What has changed at AMD that they can put things on the market so aggressively?
1 (more) thing: Apple.
And the complete alliance with TSMC and all other companies that share the costs together.
At GlobalFoundries, AMD was the most important customer for advanced processes. So AMD had to pay the most development costs.
What about TSMC?
First comes Apple, in a month for example TSMC will start production for 5nm (!), A 2nd generation EUV process, for the iApparata that I think will be presented in Sept. 20. Apple has so much money that they let TSMC make an optimized process, which no other customer gets. But yes, knowledge gained for “Apple 7nm” can certainly also be used by TSMC for AMD.
Then – historically – comes Qualcomm, a company that is easily a few times as large as AMD. But Qualcomm always plays just a little lijper than the competition, and gives Samsung a share.
After this the elephant enters the room: Huawei. It was namely not Apple (purely for change), which single-handedly paid TSMC’s first 7nm-EUV process (7N +) and was the ‘leading’ customer, no: That’s Huawei! And they have enough money again; and moreover, unlike Apple, not listed short-sighted professional jammers, but somewhat like China “5-year plans” (Huawei’s employees are the shareholders, so they don’t go for short-term gains).
Then there is the company that wanted to take over the number 2 on the list, namely Broadcom. The average PC / Router / set-top box / WiFi card is full of that stuff, most of us will have several Broadcom chips at home.
Then come the ‘little ones’, some of whom also had a huge amount of money in one go: NVidia, MediaTek, Rockchip, Allwinner, Umicore (formerly Spreadtrum), even Intel (!) For the SoFIA modems: That kind of ‘smaller’ customers.
Those entire bubs have customers together, with much, much deeper pockets than Intel together the costs for 7nm (Apple, Qualcomm), 7nm-EUV (Huawei) and 5nm (Apple) paid. Because they have pumped so much money together, and also in a low (er) wage country (Philips, by the way, founded TSMC almost single-handedly in the 1980s!), TSMC has a new process every year. Intel and previously also GlobalFoundries simply have no money / people for that.
A little more in the background are the EDA / IP companies that are happy to participate: If TSMC comes up with a new ‘node’, Mentor (Siemens), Cadence and Synopsys will make ‘electrical’ design, simulation and testing as quickly as possible. programs for the new TSMC process. Apple may find some errors in that, it reports. EDA tool is being adjusted, Broadcom who wants to design something the next day benefits.
Same for the IP market: Ceva makes (designs) a nice DSP, which they make on TSMC 14nm, for example. The next day Apple wants him, they transfer a sum of money and rename the dastardly DSP to “AI accelerator”, and stick an Apple logo on it.
The day after Huawei sees that again, they think: Hey, we also need an AI accelerator! Huawei buys a – in practice – proven AI accelerator from CEVA (Apple’s public secret is CEVA, Huawei is a hypotethical example).
Then, after TSMC successfully ejects a node every year for years, where Intel has been failing at 10nm for 5 years now, AMD comes to TSMC and asks what’s on the shelf.
Well, TSMC and their partners (they also call it the ‘ecosystem’) has shelves full! Do you want a DSP? Do you want an ISP with a discount? An ARM Trusted zone module, already ready-tested? Stukkie electronic design software especially suitable for that node, from which Apple has already taken the errors? Maybe you would like some HPC libraries as well? And half the plant that suffices with “yield monitors” for 7nm from KLA Tencor, Apple no longer needs that – because Apple goes to 5nm. Would you also like to use AMD? Is there an EUV?
And there it is, Intel, totally upset. Have to reinvent the wheel, and pay for it, and then again in expensive countries like US, Ireland and Israel! All foundry customers ran away due to the failure of 10nm (Spreadtrum, LG, Nokia and Ericsonn probably, Achronix, Tabula …) who have no one to together accelerate innovation.
AMD does not know what they experience: Where first half of the money went into GlobalFoundries, which sometimes did not even come up with a working process (14nm I look at you!), They now buy an apple and an egg ( phone) a tested, working mass production process, which is on time and improved every year.
In fact, because the GloFo – which is a bit without modem – is now swallowing less money, money remains about at AMD !!! Well what can you do with it anyway? Hmm, maybe just accept more chip designers, and separate teams each time on the current node (7nm EUV, Zen 3 – probably from July 20 / Jan ’21), the upcoming node (5nm, H2 ’21 – H1 ’22) and the node after it (3nm, “scouts”) at the same time let work! Eat that, Intel!
[Reactie gewijzigd door kidde op 13 januari 2020 20:53]
–
