Nvidia Fermi successor called Kepler next year

[ajaidev]

"Exciting stuff at the end of Jen-Hsun Huang's GTC keynote as Nvidia's CEO has presented a roadmap depicting the next two GPU architectures currently in the works at the Santa Clara company. Fermi's successors are known as Kepler and Maxwell and are scheduled to launch in 2011 and 2013, respectively."

http://www.tcmagazine.com/tcm/news/h...ming-next-year

If i recall correctly:

GTX 480:
SP: 1.35 Tflops
DP: 168 Gflops

GTX 285
SP: 0.708 Tflops
DP: 88.5 Gflops

radeon 5870
SP: 2.72 Tflops
DP: 544 Gflops

[Dimitriman]

I wonder if they will make Maxwell a huge chip or a smaller more flexible as AMD has been doing.

Kepler is most likely just a tweaked Fermi and 28nm.

EDIT: Also note the Fermi - 2009 HA HA HA good one again Jen-Hsun Huang

[RaZz!]

[...]

EDIT: Also note the Fermi - 2009 HA HA HA good one again Jen-Hsun Huang

loool

ok, so he's actually saying kepler @2012 and maxwell @2014

[Mats]

So if they're using codenames from great inverntors, how come they haven't honored James Watt? Oh wait..

Sorry, couldn't resist..

[Dark-Energy]

So if they're using codenames from great inverntors, how come they haven't honored James Watt? Oh wait..

Sorry, couldn't resist..

[Russian]

Honestly, if you try to compare the 5870 v 480 in computational power you're missing one huge chunk. The 5870's computational power is theoretical and hasn't been harnessed at all in the last generations even though they claim to have a higher 'flop' rate. In the end, the flops don't matter at all but rather how efficiently they utilize them.

[Olivon]

Maybe we gonna have something in 40 nm before Kepler ...

[iMacmatician]

Maybe we gonna have something in 40 nm before Kepler ...

GF11x?

[-Sweeper_]

time will tell if this is another fermi or NVIDIA's RV770, I'm hoping for the latter

[96redformula]

So if they're using codenames from great inverntors, how come they haven't honored James Watt? Oh wait..

Sorry, couldn't resist..

Very funny & clever. Made me laugh pretty loud, I thought I was going to have to explain to my wife why I was laughing, but she didn't hear .

[cegras]

Honestly, if you try to compare the 5870 v 480 in computational power you're missing one huge chunk. The 5870's computational power is theoretical and hasn't been harnessed at all in the last generations even though they claim to have a higher 'flop' rate. In the end, the flops don't matter at all but rather how efficiently they utilize them.

Is it because it is vapor potential, or simply unexploited?

[tajoh111]

Is it because it is vapor potential, or simply unexploited?

It a little bit of both for AMD. Although they have some unused potential, alot of it I assume is impossible to use unless for very specialized tasks. NV shaders at this point seem better to do GPU work as that has been a focus for the company for the longest time, far more than AMD anyway.

The CGPU power is much stronger in fermi compared to whatever is in firepro. Even a stripped down performs better than AMD latest and greatest by typically double and in sometimes pretty much exponential.

http://tech.icrontic.com/articles/re...-than-we-said/

It must be to be able to sell for 5 grand which is three times as much as $1499 for the firepro 8800.

There is just so much money to be found in this field that AMD would be foolish to leave it alone which it hasn't. But at the same time, their products just are not as strong, even with the paper computation advantage, thus they have to sell their pro products for alot less.

The only asswhooping NV has delivered for the last 3 years has been their professional products. They are simply a generation ahead and as a result, can charge a premium(I heard their driver support is also reason for this).

Heck even in games, which is AMD specialty this generation, we never ever see the double in FPS or triple that is the TFLOP advantage that AMD has.

The only time I have see this advantage remotely is some password application which is very very specialized and a task very suited for the shaders of AMD line.

This might change next generation, it might not, but it probably won't be done this generation coming in a couple months. Its going to take a huge overhaul.

