Got a tip for us?
Let us know
Become a MacRumors Supporter for $50/year with no ads, ability to filter front page stories, and private forums.
AMD Confirms RX 480 At $199 USD, Other APU & Polaris Announcements
- Thread starter lowendlinux
- Start date
- Sort by reaction score
I think you're the only person who seems upset about this perf per clock metric. Using the same process (28nm) NVIDIA improved their efficiency with Maxwell, delivering a massive performance increase over Kepler. Yes, the clocks ended up being higher as well, but most of the improvements came from designing their GPU cores to be more efficient. Pascal continues that trend, but NVIDIA gets a larger increase in clock speed from the new process as well. So:
- Raw performance is improving with each generation.
- Perf/watt is improving with each generation.
- Large clock speed increase from new process.
What are you complaining about exactly? Shouldn't you be asking why AMD wasn't able to get a similar increase in their clock speeds when they moved from 28nm to 14nm as well?
Also in not every GPU rendering benchmark GTX 1080 is slower than previous gen.
https://www.computerbase.de/2016-06...5/#abschnitt_gpucomputing_und_videowiedergabe
I am not seeing any cases here where the GTX 1080 loses to any other Nvidia GPU.
CPUs are very different from GPUs. Generally, CPUs focus on single threaded tasks while GPUs are massively parallel. Both AMD and Nvidia increased their clock speeds relative to the previous generation while also improving efficiency (especially Nvidia). Given todays market for mobile products I doubt neither Nvidia or AMD have lost sight of the fact that their GPUs need to scale down to small power envelopes.
If you took a GTX 1070 and clocked it at 1Ghz it's no more powerful than a two year old budget card that had a similar clock speeds and similar or better power envelope.
The current Pascals are basically equivalent to budget cards that Nvidia were able to give high clock speeds because of the 16nm process.
I have no shame admitting that as a 1070 owner. I just have to deal with it. Nvidia's "progress" should be credited to the fabrication plants more than to themselves.
Thats simply not true. The GTX 1070 and 1080 are the most efficient consumer GPUs that exist. Even if you scaled down the clocks to make it say a 100 W card it would beat all previous 100 W cards like the gtx 960.
It's 100% true.
A 1070 running at 1Ghz would not be significantly better than a 770 or 960 running at the same clock speed. Dozens of benchmarks in reviews prove this simply by looking at the performance scaling.
The only time when this does not hold true is when more VRAM is needed.
You are simply wrong. GTX 1070 clocked at 1 GB would be between GTX 970 and GTX 980.
It would also provide 3.8 TFLOPs of compute power.
I think that GPU computing has less of an issue with pipeline bubbles compared to general purpose CPUs.
Thats silly. The GTX 1070 has 1920 compute units whereas the GTX 960 has 1024. Basically the GTX 1070 would have to be half as fast per clock cycle for that to be true. Its clearly not.
Lets look at some numbers for the game witcher 3 at 1080p:
GTX 1080: Freq: 1733 Mhz, FPS: 100.3, FPS/Mhz=0.058
GTX 1070: Freq: 1683 Mhz, FPS: 82.1, FPS/Mhz=0.049
GTX 960: Freq: 1228 Mhz, FPS: 32.9, FPS/Mhz=0.027
GTX 770: Freq: 1130 Mhz, FPS: 33.2, FPS/Mhz=0.029
Clearly the GTX 1080/1070 is roughly twice as fast in this somewhat arbitrary metric. This stems from the fact that it has roughly double the number of compute units as these low end video cards. I agree that pascal is slower in IPC, but it doesn't matter very much for GPUs.
Forget about this theoretical TFLOPS calculations. All the real world tests are in the reviews and done by myself right here where I sit.
We should be judging architecture improvements by seeing how much better performance increases each year per clock.
Back when AMD was a big threat to Intel the latter company conducted a megahertz war by ramping up clock speeds with the Pentium IV. It had terrible efficiency per clock.
After it was clear that AMD was beaten, then Intel went back to the drawing board and decided to focus on not only per clock efficiency but also power efficiency.
So clock speeds have barely climbed in the last 5-6 years, yet per clock each new generation is far more efficient. Today's 3.5Ghz processor is 80% more powerful than a similar clocked processor from 5 years ago.
This is what Nvidia should be focused on. Just ramping up clock speeds without showing much per clock improvements in the last 3 years isn't impressive. It's great that the new cards can perform so well but if you can SLI or crossfire a couple of budget cards from two years ago and get the same performance then I'm not so sure about value for money.
[doublepost=1468873274][/doublepost]
Your link doesn't say the clock speeds the Pascals are tested at and the typical boost clock numbers given in the official specs are lower than actual boost clocks.
My 1070 reference card is about 1800mhz under gameplay. You should know that by now, especially since I showed 200 MHz overclock just two days ago when running 3dmark at 2ghz.
And then you have nudged him 960 and 770 boost clocks upwards. The actual clocks are.
960 : 1178
770 : 1085
And unlike the Pascal, these two older cards dont surpass their official boost clocks because of thermal throttling and stability issues.
