Nvidia Maxwell - nVidia are moving my cheese

[whitedragon101]

nVidia have been posting this graphic for quite some time showing Kepler at 4-6 Gflops per watt and Maxwell at 14-16 Gflops per watt. That is amazing. I have been imagining a doubling of frame rates and openCL performance that makes your eyes bleed etc

Now nVidia have given Anandtech this image.

http://www.anandtech.com/show/7515/

Which shows Kepler at 4-8 Gflops per watt and Maxwell at 8-16Gflops per watt.

It seems nVidia have moved my cheese as that haven't even left any daylight between them. Kepler finishes at 8 and Maxwell starts at 8. That means they could launch Maxwell at 8 Gflops per watt (hopefully not though) and we wouldn't see much change

Also the only thing I can find about Maxwell architecture is unified memory which as per the Anandtech article is about making coding easier. I hope they have some performance improvement magic going on too.

Any thoughts?

 

[Sun Baked]

Read your chart again, nothing changed expect them going from a linear scale on the first to a power of 2 scale on the second.

Which really shows that, yes they are pretty much on a straight line doubling of performance per watt every generation.

 

[whitedragon101]

Read your chart again, nothing changed expect them going from a linear scale on the first to a power of 2 scale on the second.

Which really shows that, yes they are pretty much on a straight line doubling of performance per watt every generation.

Hopefully you are right. But changing the scale and making the icons bigger means that reading the first conveys lowest performance of Maxwell is 14 Gflops per watt and the second could be saying lowest performance of Maxwell is 8 Gflops per watt.

 

[leman]

Hopefully you are right. But changing the scale and making the icons bigger means that reading the first conveys lowest performance of Maxwell is 14 Gflops per watt and the second could be saying lowest performance of Maxwell is 8 Gflops per watt.

Not nessesarily. I think the top box is what counts The reason for the logarithmic scale is simple: otherwise you would need make the plot 2 times bigger to fit the Volta on it.

 

[cirus]

Have you even looked at the number?

GTX titan - ~1500 DP GFLOPS @ 250W = 6 DP GFLOPS/W.

Its even lower for lower clocked professional products.

 

[john123]

Any thoughts?

Yeah, my biggest thought is there is no guarantee whatsoever you'll see Maxwell in a MBP. It's as clear as the day is long that Apple would like to get off the dGPU chain. Despite what Maxwell will offer, it would not surprise me if, given the 40% projected improvement that Broadwell is supposed to bring, we've seen the last generation of discrete graphics.

 

[ssn637]

Yeah, my biggest thought is there is no guarantee whatsoever you'll see Maxwell in a MBP. It's as clear as the day is long that Apple would like to get off the dGPU chain. Despite what Maxwell will offer, it would not surprise me if, given the 40% projected improvement that Broadwell is supposed to bring, we've seen the last generation of discrete graphics.

Hasn't Broadwell been pushed back to the end of 2014? If so, couldn't we hope to see a Maxwell alongside the current Iris Pro 5200 in the high-end rMBP 15" sometime next quarter? I can't help wondering if the current 750M is a stop-gap solution. After all, the difference between the dedicated and integrated graphics in the 2012 model hasn't been seen this year.

 

[john123]

Hasn't Broadwell been pushed back to the end of 2014? If so, couldn't we hope to see a Maxwell alongside the current Iris Pro 5200 in the high-end rMBP 15" sometime next quarter? I can't help wondering if the current 750M is a stop-gap solution. After all, the difference between the dedicated and integrated graphics in the 2012 model hasn't been seen this year.

Q4 is just the leaked roadmap for Broadwell-K. A lot's up in the air between now and then. If Maxwell seems early and next-gen Broadwell seems far off, what you're suggesting seems possible...although I bet it would make the execs at Apple wince. They seem pretty dead-set on getting rid of dGPUs.

 

[dusk007]

I read some rumors that they will increase ALU count primarily in Maxwell. So it will push more flops in theory like the VLIW5 arch. used to but what the actual real performance improvement will be is still up in the air. Could be like VLIW5 lots of theoretical crunching power but not nearly as much practical.
One problem that only Volta seems to address is also memory bandwidth. It is still only GDDR5 so in some cases wider busses would be necessary to really feed a twice as fast GPU. That takes DIE space (not that much of a problem with 20nm) and probably extra power which means efficiency will not likely increase here.

I have my doubts about these claims. Usually a new process nodes only offers about 30% better efficiency and this time around they cannot kill the fairly inefficient high shader clock. There was a lot worth improving in Fermi. There is less in Kepler. Intel made big jumps when Core came around. Then they Nehalem and Sandybridge made big changes to the cache and memory controller. Since then it was only the GPU they really improved. I don't see anything really significant for Maxwell that can again deliver what they hope for.
All I read in that one rumors thread seemed to suggest more of the same on the new process node.

680 vs 580 was on average only 30% faster. Also somewhat more efficient but I guess 50% more performance for a 850M would be realistic. 100% would be very optimistic and triple performance like some hoped is wishful thinking.

 

[Ice Dragon]

Q4 is just the leaked roadmap for Broadwell-K. A lot's up in the air between now and then. If Maxwell seems early and next-gen Broadwell seems far off, what you're suggesting seems possible...although I bet it would make the execs at Apple wince. They seem pretty dead-set on getting rid of dGPUs.

In the MacBook Pro, maybe... though not in the iMac.

It will be interesting to see how much of an improvement the 850M is from the 750M.

 

[Mr. Wonderful]

Hopefully you are right. But changing the scale and making the icons bigger means that reading the first conveys lowest performance of Maxwell is 14 Gflops per watt and the second could be saying lowest performance of Maxwell is 8 Gflops per watt.

There is a good chance that Maxwell won't come with a die-shrink to 20nm. If that doesn't happen, expect the performance per watt improvement to be much smaller until a 20nm refresh.

 

[Dovahkiing]

Which really shows that, yes they are pretty much on a straight line doubling of performance per watt every generation.

This.

Whenever you graph something, you're always trying to make a specific point. The first graph is linear because that puts visual emphasis on how much better Maxwell is. The second graph is log base 2 to demonstrate exponential growth.

 

[whitedragon101]

Yeah, my biggest thought is there is no guarantee whatsoever you'll see Maxwell in a MBP. It's as clear as the day is long that Apple would like to get off the dGPU chain. Despite what Maxwell will offer, it would not surprise me if, given the 40% projected improvement that Broadwell is supposed to bring, we've seen the last generation of discrete graphics.

I am hoping that Apple will keep a dGPU in the high end MBP until the iGPU can match it .

 

[thebrobrah]

broadwell k in 4q

historically, intel releases enthusiast cpus a quarter to two after the mainstreams

production is on track for q1...which means oems get them in q2

and we will see them on market in q3

 

[Vanilla35]

yes