[jtdigital]

my crystal ball says Intel will put a gpu on their cpu and Nvidia will put a cpu on their gpu, maxwell in 2013 will support direct 12, just in time for windows 8. 2014-2015 is a good time for a new rig.

[Chumbucket843]

Is it because it is vapor potential, or simply unexploited?

some of it is vapor potential as register file bandwidth only allows for ~2.2TFLOPs of sustained performance. even then you would be beating fermi by a minimum of 50%. then there is some untapped potential but it's limited to specific cases. truthfully, the shaders are very fast at graphics but gpgpu is a different story. ie compiler issues.

the reason why extra resources are there is to simplify the design. the shader architecture they have now is quite good imo, but im sure they know how to improve it. one of the major misconceptions people have is that units whether they be cores, alu's or whatever are designed for high utilization. this does not necessarily translate to a good gpu (see gf100.). the absurdly high shader performance isnt because games require it; its to save die space. nothing but furmark requires 99% shading, 1% other tasks. 3d rendering is not just shading. iirc r600 had a much lower shader to bus size and shader to tmu/rop ratio than later gpu's.

[Manicdan]

think that ati should try to go hybrid, like half shaders with high depth (maybe even higher for just simple stuff) and the other half for things that its relatively weak at?

[trinibwoy]

If i recall correctly:

GTX 480: DP: 168 Gflops

That's 672 Gflops.

[spursindonesia]

GF 100 chip in GTX 480 is only capable of 168 GFlops since it's intentionally crippled. But this particular product isn't the right one for example since it's not meant for HPC segment anyway. CMIIW.

[blindbox]

That's 672 Gflops.

DP performance is nerfed on consumer cards.

[saaya]

you got the numbers wrong ajaidev, GT200 is lower and fermi is much higher... at least when your comparing pro cards and not vgas, and this is what the graph is about...

GF100 C2070 Fermi
225W
DP: 515 Gflops

GT200B C1060 Tesla
190W
DP: 78 Gflops

radeon 5870
DP: 544 Gflops


tesla GT200B is around 190W which makes it 0.41 Gflops/Watt. it looks higher in the graph but i guess they didnt want to move the chip partially below the 0 line ^^

fermi GF100 is around 225W, that makes it 2.28 Gflops/Watt. surprisingly nvidia has GF100 hanging much lower, maybe they are using ACTUALLY achievable Gflops? or maybe they just want to make the next gen chips look better?

right now Keppler looks 5x as powerful as Fermi, but if you take the numbers and not the graphical distance into account, then Keppler with 5 Gflops/Watt is actually only about twice as fast as Fermi, and not 5x as fast as this graph makes it look

And Maxwell is then going to be 3x faster than Keppler, which sounds possible...

so overall there is a LOT of marketing bs here once again, whos surprised though, its nvidia... alleging a 5x boost over fermi with their next chip is just what they need marketing wise to get analyst confidence back, but clearly not possible...

nvidia would have to beat ati's transistor efficiency by a looooong shot AND go 28/24nm AND have a massive 5xxmm2 chip again to pull this off...
OR, they actually did split up their chip lines and have a dedicated gpu and a dedicated gppu line from now on... but even then i dont think they could pull off a 5x boost

but i havent heard nvidia announce this, and im sure they WOULD if they decided this as it would mean better perf/watt for both, their pro and their consumer parts...

and about keppler for christmas... hmmm where have i heard this again? a new nvidia card for christmas hmmmmm

So if they're using codenames from great inverntors, how come they haven't honored James Watt? Oh wait..

Sorry, couldn't resist..

hah, nice one
well instead of naming things after him, they maybe thought theyd honor him by designing their gpus in a special way ^^

nah thats not fair, their engineers are doing a good job actually given the goals they have to fulfil (gppu plus gpu in one chip)

[flopper]

It a little bit of both for AMD. Although they have some unused potential, alot of it I assume is impossible to use unless for very specialized tasks. NV shaders at this point seem better to do GPU work as that has been a focus for the company for the longest time, far more than AMD anyway.

