videocardzAMD’s Greenland/Vega 10 GPU to feature 4096 Stream Processors

[ShintaiDK]

I have no care where AMD sits. If you want to wager with yourself, go for it.

So you are not willing to commit to your statement?

 

[Head1985]

If AMD launch polaris first and cover 7770-GTX980 market its clear winner for them.They will gain a lot market share.
IF GP104 uses GDRR5X or HBM2 it wont be out intill q4 2016.Same as vega 11.
Only if GP 104 uses old DDR5 they will have few monts lead, but its only 2 cards in 400-550 market.

 

[Silverforce11]

If AMD launch polaris first and cover 7770-GTX980 market its clear winner for them.

They need notebook wins desperately.

And Polaris 11 perf/w advantages will definitely help in the entry segment, as AMD has nothing against the 750Ti and 950 for people who have crap OEM PSU without PCIE power. AMD knows it too, hence they deliberately compared Polaris 11 vs the 950 in their bench/demo.

 

[railven]

So you are not willing to commit to your statement?

Very well committed. However, I don't have a need to wager anything.

 

[Timmah!]

So now they are uninformed buyers? :thumbsdown:

So when AMD champion performance/watt and low power. They are also ignorant? Or is it because they actually know what the broad user base wants?

Look around, lower power, mobile, smaller form factors. This is what sells. Not some dinosaurs with big cases and 300W card(s).

And who is talking about mobile?

 

[crisium]

Hopefully 4096sp is Vega 11 and there's something above it. But it may be Vega 10. For all we know Polaris 10 is 2304 for the full chip. 2560 is pure speculation.

Possibilities if Vega 10 is 4096 and Polaris 10 is 2304:

Vega 10 full 4096
Vega 10 cut 3584?
Vega 11 full 3072?
Vega 11 cut 2560/2688/2816?
Polaris 10 full 2304

Vega 11 cut and Polaris 10 full would be very close in SP count if this were the case, but remember that Vega 11 would have far more memory bandwidth, and it's possible it will have more ROPs as well so there's no reason it cannot be a full tier faster.

Now I hope not, as I would like Vega 10 to be ~4600-~5600SP, but you cannot dismiss the possibility of 4096SP Vega 10 on performance overlap alone since there could still be a sizable gap between full Polaris 10 and cut Vega 11.

 

[Hi-Fi Man]

Glofo 14LPP. TSMC 16FF+ got better electrics. A10 is TSMC 16FF+ as well.

With the WSA AMD dont have a choice like the rest.

Does TSMC's node actually perform better? Even if it does it's well known it's not as dense as Samsung's and a denser node means more transistors per square mm...

This is a more important question than the whole async debate I think. The last time one of them used another fab other than TSMC was when NVIDIA used IBM's 130nm node for NV4x I believe. Those GPUs didn't really clock high and weren't very power efficient. This fab change could be a pretty big deal.

Does anybody have any sources for AMD using 14nm LPP on their next GPUs and if TSMC's 20nm FinFET actually performs better?

 

[Blitzvogel]

AMD crucially needs to fill out each market segment with products as fast as possible. I agree that AMD desperately needs a new no-external power GPU to succeed the 7750 and the few R7 250X's that don't require external power. They are already beaten by the GTX 750, and of course no where near as capable as the GTX 950.

 

[el etro]

At least AMD has the 14LPP in GloFo, probably the best process they can contract today's.


AMD said they will use the process for a 4-5 years timeframe, maybe because 10nm process will not offer reasonable improvement over 14/16nm processes we have today. I think that with a four or five year timegap in enough time to get 7nm process(this one probably with new(and expensive!) transistor architecture according to rumors) with good available capacity and yields, maybe being just one year after smartphones and FPGA/HPC adopt it. Not that bad for us PC enthusiasts.

 

[Head1985]

Hopefully 4096sp is Vega 11 and there's something above it. But it may be Vega 10. For all we know Polaris 10 is 2304 for the full chip. 2560 is pure speculation.

Possibilities if Vega 10 is 4096 and Polaris 10 is 2304:

Vega 10 full 4096
Vega 10 cut 3584?
Vega 11 full 3072?
Vega 11 cut 2560/2688/2816?
Polaris 10 full 2304

Vega 11 cut and Polaris 10 full would be very close in SP count if this were the case, but remember that Vega 11 would have far more memory bandwidth, and it's possible it will have more ROPs as well so there's no reason it cannot be a full tier faster.

Now I hope not, as I would like Vega 10 to be ~4600-~5600SP, but you cannot dismiss the possibility of 4096SP Vega 10 on performance overlap alone since there could still be a sizable gap between full Polaris 10 and cut Vega 11.

I dont think polaris 10 2304Sp is full SKU.
2304SP is 36CU it should be 40Cu so 2560SP.
2304 is wierd number

Btw i think Vega will have 8X Shader engines(SE) and 128rops.Fiji have only 4x so max 64Rops because each SE can have max 16Rops.
That vega 4096SP SKU its how FIJI should looks like in first place.
Only by increase from 4x to 8x SE vega should be 20% faster than Fiji.ADD to that polaris architectural changes and Vega 4096SP should perform 30-40% faster than FIJI with same ammount of SP at same clock.

8X Shader engines:
4096SP
256TMU
128Rops
SO it should be 20-30% faster than TITANX(More at 4K because 128rops)

Still i think AMD need bigger SKU to compete against GP100.Thats why i think this SKu is just vega 11-Tahtiti successor.SKU should be around 330-350mm2 big same as tahiti.

