Tuesday, September 21st 2010

NVIDIA Names Fermi Architecture Successors

At the GPU Technology Conference (GTC), an annual event hosted by NVIDIA, the company named the next two succeeding GPU architectures to Fermi (the current generation architecture on which are based GeForce 400 series GPUs). NVIDIA's next major GPU design change will come in the form of "Kepler", probably named after the German mathematician Johannes Kepler. The only concrete details about this architecture is that chips will be built on the 28 nanometer silicon fabrication process, and that going by the architecture's double-precision GPU compute performance per Watt represented on a graph, NVIDIA expects Kepler to be 4~5 times faster than Tesla, and over twice as fast as Fermi, again, at double-precision GPU compute performance per Watt.

Kepler is slated for 2011, though which part of the year will it be out (since AMD's answer to Fermi isn't far away), wasn't revealed. Looking much further away into the future, much like Intel mentioned Sandy Bridge's successor (Gesher) way back when unveiling Nehalem, NVIDIA talked about Kepler's successor slated for 2013. This one is called Maxwell, probably in honour of Scottish mathematician James Maxwell, with expectations of no less than three times the double-precision computation power per Watt of Kepler. These chips will be built on the 22 nanometer process.

Source: PC Perspective
Add your own comment

69 Comments on NVIDIA Names Fermi Architecture Successors

#1
Roph
Perhaps it's being held off for 2013 to allow PSU manufacturers to catch up :laugh:
Posted on Reply
#2
Tyr.1358
PVTCaboose1337 said:
I overlayed an exponential growth onto the NVIDIA graph to prove utter and complete BS that is that graph. It is an obvious estimation scheme. No real numbers, I promise you!

http://forums.techpowerup.com/attachment.php?attachmentid=38083&stc=1&d=1285098900
There is nothing wrong with the curve. See attached raster image.

As you can see, once adjusted to follow a consistent scale, this chart is actually rather conservative. They may even surpass those deadlines, regardless of what ATI offers.
Posted on Reply
#3
Melvis
Wait wait wait >(since AMD's answer to Fermi isn't far away), ummm do they mean in DX 11 or? because i think this should be there other way around lol
Posted on Reply
#4
PVTCaboose1337
Graphical Hacker
phanbuey said:
I mean... yeah obviously its an estimation. Thats 2013. How would they have real numbers on a chip that is barely on paper based on a process that isn't available yet?

Graph is based on Moore's law double the power every 18 mos, if you keep doubling then its an exponential curve. I do agree with you that its a little steep tho.
I agree with your bit on Moore's law. That is why they chose an exponential curve!

IE: 2^X where X is years.

Tyr.1358 said:
There is nothing wrong with the curve. See attached raster image.

As you can see, once adjusted to follow a consistent scale, this chart is actually rather conservative. They may even surpass those deadlines, regardless of what ATI offers.
You are correct, there is nothing wrong with the curve, however the point of my graph was to point out that NVIDIA simply did an estimation, not show actual data.
Posted on Reply
#5
sneekypeet
Unpaid Babysitter
Isn't the point about the chart sort of obvious? I mean it is a projected figure, much in the way Folding@Home projects my monthly and yearly points. Based off what they have as information at hand it is a best guess scenario. Likely they or I in folding aren't going to make those exact numbers, but it is the best guesstimate at the time.

On the flip side it could be strongly exaggerated, as any company does to boost morale in their future products.
Posted on Reply
#6
Kantastic
I'm wondering if they'll slap a PhysX chip in there and charge a premium for it since by 2011/2012 PhysX would likely be history.
Posted on Reply
#7
Benetanegia
Those estimates are not delusional by any means. First of all, remember they are talking about computing and not gaming performance. As much as people like calling Fermi a flop, it is not, not at all. Sure it's not as good as they thought it would be in gaming and power consumption, specially GF100 which is horrible, GF104 is much better, but Fermi was made with one purpose and that is to serve 3 markets as best as posible while spending as less in R&D as posible.

Now, take a look:



^^Is or is not Fermi (Quadro 5000 and 6000) a lot faster (2x to 4x) than GT200 (FX4800) -and while we are at it- Cypress (V8800, V9800) at the one thing this particular GPU was mostly designed for? And if anyone actually saw the keynote webcast, that's not all. iRay is amazing, it's an almost "instant" Mental Ray renderer that's going to make our job soooo much easier...

Also you can always have a bad chip from time to time, at least GF100 IS faster than Cypress, a luxury that AMD would have liked to have with the bigger than G80 (more transistors), extremely hot and power hungry R600, instead of being 40% slower than G80 when AA was used. And the thing is that R600 was the spark that initiated a fire that ended up on Cypress, it's not too far fetched to think Fermi and sucessors can do the same. At least the start has been similar, but at the same time better: it started better than R600 and like with R600 a few months later a refresh that does very well (RV670 ~= GF104), and unlike RV670, GF104 is not made on a smaller process.
Like R600, Fermi has a very strong uncore designed to last 2-3 generations, unchanged. Hence right now the uncore is very bulky and the core/executing units are very weak. Now we know that the architecture initiated by R600 needed somewhere inbetween 800 and 1600 SPs in order to give it's best, if Fermi is similar we will have to wait until it has 2000 SPs before any judgements are made.
Posted on Reply
#8
DaedalusHelios
Steevo said:
Date of 480 review.

