ARM grabs TSMC's 3D FinFETs for future 64-bit PC brains
Will it be a RISCy move for the v8 family?
ARM says its 64-bit ARMv8 processor architecture is a real contender for servers and PCs. But without an appropriate process from major fab partners to etch the chips, the design doesn't matter all that much.
That's why an agreement between the stewards of the ARM designs and Taiwan Semiconductor Manufacturing Corp is vital if the ARM collective is going to retain its position in smartphones and tablets while giving Intel a run for its money in data centres and desktops.
In a statement, ARM and TSMC said they have inked a deal to extend their silicon die technology beyond 20-nanometer-sized transistor gates. The financial details of the new agreement were not divulged. The two companies had agreed in July 2010 to collaborate on baking wafers with 20nm features.
All they did say was that the renewed partnership would bring future manufacturing processes to ARMv8 derivatives and to the Artisan tools that companies use to modify ARM's blueprints and optimise them for specific foundries and processes.
TSMC will be tuning its three-dimensional double-gate field-effect transistors (FinFETs) so they can be implemented with 64-bit ARMv8 processor designs. As ordinary planar transistors have shrunk, they have become relatively poor switches, leaking current even when they are in the off position.
This current leakage is a very big problem and is one of the reasons why chips suck so much juice and emit so much heat as they operate. This leakage is mitigated by going 3D, as Intel has already done with its 22-nanometer Tri-Gate process used with the current "Ivy Bridge" Core and Xeon processors. In fact, the Tri-Gate approach by Intel is an example of a FinFET design, albeit one with three gate surfaces instead of two.
Every electron out of the water, there's a fin in the sea!
The fin in a FinFET is just that: a vertical fin coming out of the silicon that presents more surface area than a planar transistor design and therefore does not have high current leakage at 20-ish nanometer sizes, which is very likely to hold for a few more future process shrinkages.
Both Intel and TSMC have been showing off various FinFET designs for the past decade, and Intel was the first to commercialise the idea. The key point with the various FinFET approaches is that the threshold voltage - the point at which the gate switches from one binary state to another - for the transistor is lower, too. So you get less leakage and lower voltage operation. And as a bonus, the lower you drop the voltage on the chip, the lower the transistor switching gate delay, which can increase processing speed.
You can see why TSMC, Intel, and others are keen on adding FinFETs to their processes. If the ARM collective is to keep its lead on Intel in the smartphone and tablet spaces and take on Intel in servers and PCs (or whatever is left of the PC market some years hence), then ARM Holdings is going to need to work with fabs to get processes ramped up to keep pace with Intel's factories.
Back in April 2010, TSMC said it would skip 22-nanometer tech and go right to 20nm in the second half of 2012. That smaller spin was planned to use planar transistor designs with enhanced high-K metal gate, strained silicon and low-resistance copper ultra-low-K interconnects; the company said at the time that FinFETs will be feasible at 20nm, but made no commitments on it.
From the ARM-TSMC partnership, we can now see that FinFETs are not coming out until after the 20nm chip phase. For TSMC, the next jump is to 14nm. TSMC has had volume issues with its 28nm manufacturing processes, impacting shipments for Advanced Micro Devices and NVIDIA among others.
So-called "risk production" of 20nm chips (a kind of beta test for the fab) started at the end of 2011 at TSMC, and the silicon will go into full production in 2013. The expectation is for a second-generation of FinFETs to come out during the 14nm phase, which will go into risk production in 2014.
That's about a year behind Intel. So the design benefits between x86 and ARM will have to make up the difference in the manufacturing gap. ®
You're both missing the point
The big advantage of ARM is not architecture, as there are other RISC processors out there that can beat it at its own game.
The big advantage is the ecosystem, where one can buy an ARM chip from just about anyone, and anyone can join and be fabless too.
That drives the prices down, turning the market into something Intel does not want to be a part of, as they have demonstrated in the past leaving markets when margins became too low.
Your comparison between a core iN and an ARM core is not adequate, the x86 core in the Intel CPU has a bajillion more features (x86 decode, SSEs, AVX, specific decode/encode functions, etc.) and wider computing units on top of a lot more cache and having 64bit support. And they're way more than twice as fast, even if pure DMIPS is about a factor 2 between a core i5 quad and an arm cortex a9 quad like the iPad cpu, you have to remember that's only integer performance and hardly representative of overall performance.
The big reason the Intel cores win that round is because they're much wider and consequently have more IPC than the ARM core which is really really tiny. You could for the purpose of benchmarking include a wider Integer Unit in the ARM core and get exactly the same DMIPS/clock as the Intel, but the CPU would still be quite inadequate.
Either way, there is NO competition between 22nm and 32nm chips, since the first cost MUCH more to produce at the moment, and Intel's process is still far from maturity, as the first Ivy launch demonstrated - let's wait till they get it right and we'll see exactly how far behind TSMC is.
Plus, as the number of bleeding edge -capable foundries shrinks, I wouldn't be surprised to see TSMC emerge as a winner and end up with more fab capital than Intel, which it's worth reminding, is only worth about 20 new fabs right now ( as the costs rise for every additional process - at the moment at least - Intel could very well end up facing too high a capital cost to remain one node ahead).
Fab Less...Isn't that Welsh for 'Awesome'
"...the design benefits between x86 and ARM will have to make up the difference in the manufacturing gap"
I can't imagine that will be too difficult - the x86's internal CISC->RISC translation must be quite a bloat, adding latency to branches, and code density will always be crap thanks to the register non-orthogonality, no matter how clever internal register renaming and other workarounds get at optimising the micro-instructions thereafter.
I was talking about mobiles and tablets (if you wanted a low power supercomputer you certainly wouldn't use x86 cores or GPUs but a purpose designed ARM core similar to the undisputed #1 BlueGene/Q in the Green500). What you'll see next year is fast Cortex-A15 cores paired up with tiny power efficient Cortex-A7 cores. When you don't do anything CPU intensive it will automatically switch to the A7 core, thus saving power.
ARM already supports up to 8 cores with the A15, so extending that to 16 or more doesn't seem too far fetched. Calxeda designed their own interconnect between chips (like Hypertransport) which is different and more difficult.
But what I really meant with designing cores much faster is that due to the inherent simplicity of RISC it is possible to create new CPU designs with less effort. And it's not just ARM, Qualcomm, Marvell, and AMCC doing the same. This would not be possible if it was x86.
ARM don't make the money to sub TSM for a new plant or really for much of anything. So this must either be PR or ARM have given TSM an exclusive, limited time offer.
Inquiring minds wish to know which it is!