PowerColor's Next-Gen Radeon R9 390X Pictured

Seems its hard to keep all the info on these cards in any sort of real order. To the best of my knowledge, based on both official info and rumors:

The GPU known as Fiji (HBM all that jazz) is internally known as the Fury X. Whether it will keep that name in retail is still unknown to us plebes.
There are potentially a few boards which will *possibly* be the Radeon 390 X/XT/PRO/Ultra-mega/ePeen edition, whatever. These are Hawaii architecture respun on Global Foundries' Mo' bettah 28nm process (original Hawaii were TSMC) and paired with 8 or 4GB of GDDR5. The pinouts and power reqs should be the same, or less in the case of the power draw, so the same PCBs would work just fine for reference boards.

As for that PowerColor board, reading the article they state that it could change some (aesthetics) as it won't be available for a couple of months still.
 
The rumor has always been a fab change for Hawaii as this will not be the TSMC 28nm Hawaii but Glo-Fo or Samsung fab which was said to offer 20-25% performance without the magic tricks .. it will take the 980GTX on as this is Hawaii's refresh we have all longed for..

So that leaves AMD two more cards to battle 980Ti and Titan X..
 
the process is supposed to be HP+, 25% is a bit of a stretch though.
 
If they could pull 20% out of just from a fab change then is see the 980GTX having a player to deal with.. also the new is die is said to overclock very high so 980GTX may lose that edge also.
 
if they can get 20%, then Antigua (380/x tonga refresh) should compete with the 980 vanilla.

Which means Grenada (390/x Hawaii refresh) should fly past the 980.
 
if the process increase of 20% is true, i would expect the 380x to compete with 980.
 
If these numbers hold up, and they seem reasonable, AMD is sitting on a massive performance lineup upgrade along with HBM to unveil at E3! :eek:
 
Hawaii was rushed to market because Nvidia forced there hand and why the big power draw and heat as I think the plan was to have Glo-Fo fab Hawaii but it was no where ready for 28nm and TMSC was and they did what they had to do..

I still think Hawaii has a lot more to offer as clock speed has been the hold up because of the heat and power .. I think this new 28nm fab is what Hawaii was to be in the first place as I will run cooler and faster.
 
Hawaii was rushed to market because Nvidia forced there hand and why the big power draw and heat as I think the plan was to have Glo-Fo fab Hawaii but it was no where ready for 28nm and TMSC was and they did what they had to do..

I still think Hawaii has a lot more to offer as clock speed has been the hold up because of the heat and power .. I think this new 28nm fab is what Hawaii was to be in the first place as I will run cooler and faster.

thats rediculous.

do you have any idea what the lead time on a tape out is?

they would have to do 6 months of work in 2.

tapeout means the silicon is complete, and the boards are going to manufacturing.

not physically possible.

unless the boards are already in manufacturing and assembly, there is no possible way to speed up the process.
 
I wonder if you can crossfire a 390x with a 290x, like the 7970 xfire with the 280x.

If it's a completely different process from GloFo I strongly doubt it.

thats rediculous.

do you have any idea what the lead time on a tape out is?

they would have to do 6 months of work in 2.

tapeout means the silicon is complete, and the boards are going to manufacturing.

not physically possible.

unless the boards are already in manufacturing and assembly, there is no possible way to speed up the process.

They might be plug in replacement for Hawaii and work on the same PCB?
 
thats rediculous.

do you have any idea what the lead time on a tape out is?

they would have to do 6 months of work in 2.

tapeout means the silicon is complete, and the boards are going to manufacturing.

not physically possible.

unless the boards are already in manufacturing and assembly, there is no possible way to speed up the process.


So you think they just woke up one morning and said we are switching fabs?

It's not AMD vs Nvidia on TMSC anymore..it's Glo-Fo/Samsung vs TMSC/Intel now..
 
thats rediculous.

do you have any idea what the lead time on a tape out is?

they would have to do 6 months of work in 2.

tapeout means the silicon is complete, and the boards are going to manufacturing.

not physically possible.

unless the boards are already in manufacturing and assembly, there is no possible way to speed up the process.

Tape out does not mean that "silicon is complete"...It means that the design is finalized and the design is being sent off to have the masking completed, which adds another ~3 weeks to 1.5 months depending on how many masking layers are needed.
 
Tape out does not mean that "silicon is complete"...It means that the design is finalized and the design is being sent off to have the masking completed, which adds another ~3 weeks to 1.5 months depending on how many masking layers are needed.

Correct. I will also add that "tapeout" or "design is finalized" means that the final GDS2 gets sent to the fab and then the fab builds the masks. Which yes, does take a few weeks to a couple of months. But if you wanted to switch fabs at this point you likely need to switch cell libraries. That could involve laying out the entire chip and redoing static timing. I think this is the extra time that The Mac was talking about.

Source: Silicon design engineer, 15 years experience, > 10 tapeouts
 
Correct. I will also add that "tapeout" or "design is finalized" means that the final GDS2 gets sent to the fab and then the fab builds the masks. Which yes, does take a few weeks to a couple of months. But if you wanted to switch fabs at this point you likely need to switch cell libraries. That could involve laying out the entire chip and redoing static timing. I think this is the extra time that The Mac was talking about.

Source: Silicon design engineer, 15 years experience, > 10 tapeouts

Someone with your level of experience is useful to have here. I have a question: as I understand it, two fabs could have the same transistor width, but would use vastly different fab methods. Would this mean that a chip design that was designed for Fab A's 28nm process could not just be dragged and dropped onto the Fab B's 28nm process?

I'm just wondering how realistic these rumours of a TSMC vs GloFo Hawaii really are.
 
Same transistor width means nearly nothing. The cell library is what matters and those are almost assuredly different between the fabs. Also, AMD likely uses custom cells for some logic and those would need to be redesigned/simulated as well.

Moving from TSMC to GloFo would mean synthesizing the RTL into the new cell library and going through the entire layout and static timing process again. There is merit to doing this with a proven design but it is still a lot of work and is damn expensive. Unless they get huge production cost savings and GloFo covers some of their upfront NRE costs it makes no sense to tapeout the same chip again on the same node at a different foundry.
 
if they were going to upgrade the functionality to GCN 1.2 it might be worth it as theyd have to redesign anyway.
 
AMD was already paying for fab time and not building anything as some agreement they had with Glo-Fo as they had to buy so many waffles as part of the agreement but using TMSC to build GPU's so switching will save them a lot of money and should turn a profit not having to pay both Glo-Fo and TMSC.
 
AMD was already paying for fab time and not building anything as some agreement they had with Glo-Fo as they had to buy so many waffles as part of the agreement but using TMSC to build GPU's so switching will save them a lot of money and should turn a profit not having to pay both Glo-Fo and TMSC.

I laughed at waffles. Also gonna see if I can slip that by anyone in a conversation at work.

"Waffle lot 27 is way down in the slow-slow corner. Yield is gonna be shit."
 
Back
Top