is TSMC going to do what Intel did for 22nm?

[Borealis7]

We have heard many things about 3D Tri-gate transistors, FinFETs, and other scavanged Alien Technology when Intel announced their 22nm process back when Ivy Bridge was released, but I believe we haven't "fully enjoyed" all the benefits that these technologies bring with them in terms of reduction in power consumption, heat and greater switching speed on the CPU side - core speed has been stagnant in Intel chips for 3 generations.
Now that TSMC is talking about future nodes and processes, i wonder what is going to be the effect of these technologies on GPUs.

GPU makers can choose a maximum power consumption number and stick with it, utilizing a decrease in gate leakage and increase in transistor density to cram in more transistors and thus increase performance, or make GPUs with roughly the same number of transistors as we have today, tweak the architecture, and design smaller-die lower-TDP and cooler running chips for the midrange.

If GPU makers do elect to utilize every Watt available to push performance in their high-end chips, what is the Overclocking headroom going to be like? likely it will suffer from lack of "electric oxygen" and we won't see 20-25% OCs we see today.

Core clock speed is steadily rising, 1GHz core is a common frequency these days, but memory speed is slightly plateauing with 6-7Gbps GDDR5 chips. Memory modules are not produced at the same process as GPUs so it could be a while before we see any major improvement there.

GPU makers are always competing for the performance crown, but i really don't want to get used to 95C operating temperatures as the norm.

what do you think?

 

[ShintaiDK]

When you talk about core speed, I assume you mean frequency. Because the core performance development is not stagnant.

TSMC only got 1 processnode for the future. And it ends with SOC. First 20nm 2D SOC then later 16nm(20nm) 3D SOC.

So I wouldnt put my hopes up for anything frequency related.

GDDR5 is last of its kind as well. Stacked memory will be the next big thing. And for 20/16nm GPUs, we will still see the same GDDR5 as today.

Core speed of a GPU doesnt say much as such either. IGPs already run at up to 1350Mhz. That doesnt make them very fast, yet they are still very power efficient. Its simply all about design. GPU makers will have to rely more on improving designs (As we can see with nVidias Maxwell). Rather than depend on node shrinks to make any relevant progress. But that increases R&D costs.

95C isnt the norm. Its only the norm on 1 chip used in 2 cards.

 

[Mand]

what is the Overclocking headroom going to be like? likely it will suffer from lack of "electric oxygen" and we won't see 20-25% OCs we see today.

I've never understood why that headroom exists.

If you can just turn up the numbers with even a modest overclock and not suffer any lifetime or reliability problems...why not just increase what's considered "stock" speeds?

Having so much headroom just seems like a design error at this point, one perpetuated to give us folks something to have a hobby on. I don't see why it's a good thing from a technical perspective.

 

[ShintaiDK]

I've never understood why that headroom exists.

If you can just turn up the numbers with even a modest overclock and not suffer any lifetime or reliability problems...why not just increase what's considered "stock" speeds?

Having so much headroom just seems like a design error at this point, one perpetuated to give us folks something to have a hobby on. I don't see why it's a good thing from a technical perspective.

Power consumption, heat etc. Also you assume there is no problems involved in it in terms of life and reliability. Thats just a false assumption.

 

[Will Robinson]

I think he was implying "given those obvious parameters" ie: H/P/N reliability etc why leave performance on the table?

 

[Borealis7]

because the percentage of customers who OC is infinitesimal at best, and it leaves the companies an option to release higher core frequency variants in the future and milk the cash cow.

 

[Haserath]

They have to mass manufacture thousands of chips with less time to tune each and yields to be maximized.

They compromise between speed, yield, power, temperature, size, noise, and versatility(blower vs open).

If every GPU could hit a 1500 core clock in the determined TDP but none could reach 1550, that would kill overclocking that generation.

 

[Haserath]

And stacked memory is coming to Nvidia's next generation.

It should improve the memory performance at least a generation's worth, if not two. The power efficiency is supposed to be tremendously better.

The Volta generation of cards might be the one's to get if games can use the,.