TSMC Roadmap Update: Will Be Ready To Take 7nm Orders By April 2017, Volume Production Scheduled To Begin By 2018
The silicon industry is nothing without the foundries that produce the brains for pretty much everything powered by technology; which is why the progress of these fabs in terms of technology and schedule is of immense importance to enthusiasts. We have just received word that TSMC, a leading pure-play foundry, will be ready to take orders for products on the 7nm node by as early as April 2017. This is, of course, a pretty significant announcement because it lets us know the approximate time by which we can expect GPUs and CPUs on the same node.
Roadmap update: TSMC’s 10nm process landing by year end, 7nm will begin trial production in 2017
At a domestic event, reports from the TSMC Research Unit have revealed the roadmap of the company for the next few years. According to senior officials, the company will be switching to 10nm process technology by the end of this year and will start 7nm trial production in 2017.The 16 FFC process (a much more refined version of the 16FF+) will also be landing this year. The 7nm process will allow it to offer chips with greatly increased power efficiency (think clocks of around 3.8 Ghz at 1V vcore) with the threshold voltage values as low as 0.4V. The operating temperature of these chips is also expected to be around 150 degrees.
The 10nm FinFET based process will be able to provide a 50% reduction in chip size as compared to 16FF+, as well a 50% increase in performance or a 40% decrease in power consumption (The 10% difference is due to how power efficiency scales across clocks). 7nm on the other hand (presumably with FinFETs) will offer a performance increase of around 15% or a decrease in power consumption of 35%. The transistor density, however, will be increased by 163%. This jump is not as impressive as the jump from 10nm because as I have stated multiple times before, there is a difference between the foundry’s marketed process and the actual physical process. Very roughly speaking, the 10nm FinFET process should be equal to Intel’s 14nm process and the 7nm process should be equal to (or worse than) Intel’s 10nm process. For a much more detailed comparison of the process nodes of Intel, TSMC and GlobalFoundries, go here.
AMD has already dropped hints in its roadmaps that it would be moving directly from the 16nm node to the 7nm node. But it of course, for that to happen, the foundries that it will source its designs to need to have a fully functional 7nm process. According to recent reports, GlobalFoundries is scheduled to begin trial production on the 7nm node by 2018, this is almost a full year later than when TSMC should be ready with its own 7nm process (volume production is slated to begin by 2018). Thanks to the amended WSA, however, AMD can choose to go with any foundry and is not limited only to GlobalFoundries as its fab partner. If TSMC is actually able to beat GloFo by a significant margin, it is entirely possible that the Sunnyvale chip manufacturer will decide to pitch the camp with TSMC.
In terms of the exact products, according to the information revealed so far, the AMD’s Vega 20 powered graphics card will be built on the 7nm process. Of course, GPUs are not the only thing AMD is working on. When talking about the 7nm architecture, Zen+ (the successor to Zen) architecture must be talked about as well. With the information released so far, the company should have at least one CPU/APU product on the 7nm node that will be powered by the Zen+ architecture in 2018. The AMD Starship CPU and the Gray Hawk APU are two such examples.
At this point, however, it is prudent to add two very important aspects to why a node launch date doesn’t necessarily correlate with product orders being accepted by TSMC as far as Nvidia/AMD are concerned. Historically speaking, 1) TSMC has always favoured orders from apple (much to the chagrin of AMD and Nvidia) and we have no reason to believe that it won’t continue to do so and 2) the node at the beginning of volume production isn’t usually mature enough for high-performance ASIC production (although this is valid for any foundry.)
Now let’s talk about Nvidia. Just like AMD, Nvidia has diversified its chip sourcing capacities away from TSMC and also has listed the help of Samsung. Considering the fact that Samsung’s nodes are approximately the same as GloFo’s (they aren’t as similar as Intel’s Copy Exactly philosophy but they are pretty darn close), it looks like neither chip maker will have any advantage where the foundry is concerned for the next few years now. Both will have access to TSMC and GlobalFoundries/Samsung process nodes. This is something that is good for the consumer because it means the competition will boil down to who designs the better chip and nothing else. Nvidia also has quite a few architectures left up its sleeves. The company hasn’t maxed out its Pascal architecture on the 16nm node for the mainstream market and its next generation Volta based GPUs should be announced by May 2017.