The CGPU power is much stronger in fermi compared to whatever is in firepro. Even a stripped down performs better than AMD latest and greatest by typically double and in sometimes pretty much exponential.

http://tech.icrontic.com/articles/re...-than-we-said/

It must be to be able to sell for 5 grand which is three times as much as $1499 for the firepro 8800.

There is just so much money to be found in this field that AMD would be foolish to leave it alone which it hasn't. But at the same time, their products just are not as strong, even with the paper computation advantage, thus they have to sell their pro products for alot less.

The only asswhooping NV has delivered for the last 3 years has been their professional products. They are simply a generation ahead and as a result, can charge a premium(I heard their driver support is also reason for this).

Heck even in games, which is AMD specialty this generation, we never ever see the double in FPS or triple that is the TFLOP advantage that AMD has.

The only time I have see this advantage remotely is some password application which is very very specialized and a task very suited for the shaders of AMD line.

This might change next generation, it might not, but it probably won't be done this generation coming in a couple months. Its going to take a huge overhaul.

I agree with that assesment as it reflect the what has been.
Nvidia is a big mouthful that havent been able to compete and have products out so they do more of smokescreens to try hold their shareholders happy.

People tend to negate and delete portions of information that isnt good so this means if the company is doing badly then people tend to delete it does until its to late ofc.

Announcing future line of products, shows they have a lot of issues and they now also are afraid of the upcoming line from amd and they are really afraid.
Cant be a coincidence that amd spent time with presenting cards to aib and now Nvidia do a paper prediction of upcoming line?

Might be whoop ass can another year.

[duploxxx]

It a little bit of both for AMD. Although they have some unused potential, alot of it I assume is impossible to use unless for very specialized tasks. NV shaders at this point seem better to do GPU work as that has been a focus for the company for the longest time, far more than AMD anyway.

The CGPU power is much stronger in fermi compared to whatever is in firepro. Even a stripped down performs better than AMD latest and greatest by typically double and in sometimes pretty much exponential.

http://tech.icrontic.com/articles/re...-than-we-said/

It must be to be able to sell for 5 grand which is three times as much as $1499 for the firepro 8800.

There is just so much money to be found in this field that AMD would be foolish to leave it alone which it hasn't. But at the same time, their products just are not as strong, even with the paper computation advantage, thus they have to sell their pro products for alot less.

The only asswhooping NV has delivered for the last 3 years has been their professional products. They are simply a generation ahead and as a result, can charge a premium(I heard their driver support is also reason for this).

Heck even in games, which is AMD specialty this generation, we never ever see the double in FPS or triple that is the TFLOP advantage that AMD has.

The only time I have see this advantage remotely is some password application which is very very specialized and a task very suited for the shaders of AMD line.

This might change next generation, it might not, but it probably won't be done this generation coming in a couple months. Its going to take a huge overhaul.

and now they have priced there 6000 series so high that anybody will think again if that extreme expensive card is worth the additional money, the 5000 series has a much better price/performance ratio.

The fight between quadro and firepro will be for quadro 4000-5000 against firepro 8700-8800 exactly where the price and performance is comparable and the battle is.

[fellix_bg]

you got the numbers wrong ajaidev, GT200 is lower and fermi is much higher... at least when your comparing pro cards and not vgas, and this is what the graph is about...

GT200 performs DP op's in a dedicated ALU -- one per multiprocessor (30xDP vs. 240xSP). During DP processing the SP units are idling, so you have to factor this in the overall load-time TDP, which brings the GT200's DP-per-Watt ratio closer to Fermi than you think.

[570091D]

i am now able to see the future... i see that nvidia will produce another gpu... i see that it will be released 5 years late... i see that it will consume 5 billion watts... but this will not kill nvidia, rather demand for thier products will GROW! this will infuriate the fanbois, war is coming.

[[XC] Oj101]

So if they're using codenames from great inverntors, how come they haven't honored James Watt? Oh wait..

Sorry, couldn't resist..

Is your sig a redneck inpaling himself?

[STEvil]

Is your sig a redneck inpaling himself?

thats an avatar.

its sun rising above a building which conveniently does not look to be as such.