News GlobalFoundries unveils 7nm roadmap

If they can pull this off, could this allow AMD to actually get the jump on Intel for once? Will this also bring benefits for AMD's GPUs and allow then to get ahead of Nvidia? Or will GPU node size still lag well behind CPUs? I hope AMD, working with GlobalFoundries, can capitalise on this technology and give Intel and Nvidia more serious competition.

 

I agree. It'll be a firm back-handed slap in Intel's face if GF can pull this off. For way too long, Intel has been seen as invincible. First they give up fighting against ARM in the phone market, then they get reminded that nobody wants their microcontrollers either, and then this comparatively minuscule (with 11x less revenue) company may manage to out-pace Intel. If Zen turns out to be pretty competitive and manages to get the 7nm node before Intel, this could be like the Athlon 64 days all over again.


Anyway, I thought I remembered reading somewhere that 7nm is the theoretical minimum size for a transistor, where if you get any smaller electricity starts leaking too much. Perhaps this changes depending on what dopes the transistors, or maybe FinFET is an exception. But things are not looking so good. Whether or not 7nm is the limit, we're definitely getting close to it, and there's not a whole lot else that can be done to improve upon the latest CPU architectures without increasing wattage.

 

Much more likely they'll still end up as late or later then their competitors, in the mean time AMD will be stuck with a rubbish 14nm node while everyone else gets 10nm. Unless AMD want to use someone else for 10nm in which case they have to pay TMSC (or whoever) to use that and in addition pay GloFo not to use them (yes that is written into the terrible contract AMD has with GloFo).

 

Node size will mean nothing if Zen doesn't deliver the IPC goods. It's all about the architecture with CPU's and unlike GPU's throwing more transistors at the job doesn't help (consider that the R9 Fury X has 1.7B more transistors than the Intel 22 core E5-2699 v4).

 

How exactly do you think IPC is achieved if neither node size nor transistor count will help? If we ignore logical threads, the only way to improve performance is to find the right balance of pipeline length and add more instructions. In AMD's case, they have shorter pipelines than Intel and fewer implemented instructions, so yes, adding more transistors is in fact a solution for them.

The way I see it, the real problem at hand is x86 has no equivalent of big.LITTLE, except more of a focus on performance rather than efficiency. CPUs should be divided in half, where one set of cores uses low clock speeds and large pipelines, while the other half uses high clock speeds and short pipelines. Or maybe the CPU could be divided into thirds, where the last set of cores are used for very fast non-float calculations. Not all calculations are equal, and x86 kind of became this behemoth of an architecture containing more instructions than developers know what to do with. Meanwhile there are things ARM can do much more efficiently, simply because it lacks the burden of so much extra clutter.

 

big.LITTLE on a desktop is pointless, The main goal of big.LITTLE is power saving in mobile. Andriod has had big.LITTLE for nearly 3 years now, Andriod is still not setup to take full advantage of this. anandtechs own testing proved as much.

Everyone is on Windows, Which would need to be programmed to take advantage of big.LITTLE.

Efficiency on a desktop pc is not fun to me, Do not really care what it draws in power, I expect it to deliver on performance though.

 

Talking of windows, I'm sure MS updates will negate any and all advancement engineers and scientist can come up with. (plus add even more tracking, but that's another story.)