Amd Volcanic Islands details

[Kersh]

Maxwell this year less than 7 or 8 months later than the gtx and titans? and 2x titan power?

I don't simply believe that, maybe for February or something a card a little faster than titan and again limited supply and 1000€ could sound reasonable but I don't see much more.

[Tao~]

We don't even have authentic numbers for the upcoming RedTeam part and people are already talking how the next GreenTeam product is faster . Fanboism is on a new high ?

First maxwell will be released on 28 nm .... and this year

Maxwell is definitely not coming this year. Maybe we'll see the performance part (GM104?) on 28nm in 1H14 if the delays with 20nm persist , however aren't chip designs inflexible wrt manufacturing process ? I mean if nV started designing Maxwell for 20nm 2 years back then they can't suddenly scale their design to 28nm because 20nm isn't available in time - a lot of aggressive optimization on the design means they have to redo the whole process for each node.

[AbortRetryFail?]

It did not ship any meaningful amount of GK110s to Oak Ridge until late Q3 '12.

My mistake, then

My understanding was it fully came online by October, and was ranked #1 in November ...

[PedantOne]

Maxwell this year less than 7 or 8 months later than the gtx and titans? and 2x titan power?

I don't simply believe that, maybe for February or something a card a little faster than titan and again limited supply and 1000€ could sound reasonable but I don't see much more.

kepler came with midrange cards at first, maxwell will be the same. Maxwell is Very special GPU with sophisticated power management and great performance, by the Nvidias materials.

[zalbard]

First maxwell will be released on 28 nm .... and this year

TSMC will not be able to produce GPU @20nm until the last quarter 2014 (Supply problem because of Apple SoCs). AMD and nVidia have two choices:
- waiting;
- to turn an another foundry.

Your proof is lacking so I will go ahead and disregard your opinions which you are trying to pass over as facts.

[boxleitnerb]

20 nm is extremely expensive right now. I have it on good authority that a wafer costs about $20.000 right now. That makes 20 nm impossible for the foreseeable future.

[Alex-Ro]

how much does a 28 nm waffer cost right now?

[boxleitnerb]

I don't know for sure, but my guess would be $5000.

[LordEC911]

My mistake, then

My understanding was it fully came online by October, and was ranked #1 in November ...

Yeah, they used Fermi back in Dec 2011 since Kepler was no where near ready. They upgraded 10 cabinets, called TitanDev, to test out their code, CUDA.
I'm sure they received a number of early/ES GK110s at some point but there really is no documentation of that.

Oakridge supposedly received ~32 K20x at the beginning of Sept.

The article states they were planning on receiving about 1k more that week, I would assume it was a couple thousand before the 13th of Sept.

They started installing K20x for the Titan upgrade Sept 13th and finished sometime in Oct.

They didn't finish acceptance testing until 11th of June 2013, since Cray had to repair all the mobos.

Edit- Ahhh... don't talk about wafer pricing. Highoctane will go crazy if you imply a new node is more expensive per wafer than an old node.

[Dimitriman]

90% of people talking up Maxwell only have this slide in their minds:


Not much is known about the product, to say it is 100% faster than Titan, and even worse, at 28nm... is borderline stupid.

PS: Btw the pretty slide above has been revised to a much more down to earth version below:


I think we should get back to talking about Vulcanic Islands!

[PedantOne]

no, i have many more info about maxwell forget this slide

[paulbagz]

Well then spit it out

-PB

[boxleitnerb]

Seconded

[Fottemberg]

Your proof is lacking so I will go ahead and disregard your opinions which you are trying to pass over as facts.

I have my chinese sources.

For the moment, take this article: http://www.wantchinatimes.com/news-s...00001&cid=1206
TSMC's 20nm is reserved for Apple until 3Q2014, but I can't tell anymore about this theme.

[Iconyu]

90% of people talking up Maxwell only have this slide in their minds:
Not much is known about the product, to say it is 100% faster than Titan, and even worse, at 28nm... is borderline stupid.

PS: Btw the pretty slide above has been revised to a much more down to earth version below:
I think we should get back to talking about Vulcanic Islands!

