Balancing AMD’s Future on the Edge of a Silicon Wafer | Hacker News Reader

Balancing AMD’s Future on the Edge of a Silicon Wafer

nextplatform.com

41 pointsrbanffy8 comments

The article refers to orders of 50000 or 100000 CPUs as so large that AMD might have trouble fulfilling them. Is it plausible? How long does it take to manufacture them? That is, how many wafers per day can TSMC and/or Global Foundries process with a plausible portion of their capacity?

The demand is already booked at the fabs. This is an extra 50-100k on top of that.

Each chiplet (core complex die) on the AMD Epyc is 74 mm² plus the I/O die, which is built on a 12nm process is 125 mm². [0]

There is roughly 70k mm² of area on a 300mm wafer. Let's say for math's sake that after yield loss and so on, we end up with 37k mm^2 of chips that work; That's 500 chiplets per wafer.

Each CPU has 8 chiplets, so that's 62 CPUs per wafer.

So 50k CPUs is ~800 wafers and 100k CPUs is ~1600 wafers.

Cycle times for 7nm are ~80 days [1]

So you have at least an 80 day lead time, plus whatever time is spent waiting for open capacity.

TSMC capacity is said to be 130k wafers/month @ 7nm [2], so the volume needed is only ~1% of total monthly fab allocation. According to the same source, AMD already has ~20% allocation at TSMC, so it's 5% of AMDs total production for a single month for that order.

[0] - https://en.wikichip.org/wiki/amd/microarchitectures/zen_2#Co... [1] - https://semiengineering.com/battling-fab-cycle-times/ [2] - https://wccftech.com/amd-7nm-wafer-production-set-to-double-...

So this basically comes down to TSMC having some sorta advantage in chip design that no one else has discovered yet. And AMD was able to capitalize on that by not having its own chip maker, so when GF failed to produce good enough chips AMD just shopped around.

The AMD chiplet design is also a big advantage, they can mix 10nm and 14nm parts, they get much better yields then monolithic designs also read that the instructions per clock is also very good, the only missing things is frequency.

AMD is still locked into an agreement to buy wafers from GloFo that dates back to their spinoff, and the chiplet design allows them to use GloFo 14nm for the I/O die which helps to satisfy some of their wafer purchasing requirements.

It's a massive win all around.

Just to be a little more clear - I believe the original lock-in agreement with GF has expired, but they negotiated a new one. So they have an agreement about chip purchasing, but they aren't still in the agreement from the spinoff.

I have heard rumors that AMD had a closer relationship with TSMC than NVIDIA. I have found that a bit weird as NVIDIA seems to move some higher margin silicon despite vague claims that NVIDIA is hard to work with. These rumors show NVIDIA either moving or being pushed to Samsung 8nm [0].

This article brings up the interesting point that AMD is spread thin trying to manufacture CPUs, APUs, GPUs, and has products for xbox and playstation both last and current gen. It paints it like a problem, and I would agree that predicting demand is challenging in the best of times. But maybe the diversity of the lineup is the real drive of TSMC partnering more closely with AMD; TSMC is even willing to make special variations of its process for AMD as seen in [1].

I think its clear why AMD wants to partner with TSMC, but I was always wondered whats the incentive the other way around that makes AMD attractive to TSMC.

One possible point is also mind-share as AMD has been getting many headlines but as TSMC is B2B, so I find that unlikely as TSMC's main value add from AMD over other manufacturers.

Foundries are expensive business, and TSMC likely wants big commits to make sure its bets pay off. TSMC only cares about what they are paid and NVIDIA likely isn't passing on much of its margin, so AMD's money is just as good. Except its not just AMD's money, its also Sony's and Microsoft's money. I would be shocked if there weren't minimum volume clauses in the contracts for making those console chips, custom silicon is expensive and AMD probably wants to have a guarantee it will break even. The TLDR is AMD can feel safer in making bigger commitments to fab time from its console agreements and product lineup, and TSMC can feel safer that this capacity will actually be used.

[0] - https://www.techpowerup.com/269347/nvidia-geforce-ampere-gpu... [1] - https://www.pcgamer.com/amd-zen-4-specific-5nm-enhanced-node...

Tsmc needs to have customers for their products. In theory they could convince my company to build our chips with them (though I think we are happy with whatever process our current suppliers have which is probably in the 60nm range - obviously if I had insight into that side of the business I couldn't talk), but out column is only a few thousand chips a year. They would love to take our money, but they won't build a plant to make chips just for us. However if they think they can get AMD as a customer, that is enough to make a plant for as even worst case there will be enough volumn to pay off the loans (that is break even), and that lets them sell capacity to people like us who might be willing to pay a higher margin.

I would bet that Tsmc was courting amd for years. It is too expensive to build a foundry without some big customers committed. You can build a foundary for someone who won't order for a few years and you have to manage the risks of all the big customers (they might go bankrupt, they might never commit to...), so the more the better.

Reply on news.ycombinator.com