Mar 19, 2010, 11:30 AM


They show Fermi as 2009, they were how many months late? So to say they plan on 2011, means perhaps they will get product to reviewers in 2012?
That was due to fab issues that cannot fully be predicted no matter if your name is AMD, Nvidia, Intel, or Via. Although the funny part is that via would probably solve it fastest thanks to the place of fab being in the same place as the engineering team last time I checked. But of course via is all 65nm last time I checked aswell which is pretty far behind.
Posted on Reply
#9
PVTCaboose1337
Graphical Hacker
sneekypeet said:
Isn't the point about the chart sort of obvious? I mean it is a projected figure, much in the way Folding@Home projects my monthly and yearly points. Based off what they have as information at hand it is a best guess scenario. Likely they or I in folding aren't going to make those exact numbers, but it is the best guesstimate at the time.

On the flip side it could be strongly exaggerated, as any company does to boost morale in their future products.
If Moore's law is true today (it was until about a year ago) then I will say ok, maybe NVIDIA, but it really is an unrealistic estimate. Do a graph of past cards. Not exponential.
Posted on Reply
#10
theonedub
habe fidem
Do the numbers take into account the intentional capping of double precision performance Nvidia imposed on Fermi chips?
Posted on Reply
#11
KainXS
the difference in compute is probably because Nvidia when it comes to computing, Nvidia has almost always been ahead of AMD/ATI, but I would wait to see what AMD has in store first before I pass judgement, the architechture on the R870 is how old now, it hasn't changed much on the SP's themselves since the 2900XT was released which is probably why they are redesigning the SP's now, don't know what will happen but its suprising to see Fermi has a successor, a 768SP version of the current GF104 to replace the GTX480 would be more than enough to keep AMD at bay.

thats probably what kepler is
Posted on Reply
#12
ebolamonkey3
Tyr.1358 said:
There is nothing wrong with the curve. See attached raster image.

As you can see, once adjusted to follow a consistent scale, this chart is actually rather conservative. They may even surpass those deadlines, regardless of what ATI offers.
Your chart increases in 1 year period while theirs increases in 2 yr period.
Posted on Reply
#13
[H]@RD5TUFF
*looks at my 4 GTX 470's*

sigh too soon. . . .

Oh well , we shall see, my only request is that they run cooler, also we will have to see what the 6xxx series does.
Posted on Reply
#14
jmcslob
This seems to me to be more of a "We better say something..Anything or we're done" type of press conference...You know like what ATI used to do...Lol
Posted on Reply
#15
Bjorn_Of_Iceland
"omg, investors are being scrooges. Time for some fancy slides and up the optimism!" XD
Posted on Reply
#16
cheezburger
kapler will be feature 5120 cuda core/800TMU/256 rops and 1024bit bus that build on 22nm with only 484mm^2 die size. pretty reasonable spec for high end in 2012. this would definitely crush any of amd line.
Posted on Reply
#17
Athlon2K15
HyperVtX™
cheezburger said:
kapter will be feature 5120 cuda core/800TMU/256 rops and 1024bit bus that build on 22nm with only 484mm^2 die size. pretty reasonable spec for high end in 2012. this would definitely crush any of amd line.
pics or it didnt happen:D
Posted on Reply
#18
cheezburger
AthlonX2 said:
pics or it didnt happen:D
jen-hsun huang will grab it in right front of camera in end of this year after cayman launch
Posted on Reply
#19
afw
What does this graph actually mean ...??? is it about processing power or performance on the whole (including gaming)

The y-axis is Giga-flops per watt ... so it means kepler will consume less power and can do more processing ... rite ???

(wonder whats the Giga-flops per watt value of 5870 .. hmmm )

EDIT: So 5870 has a Teraflop value of 3 ... and TDP is 200w so that means the 5870 has a 'Giga-flop per watt 'value of 15 ... and equal to maxwell :confused: )

Im confused ... just show me some gaming bechmarks ... :laugh:
Posted on Reply
#20
pantherx12
cheezburger said:
kapler will be feature 5120 cuda core/800TMU/256 rops and 1024bit bus that build on 22nm with only 484mm^2 die size. pretty reasonable spec for high end in 2012. this would definitely crush any of amd line.
Unless of course, ATI want to make a stupidly expensive card as well.
Posted on Reply
#21
yogurt_21
dost mine eyes decieve me? fermi in 2009? did they forget their own launch date?
Posted on Reply
#22
TIGR
btarunr said:
Kepler is slated for 2011....
So start saving money in 2012, everyone.
Posted on Reply
#23
wahdangun
yogurt_21 said:
dost mine eyes decieve me? fermi in 2009? did they forget their own launch date?
maybe their refer to mock-up fermi version, with fake PCB


i hope they don't do that again when SI was launched
Posted on Reply
#24
Tatty_One
Senior Moderator
Steevo said:
Date of 480 review.

Mar 19, 2010, 11:30 AM


They show Fermi as 2009, they were how many months late? So to say they plan on 2011, means perhaps they will get product to reviewers in 2012?
Those dates/years may be in production terms, not released to retail.... IDK.
Posted on Reply
#25
Jizzler
Sonabitch. I'm a Kepler.

[loop]
Get a perverted feeling of glee when nVidia fails.
But don't want to see my namesake tarnished.
[/loop]
Posted on Reply
Add your own comment