Are GPUs ending up like CPUs?

[inachu]

within 5 years all video card norms will all be external hardware devices to keep the pc cool.

 

[jpiniero]

You have to remember that Big Maxwell will probably have 1/2 DP, so whatever it costs in terms of adding more transistors to get there is going to eat into the budget. Note that GK110 is a 90% bigger die than GK104 yet only had 50% more cores, and Titan was only 1/3 DP. So a theoretical Big Maxwell that has a 50% bigger die than GM204 (~600 mm) might only have 20% more cores.

 

[CrazyElf]

You have to remember that Big Maxwell will probably have 1/2 DP, so whatever it costs in terms of adding more transistors to get there is going to eat into the budget. Note that GK110 is a 90% bigger die than GK104 yet only had 50% more cores, and Titan was only 1/3 DP. So a theoretical Big Maxwell that has a 50% bigger die than GM204 (~600 mm) might only have 20% more cores.

Not sure about that one. Let's do an analysis.

Hmm ... 680
8 SMX at about ~294mm^2 = 1 SMX per 36.75mm^2

While Titan Black
15 SMX at about ~561mm^2 = 1 SMX per 37.4 mm^2

I don't think that the double precision took that much performance out of the budget.

Each 680 had SMX had 192 single-precision (SP) Cores, and 8 double-precision cores. Against this, the Titan Black, there were 64 DP Cores.

Of course, in this analysis, we're not factoring the memory controller (384 bit on Titan vs 256 bit bus on 680, which also eats up die space), command processor, and setup pipelines.

All in all, I'm not too sure that the compute functions took that much away from the gaming capabilities. Granted in terms of gaming performance per mm^2, the 680 does outperform the Titan Black, but it's not by a huge amount.

within 5 years all video card norms will all be external hardware devices to keep the pc cool.

Fermi is probably as hot as it will realistically get.

 

[Enigmoid]

You have to remember that Big Maxwell will probably have 1/2 DP, so whatever it costs in terms of adding more transistors to get there is going to eat into the budget. Note that GK110 is a 90% bigger die than GK104 yet only had 50% more cores, and Titan was only 1/3 DP. So a theoretical Big Maxwell that has a 50% bigger die than GM204 (~600 mm) might only have 20% more cores.

Someone correct me if I'm wrong but Kepler used separate SP and DP shaders while Maxwell will reuse DP shaders in SP mode (why it went from 1/24 to 1/32 DP rate).

 

[bunnyfubbles]

GM204 is 5.2B transistors on 28nm and is performing similarly if not superior to the 7.1B transistor GK110 on 28nm, all with power consumption performance similar if not superior to the 3.5B transistor GK104 on 28nm...

whereas the GK104 had a new 28nm advantage over 40nm GF110, and thus GK104 actually had more transistors vs. GF110, (3.5B vs. 3B) as well as much higher clock rate (and also much faster GDDR5 to nullify bandwidth advantage, 256bit vs. 384bit but both @ 192GB/sec)

any shortcomings by GM204 is pretty obviously due to being stuck on 28nm

 

[BrightCandle]

CPUs and GPUs both became thermally limited a few versions ago and power saving become a big issue for them to being able to fully utilise the increased transistor density. They have at least managed to get this sort of under control to be able to get some benefits from improved process and design but its not really been the same impressive rate of progress compared to what we saw before.

To some extent GPUs have a different issue to CPUs today however. A CPU really can't accelerate a single thread a great deal more without more clock speed. Sure you can do some improvements to architecture that improves instruction performance but fundamentally the performance comes from clock speed, and they can't increase that because of the thermal issues with doing so. They could take the extra transistors and put them into more cores, but because software isn't written to use it (and its hard to write software to do so) in a lot of cases its wasted transistors in a lot of programs. So more transistors aren't really bringing much in the way of actual performance improvements because we can't extend the pipelines any further, we can't improve the instructions easily and adding more cores just doesn't bring more performance without big software changes. Making progress here is just hard.

GPUs on the other hand can scale to a lot of cores. At the top end of scaling you could imagine having a GPU with 1million or more processing cores in it and every shader program being run across all the pixels on the screen at the same time. Scaling past that point becomes tricky but we have 3 orders of magnitude of growth before we get there (we are around 2000 cores today) that should bring enormous gains to performance. We potentially have memory bandwidth issues long before that point and a whole host of other scaling issues to solve but the problem of rendering is a hugely parallel activity and we know it scales very very well per pixel. More transistors = more cores = more performance and we don't need clock speed or other architecture improvements other than to support the computation performance growth that comes with more cores and having to feed them all with data.