These charts don't say Maxwell is 100% faster, just that it produces around 50-100% more DP Gflops/watt. The second chart isn't more realistic it's just a more even but less accurate scale, because Volta is on it now. Instead of promising between 14 and 16 Gflop's per watt, it's now between 8 and 16. There is nothing about it that makes Maxwell look less powerful than the previous slide.

[nossy23]

These charts don't say Maxwell is 100% faster, just that it produces around 50-100% more DP Gflops/watt. The second chart isn't more realistic it's just a more even but less accurate scale, because Volta is on it now. Instead of promising between 14 and 16 Gflop's per watt, it's now between 8 and 16. There is nothing about it that makes Maxwell look less powerful than the previous slide.

To me the change in slide suggests that we should not expect Maxwell to be between 14 and 16 Gflops. I mean, also look at Kepler, in reality the Geforce Titan is only 4.5 Gflops which is almost the lowest possible performance according to that slide (4 to 8 Gflops). So my guess is Maxwell will be 8 Gflops or slightly more.

[zalbard]

Btw the pretty slide above has been revised to a much more down to earth version below

They are about the same. Fermi is still ~2x, Kepler is still ~6x, Maxwell is still ~16x.
They just had to rescale the graph.

What's more important is that these are DP FLOPS, and they do not equal gaming performance.

I have my chinese sources.

For the moment, take this article: http://www.wantchinatimes.com/news-s...00001&cid=1206
TSMC's 20nm is reserved for Apple until 3Q2014, but I can't tell anymore about this theme.

Do you even read your own links?
All it says is,

TSMC aims to launch its 20-nm wafers as soon as possible and before its competitors because it hopes to obtain orders from Apple for its ARM microprocessor

which has little to do with your empty claims.

[Lanek]

Dp Flops / watts, is what they are trying to aimed for, but we know Fermi have not end as it what promised ( even the first Tesla C20 was really under the promise made with only 448Sp and 0.515Tflops DP )..

Nvidia since then really like use thoses slides, but really, it is really hard to have a real figure based on it..

And i agree with you on the Apple thing... its just financial analyst to like push everything as they can.. Nobody know exactly what is happend with 20nm right now.. If we follow them, Samsung have allready a contract with Apple for the processors after the one produced by TSMC ( nobody knows yet if it is really the A7 or not, or maybe just the Iwatch processors ( why Apple will transfer the production of thoses processors to TSMC for 1 series, then come back to Samsung in 2015... ).. so Apple is allready not happy with TSMC ?.. You know financial peoples like more the rumors of enthusiast gamers and overclockers, because rumors means money: the stocks go up and down.

[Fottemberg]

Do you even read your own links?
All it says is,

which has little to do with your empty claims.

The deal is done. They will publish the notice in a little bit (some weeks). Wait and see.

[informal]

Although this is a topic about AMD VI GPUs all I can read is NV Maxwell this,NV Kepler that...

[Iconyu]

Although this is a topic about AMD VI GPUs all I can read is NV Maxwell this,NV Kepler that...

Isn't that AMD's precise problem? If the card costs more than a certain amount it'll cause people to consider waiting for Maxwell instead of comparing it directly to Kepler. And if you're comparing a product to an imaginary card some ways off the real card almost always suffers.

[Flanman]

I think what he is getting at is the fact that things are off topic, I'm checking this thread for news on vi, not Maxwell.

[highoctane]

Edit- Ahhh... don't talk about wafer pricing. Highoctane will go crazy if you imply a new node is more expensive per wafer than an old node.

Way to go trying to drag my nic into something for absolutely zero reason, great wtf moment.

I'm not involved in your game of conjecture.

[LordEC911]

Way to go trying to drag my nic into something for absolutely zero reason, great wtf moment.

I'm not involved in your game of conjecture.

No problem at all...

40nm wafers are more expensive than 65/55nm wafers...

Ok, how much more .01, $10, $100, or maybe $200000?? I mean we can assume costs are what we want to imagine them to be unless somebody slipped the invoices out to somebody who leaked real cost its all assumed.

Short term memory?

[Fottemberg]

I'm just saying that we will not see AMD/nVidia GPU @20nm until the 3Q2014 (if we are lucky, 2Q2014).