AMD will transition Ryzen and Vega to GloFo 12nm LP next year

[HEXUS]

Shrink is expected to boost performance by 10 per cent or so.

Read more.

[Platinum]

Still a low power optimized node then, will this still not be great for Ryzen / Vega?

[Platinum]

I also wonder if this would still have been called 14nm+ if TSMC hadn't announced there 12nm process.

[Corky34]

At risk of sounding really dumb here but if you don't ask you don't know and my Google-fu has been letting me down.

What makes a LP or LPP FinFET different from a *normal FinFET?

*Normal as in what Intel uses.

[Platinum]

Not sure it makes any difference, my understanding is LP stands for Low power, I could be wrong though

[Corky34]

Yea LP and LPP means low power and low power plus, the thing is i don't have a clue what makes them low power versus other types of FinFET's, is it the way they're designed, the material that goes into them, or something else.

[DanceswithUnix]

*Normal as in what Intel uses.

I wouldn't call what Intel does "normal"

Intel's process is very good at producing fast CPU parts, but pretty bad at anything else, as shown by the lack of foundry customers they seem to currently have. Not being optimised for low power parts wouldn't be a problem for Intel if it wasn't for their wafers being reported as way more expensive.

is it the way they're designed, the material that goes into them, or something else.

All of the above, including metal layers layout

[DevDrake]

i cant understand why they put a graphic card icon for the Vega GPU and the marketing icons for CPUs and everything else.
Looks not professional for me.

[me-yeah]

Yea LP and LPP means low power and low power plus, the thing is i don't have a clue what makes them low power versus other types of FinFET's, is it the way they're designed, the material that goes into them, or something else.

leading performance https://www.globalfoundries.com/news-events/press-releases/globalfoundries-introduces-new-12nm-finfet-technology-for-high-performance-applications

[deksman2]

This LP however is not low power anymore.
The 12nm process seems more performance oriented this time around.

[HW90]

I'd imagine they're comparing to 16nm finfet because it was considered to be higher performing than their 14nm process

[Corky34]

leading performance https://www.globalfoundries.com/news...e-applications

That's even more confusing, they call LP "leading performance" and LPP "low power plus".

[DanceswithUnix]

leading performance https://www.globalfoundries.com/news...e-applications

Nice, I hadn't spotted the name change!

So, if a Ryzen can go 10% faster, does that make much difference to benchmark tables.

[scaryjim]

... So, if a Ryzen can go 10% faster, does that make much difference to benchmark tables.

Looking at the Ryzen 5 review, not really. The difference in ST performance between Ryzen and Kaby Lake is much bigger than that, and Ryzen is already a better value proposition in MT. The difference is marginal, and doesn't change the overall picture - i.e. Intel is still well ahead in ST workloads but Ryzen is better for MT workloads.

[deksman2]

Not sure it makes any difference, my understanding is LP stands for Low power, I could be wrong though

Yea LP and LPP means low power and low power plus, the thing is i don't have a clue what makes them low power versus other types of FinFET's, is it the way they're designed, the material that goes into them, or something else.

Actually, 12nm LP process that AMD will switch to for Ryzen+ and Vega refresh next year apparently is NOT low power.
There are differences between the nodes it would seem as this process is based more on the process that IBM is working on (7nm) which would technically allow a CPU to run at 5GhZ and the one that AMD will be using from GloFo for Ryzen 2 (however, for Vega or Navi at 7nm, it seems that AMD is looking into TSMC as Glofo process is not particularly suited for higher clock speeds - not sure how the 7nm process might work for GPU's, but it could easily be different for CPU's - GPU's might end up with lower quality yields if done on Glofo process, and due to some issues AMD experienced in this exact department, they seem to looking to TSMC for their GPU's on 7nm).

AMD's main issue right now in terms of power consumption for Vega is rooted in too high Voltages and possibly core clocks.
The manuf. process is not suited for high clock speeds, and AMD sets too high voltages to increase yields.
Undervolting Vega 56 managed to drop power consumption to just at or lower than 1070 levels which also allowed Vega 56 to reach and maintain its advertised boost clocks (otherwise it would throttle and guzzle down much more power than it needs), and surpass 1070 easily in performance.

Similar thing happened with Vega 64, but because 64 is clocked at higher core settings, it seems to be putting more strain on power consumption.

It was also noted that Vega usually benefits more from HBM overclocking as it seems to be bandwidth starved... and also overclocking the HBM also resulted in minimal increas in power consumption (about 5W increase for every 100-150MhZ).

[Corky34]

Yea they increased the voltage of Vega quiet a way past its peak efficiency, who knows whether that's for yields or performance, my guess is yields as the increased voltage didn't net them a big gain in performance.

I'd still love to know how LP and LPP differ structurally, i guess that sort of info isn't something they want to talk about though as things like gate sizes, drains, sources, and other structural differences isn't something you want your competitors to know.