 

[Mopetar]

AMD said they will use the process for a 4-5 years timeframe, maybe because 10nm process will not offer reasonable improvement over 14/16nm processes we have today.

I think it has more to do with the current 14 nm process still being relatively new so there's plenty of room for improvements as it matures. Also, after the delays from the 20 nm process that didn't pan out, I don't think either company is going to make plans to move as there's far greater degrees of uncertainty moving forward.

 

[The Alias]

I think it actually is Vega 11 because it would bode extremely well for Polaris 11 being faster than furyX. You have to understand this version of gcn to the last is likely to be like Maxwell was to Kepler with the added benefit of a node shrink.

 

[poofyhairguy]

Couldn't this be plenty of sps if there is more ROPs? Seems like Fiji had a ROP problem.



People sure expect a lot out of the first batch of chips on a new node.

 

[William Gaatjes]

I think its strange that after launching Fiji with HBM, they're skipping that product line with the polaris generation/arch and then bringing it back with Vega.

Well, like most people i am also interested in hyped up technology.
when i read about hbm2, it seems that hbm2 is not so much a lot faster, it also does not have the limitations that hbm memory seem to have. I do not know enough about it to go into detail except this link :

http://anandtech.com/show/9969/jedec-publishes-hbm2-specification

Perhaps, besides the price point, gddr5 memory is not such a bad choice for gpu's that can never make use of the full bandwidth of hbm.
When a midrange gpu can never process data fast enough, it is just a very bad financial and bad design decision to use hbm for a "midrange" gfx card.
It also allows for AMD to perform tweaks in the gcn cores of vega, to solve issues found in polaris.

Vega is going to be the beast that is bandwidth hungry.

Actually we do know based on simple logic. This is the chip using HBM2..which is for high-end skus only. They all used the nomenclature "a full 4096 sps" which means its maxed out...as in the word 'full'..So logic would tell us there isn't another chip with more SP's and that this is Vega 10, the big chip, project greenland, which has been talked about forever.

They wouldn't say FULL 4096 sp if it could have more and they wouldnt make such a big deal out of project "greenland" if it wasn't the flagship gpu.

Pretty simple.

I would think so too.

 

[Adored]

Fiji doesn't have a ROP problem, the problem is the 4096 SP's can't be saturated unless async is used. Vega will fix the issues with that meaning the same 4096 SP's will start off around 20% faster anyway. With clock increases and other IPC gains AMD's next gen could easily be 50% faster per SP.

 

[Head1985]

Fiji doesn't have a ROP problem, the problem is the 4096 SP's can't be saturated unless async is used. Vega will fix the issues with that meaning the same 4096 SP's will start off around 20% faster anyway. With clock increases and other IPC gains AMD's next gen could easily be 50% faster per SP.

i didnt say fiji is rop bottleneck.Its SE bottleneck.Fiji cant utilize 4096SP.8x SE will fix that.More Rops is bonus.
Btw 50% more performance per Sp is sci-fi.That will be HUGE.

 

[Adored]

i didnt say fiji is rop bottleneck.Its SE bottleneck.Fiji cant utilize 4096SP.8x SE will fix that.More Rops is bonus.
Btw 50% more performance per Sp is sci-fi

I didn't say you did.

How much more perf does Maxwell have per SP compared to Kepler without the node shrink? 980 Ti is 43% faster than the 780 Ti at 4K according to TPU.

http://www.techpowerup.com/reviews/ASUS/GTX_980_Ti_Matrix/23.html

 

[ShintaiDK]

I didn't say you did.

How much more perf does Maxwell have per SP compared to Kepler without the node shrink? 980 Ti is 43% faster than the 780 Ti at 4K according to TPU.

http://www.techpowerup.com/reviews/ASUS/GTX_980_Ti_Matrix/23.html

And 980TI got 2x the ROPs (48 vs 96). All Maxwell cards over Kepler got 2x ROPs. The 64 ROPs is a weak point in Fiji. And why something like a 390X gets so close now and then. Vega 10 should hopefully fix this with 128 ROPs.

 

[Adored]

And 980TI got 2x the ROPs. All Maxwell cards over Kepler got 2x ROPs.

And less TMU's. Compare it to the 980 if you want, it's still gonna be 40%+

I think we all know Fiji is a badly balanced architecture and I see no reason why AMD doesn't know it as well.

 

[ShintaiDK]

And less TMU's. Compare it to the 980 if you want, it's still gonna be 40%+

Not according to your link. And GTX980 got 64 ROPs at 1126Mhz+, while 780TI got 48 ROPs at 876Mhz+.

780TI=42Gpixels
980=72Gpixels
980TI=96Gpixels

390X=67Gpixels
Fury X=67Gpixels

Raja also said something about this at the PR event

 

[Head1985]

Also Maxwell runs at much higher clocks
780TI-900Mhz boost
980TI-1200Mhz boost

 

[Adored]

Yeah the 980 is even higher per cuda core, it has 40% less cuda cores and is still 15% faster at 4K compared to the 780 Ti.

I mentioned clock speeds as part of the 50% uplift.

 

[Head1985]

Yep but clock difference is 33% so maxwell have around 15% better "IPC"
Btw GTX980 have boost 1250-1260Mhz thats 40% difference.And 980 is 14% faster than 780TI.

 

[AtenRa]

390X=67Gpixels
Fury X=67Gpixels

So at 4K they have the same performance ?? :sneaky:

 

[ShintaiDK]

So at 4K they have the same performance ?? :sneaky:

Is that how you see bottlenecks as absolutes?

And I guess you forgot what Raja spoke about at the PR event.