's a full year's delay, not "several months".
