Vega

>>58307017
Vega is the first major layout change to the CUs since GCN :^)
This will help actually utilize all dem juicy TFLOPS :^^^^)

http://www.gputechconf.com/

May 8, Jen-Hsun will announce Volta

AYYMD will cry on that day

>>58307084
>announce Volta
>when they still haven't released full fat Pascal on the consumers
>when they still haven't released the 1080Ti
>when Volta is Maxwell 3.0

At first, I thought it was a reference to the retarded PCI-E powerdraw. Then I remembered it wasn't voltage. Then I thought about VRM issues on EVGA cards, but it's not really a general issue so eh.

And it's kinda retarded to talk shit about a new architecture that will have a die shrink, since it's almost certain Vega won't reach the efficiency, no matter how much they improve now.

>>58307176
>since it's almost certain Vega won't reach the efficiency, no matter how much they improve now.
Polaris is actually efficient my anon. Desktop got the shitty parts.
http://www.anandtech.com/show/10710/amd-announces-embedded-radeon-e9260-e9550
I don't like Jayz, but he got sent an amazing card:
https://www.youtube.com/watch?v=zWASNajSdpg

>>58307248
Yeah, there still are improvements planned for Volta which, even without much architectural improvement, should achieve higher efficiency if only with the die shrink. And the planned use of HBM, like Vega, which last I heard of was still planned with a 14nm process.

And now that I googled back, it seems that Volta won't have a die shrink, so it's actually possible for AMD to shoot first for once.

>>58307308
Both Volta and Vega shouldn't be getting a die shrink unless you want to beleive the 7nm shrink rumor. AMD however, should get a hold of a better process by not going GloFo.

>>58307176
In no way will Volta be manufactured at 10nm.
Not at the die sizes required for Desktop GPUs in 2017.
The earliest TSMC will have a mature 10nm process with more than 1.7% ready, will be 2018.

Also, Vega finally introduces variable width CUs, something nVidia did for quite some time.
This will save a massive amount of power, while increasing performance.
Vega also finally implements power gating certain CUs, maybe even single SIMDs/ALUs.

>>58307248
It actually pisses me off so much that AMD once again got fucked by GloFo's absolutely shit process.
I got myself a really well binned RX480 (does 1450MHz at 1050mV) and it uses way less than what most reviewers put out there.

>>58307364
Along with this, it should put AMD in a good spot to tackle Nvidia's dominance.

>>58307059
GCN is like Forrest Gump before he lost the leg braces

NCU is after

e-except not retarded...

>>58307424
>It actually pisses me off so much that AMD once again got fucked by GloFo's absolutely shit process.

>Implying you have any idea what the fuck you're talking about.
There is nothing wrong with 14nm LPP, and it is Samsung's process, not GloFo's.
AMD set very wide margins on what is considered a passable RX 480 die. Stock voltages are set generously high to account for the wide performance range of ASICs, because AMD chose to do no clock binning beyond this. This has nothing to do with the process, and everything to do with a business decision by AMD to virtually flood the market with their cards.

>>58307509
>oh god this is bad
>if nvidia doesn't sell enough GPUs they'll shit can me
>quick I need to find some angle
c-communism...

Reminder to report and ignore, even if /g/ is dead.

>>58307499
And Samsung's is shit too, obviously.
Just take a look at Apple's iPhone SoC on TSMC and Samsung.
>inb4 b-but apple released a statement saying there was nothing wrong.

Also, even though it's basically the same process, the yields across fabs can be completely different, even if you deploy the exact same manufacturing process.
There is a reason why there is an absurdly long certification process for critical security products in the semiconductor industry.
You can't just say fuck it and move production of a part to another fab.

>>58307687
>And Samsung's is shit too, obviously.
You're not understanding. It is the same process.

>Just take a look at Apple's iPhone SoC on TSMC and Samsung.
You feel for forum shitposting, clickbait, and simple lot to lot variation.
http://www.tomshardware.com/news/iphone-6s-a9-samsung-vs-tsmc,30306.html

The one consistent aberration being reported was with the Geekbench battery life estimation, not in actual rated battery life, and the result was not consistently appearing in other battery life tests. It was found to be an issue with Geekbench. It was discussed in a huge thread on Anandtech back when it was relevant. Other results, like one chip being 1c hotter than other go back and forth, and this is normal variation in any mass produced good. Some Samsung A9 phones used marginally less power, some TSMC A9 phones used marginally less. Neither lot of chip is the top binning running at their highest perf/watt settings. Each are set to the same nominal settings.

>Also, even though it's basically the same process, the yields across fabs can be completely different, even if you deploy the exact same manufacturing process.
Test wafers are run across all 4 facilities periodically to make sure they always yield consistently.
>You can't just say fuck it and move production of a part to another fab.
Yes, they can. That is the benefit to the deal between Samsung and Global Foundries.

You are very, very obviously an emotionally invested shitposting fanboy. I want you to calm down, stop acting like a 13 year old girl throwing a bitchfit on myspace, and calmly explain why you pretend to know what you're talking about when it comes to semiconductor fabrication. I want to know what compels the mind of a fanboy autist to just talk out of their ass and expect to not encounter someone more knowledgeable to call them out.
Give me an honest answer. Explain why you just pull things from your ass.

>>58307823
This is my main hobby and I spend so much time reading that I just respond with what I think I know. If I'm wrong then I'm wrong, but I won't be wrong next time.

>>58307017
Wow, AMD finally hired some marketing people.

>>58307084
Volta will only be for deep learning servers. It will not make it into the hands of consumers this year.