So whereas CPUs don't get much benefit from the increase in transistor budgets due to the types of problems they solve, GPUs can and do get a lot of benefits. But then along comes the problems with 20/16nm from TSMC and global foundries etc and the next generation of cards are late and we are calling doom and gloom. I think the GPU will continue to grow at a pretty decent rate of performance once those new processes come out, its not the same problem as we see on CPUs they scale just fine with more transistors. Physics might be kicking TSMC down a bit right now but it'll get fixed and we'll see genuine improvements in cards, we are not going to be stuck at 10% gains like with CPUs until they genuinely call dead on the silicon process improvement game.

 

[3DVagabond]

You have to remember that Big Maxwell will probably have 1/2 DP, so whatever it costs in terms of adding more transistors to get there is going to eat into the budget. Note that GK110 is a 90% bigger die than GK104 yet only had 50% more cores, and Titan was only 1/3 DP. So a theoretical Big Maxwell that has a 50% bigger die than GM204 (~600 mm) might only have 20% more cores.

Hawaii is 1/2 DP and isn't larger because of it.

 

[UaVaj]

there are folks who will spend money at the next tech upgrade. rather that upgrade be 1) performance gain or 2) wattage reduce or 3) just bell and whistles or 4) just new label (a rename of the same gpu). this business as usual. all for profit.

gpu performance is clearly near a wall. little room left to go. so why not give these consumer hungry folks something else. for now - how about some wattage drop.

once the wattage cow get milked dry. the next profit cow might be. all gpu are dual gpu card. then us consumers can run octa cfx/sli to match our octa cpu.

 

[jpiniero]

Hawaii is 1/2 DP and isn't larger because of it.

Is it? It's not as big as Big Kepler, but it's still rather big. Hell, look at Tonga. It's not that much smaller than GM206 yet is much slower. Even fully enabled and clocks closer to what Maxwell is doing it would still be quite a bit slower. AMD for some reason hasn't disclosed what the DP rate is on the FirePro version though.

 

[Techhog]

You're comparing mid-range Maxwell to high-end Kepler, both on 28nm. That's why there isn't a huge jump. If this were 20nm or GM200, there would have been a bigger jump.

 

[Silverforce11]

Maxwell is a massive jump, you simply don't understand it properly to compare it to the stagnation on the CPU side.

For the past few generations, Intel release a new CPU with ~5% gains and calls it a day.

 

[f1sherman]

Maxwell is a massive jump, you simply don't understand it properly to compare it to the stagnation on the CPU side.

For the past few generations, Intel release a new CPU with ~5% gains and calls it a day.

If someone switched your latest Haswell with Intel's 2,3 years old Ivy/Sandy, you'd hard pressed to notice the difference under any workload.
Most likely you'd need benchmarks to see the difference.

http://www.anandtech.com/bench/product/287?vs=836

That's how small are CPU gains are.

Try saying with the straight face that 60% of gaming perf lead that GTX 980 has over 2yr old 680
(and which is 99% guaranteed to increase over time) is similarly hard to notice.

 

[Pariah]

If someone switched your latest Haswell with Intel's 2,3 years old Ivy/Sandy, you'd hard pressed to notice the difference under any workload.
Most likely you'd need benchmarks to see the difference.

http://www.anandtech.com/bench/product/287?vs=836

That's how small are CPU gains are.

Try saying with the straight face that 60% of gaming perf lead that GTX 980 has over 2yr old 680
(and which is 99% guaranteed to increase over time) is similarly hard to notice.

Not sure why you chose the 4770k. This would be the actual latest Haswell (4790k) compared to SB:

http://www.anandtech.com/bench/product/287?vs=1260

Gains are about 30-50%. In actual CPU heavy apps, you would definitely notice the difference.

Granted, once you OC both, the gap will be much smaller. Which is why I am still content with my SB 2600k purchased almost 3 years ago.

 

[f1sherman]

2600K/4770K being the most common mid-high end CPU and first in pipeline.

2600k to 4790K is 3.5 years.

That's almost(4 months short) as GTX 580 -> GTX 980

Intel's 30-50% vs Nvidia's 300%-400% is not even funny

 

[Silverforce11]

2600K/4770K being the most common mid-high end CPU and first in pipeline.

2600k to 4790K is 3.5 years.

That's almost(4 months short) as GTX 580 -> GTX 980

Intel's 30-50% vs Nvidia's 300%-400% is not even funny

Intels gains for the purpose of us gamers, are close to irrelevant, since in games, you would be extremely sensitive to notice between 2600K vs 4790K. Even in apps, the gains over multiple generations are pathetic.

I have a spare rig with a 2500K in it, I know full well how much of a NON upgrade a 3570K or 4670K are.

This is stagnation.

This has not occurred for GPUs and unlikely to occur in the short to medium term.

 

[Pariah]

2600K/4770K being the most common mid-high end CPU and first in pipeline.

2600k to 4790K is 3.5 years.

That's almost(4 months short) as GTX 580 -> GTX 980

Intel's 30-50% vs Nvidia's 300%-400% is not even funny

Ok. Well the GTX 780 is the most common NVidia enthusiast card and first in the pipeline when compared to 980. Obviously older products are going to be more common than recent releases. The price difference between the 4770k and 4790k is $10. The only reason to choose the 4770k is to fudge the numbers in your favor. The 2600k was the fastest consumer CPU available when Intel launched. The fastest CPU available today is either the 4790k or the 5960x depending on workload. With highly parallel work loads, a 5960x will pummel a 2600k. The 980x being only a week old is skewing the results for video cards.

A 4790k is quite a bit faster than a 2600k. The fact it is about 0% faster in games is not the fault of Intel, they don't decide where to program in the bottleneck for games.

 

[f1sherman]

Ok. Well the GTX 780 is the most common NVidia enthusiast card and first in the pipeline when compared to 980. Obviously older products are going to be more common than recent releases. The price difference between the 4770k and 4790k is $10. The only reason to choose the 4770k is to fudge the numbers in your favor.

LMAO

I already chose your 4790K, look:

2600k to 4790K is 3.5 years.

That's almost(4 months short) as GTX 580 -> GTX 980

Intel's 30-50% vs Nvidia's 300%-400% is not even funny

Now add 10%(4months/3.5years) to Intel's 30-50%(haha 50% like hell it is, but nvm lets trust your numbers), and again compare that to Nvidia's 300-400% performance gains, look:

Intel 40-60% vs Nvidia 300-400%

 

[CrazyElf]

GPUs on the other hand can scale to a lot of cores. At the top end of scaling you could imagine having a GPU with 1million or more processing cores in it and every shader program being run across all the pixels on the screen at the same time. Scaling past that point becomes tricky but we have 3 orders of magnitude of growth before we get there (we are around 2000 cores today) that should bring enormous gains to performance. We potentially have memory bandwidth issues long before that point and a whole host of other scaling issues to solve but the problem of rendering is a hugely parallel activity and we know it scales very very well per pixel. More transistors = more cores = more performance and we don't need clock speed or other architecture improvements other than to support the computation performance growth that comes with more cores and having to feed them all with data.<br />

<br />
<br />

It's true that GPUs are more parallel.

There's one problem with your argument though. This whole thing is dependent on transistors of the next process being cheaper in terms of price per transistor than the one before it. 28nm is the cheapest process.

After that, the cost per transistor increases. I don't think you understand my question. The question is, can GPU performance continue to scale in a world where the price per transistor increases with each smaller node?

 

[Pariah]

I'm not arguing that GPU's have advanced more than CPU's. The parallel nature of their workload makes it much easier to increase performance than for a general purpose CPU. Just pointing out you don't need to fudge your comparisons to try and make them look better. Though again, the comparison still is not apples to apples, as the 2600k was the fastest consumer available CPU at launch. Faster than the previous generation $1000 6 core Xtreme processor on average. The 580 GTX by comparison was not the fastest consumer level video card at launch.

 

[f1sherman]

Among all this fudging... I lost you.

Anyway 580 -> 980 is more like 200-250%, not 300-400% like I previously mentioned.

 

[tidy]

.

 

[easp]

Are you focusing on the right things? Smaller dies generally make for cheaper chips. The new generations are also more power efficient.

GPUs are generally power and cost constrained, so improvements in both those metrics are, fundamentally, the only progress that really matters. Cheaper, lower power means you can buy and run more GPU.

 

[toyota]

580 to 980 is only about a 115-125% improvement since its just a little over twice as fast overaal.

 

[njdevilsfan87]

Nvidia has the compute market. Intel has seen that a battle lost with just a single CPU, hence Xeon Phi.

But they definitely are becoming like Intel CPUs in regards to release cycle. Mainstream parts first followed by the big chips later on.