Of course I exaggerate a little because the 1070 is on average more like 120% better than a 960. So at the same clock speed the 1070 would be a little better, but certainly nothing to write home about.
I'd just like to be able to afford to buy an RX480 (equivalent to $330 for 4 GB version) or 1080 (equivalent to $1000 for 8 GB version (http://kakaku.com/pc/videocard/itemlist.aspx?pdf_Spec103=420)) here in Tokyo without requiring a second mortgage and that would ACTUALLY WORK under macOS in a cMP. At the moment, nearly all this thread—in a Mac forum remember—is pure speculation based on Windows. What we need is cards for Macs or at least drivers and efi roms.
You've probably got a long wait ahead of you.
AMD doesn't provide OS X drivers at all, so we're at the mercy of whenever Apple will include them, probably based on whenever they start adding newer AMD chips in their hardware. Nvidia provides OS X drivers, but if we take Maxwell as an example, it's going to be about 6 months.
I don't think down lock the 1070 is a correct way to compare it's performance to the previous generation card.
All of them a GPU, but with different architecture, and different optimum configuration. You can drive an aircraft (jet) on the ground with just 20km/h and say that's a very poor machine, even worse than a Toyota. Very noise and high fuel burn. Because that's a transportation tool design to run at nearly speed of sound.
And more importantly, the Toyoya cannot be speed up to speed of sound with linear fuel consumption consumption / performance etc.
All of them a GPU, but with different architecture, and different optimum configuration. You can drive an aircraft (jet) on the ground with just 20km/h and say that's a very poor machine, even worse than a Toyota. Very noise and high fuel burn. Because that's a transportation tool design to run at nearly speed of sound.
And more importantly, the Toyoya cannot be speed up to speed of sound with linear fuel consumption consumption / performance etc.
What are you getting at...? So the new 10x0 cards by nvidia are not only faster per clock but more importantly by absolute terms they are way more powerful than anything out there.
FPS/clock is useless metric anyway. It is doesn't matter. I think what you should be thinking about is FPS/W and based on what I've read so far it seems that the GTX 1060 is actually going to be better at that metric than the RX480. So AMD is once again losing in pretty much every measurable metric. Which of course, from a customer point of view is a real shame because over time NVIDIA will also start showing slow performance increases like Intel with no real competition.
I do think about power consumption too, everything I said I mention that.
But it doesn't look good when a Nano at 1Ghz can beat a 1080 at 2Ghz in video and CG based OpenCL scenarios. They are in the same power envelope range.
Saying 'use CUDA' is bad. This forum hates proprietary corporate standards because it results in monopolies, less user choice, reduced performance, lockdowns on user options, and eventually the corporations want your user data so you can be tied to the CUDA powered car/house/robot, etc
Last edited:
You do realise you are on an Apple forum right?
Regarding OpenCL:
"Throughout our entire suite of tests, the GeForce GTX 1080 outpaced all of the other GPUs we tested, in every application / game and at every resolution, with the sole exception being the OpenCL-based benchmark, LuxMark."
Read more at http://hothardware.com/reviews/nvidia-geforce-gtx-1080-pascal-gpu-review?page=9#45rLSzGeMQkCMCoT.99
So you find the absolute worst possible scenario for 1080 and draw your assumptions based on that?
You do realise you are on an Apple forum right?
Regarding OpenCL:
"Throughout our entire suite of tests, the GeForce GTX 1080 outpaced all of the other GPUs we tested, in every application / game and at every resolution, with the sole exception being the OpenCL-based benchmark, LuxMark."
Read more at http://hothardware.com/reviews/nvidia-geforce-gtx-1080-pascal-gpu-review?page=9#45rLSzGeMQkCMCoT.99
So you find the absolute worst possible scenario for 1080 and draw your assumptions based on that?
That is a VERY important scenario that has been brought up by many people on this Apple forum. Windows gaming is secondary or tertiary importance here.
Because you are not active in this discussion or can't read my caveats and nuances you take a rigid contradictory position against me.
I said Pascal for gaming muddafuggin fantastic. For OpenCL workflows not good. Megahertz wars not good. Performance per clock and per watt are both important.
That makes the Nano currently best all kind performer per clock, especially in OpenCL.
Last edited:
Nano is good, but it's still 28nm and puts out a lot of heat. A 14nm Nano with HBM 2, now that would be very interesting.
And again, arguing about whatever is good is a little immaterial if we can't even use the Nano, or anything recent by AMD for that matter. This may change by the time Sierra comes out, but for now, the best we have is Maxwell.
And again, arguing about whatever is good is a little immaterial if we can't even use the Nano, or anything recent by AMD for that matter. This may change by the time Sierra comes out, but for now, the best we have is Maxwell.
What SoyCaptain is saying is that performance per clock on Pascal is smaller than is on Maxwell. Compare for example similar core count GTX 980 Ti to GTX 1080. 2816 vs 2560 CC's. IF you would declock GTX 1080 to similar performance level as GTX 980 Ti it would be slower than Maxwell part.
.
If that last bit is true it might be true that Vega 10 part might have 90 GFLOPs/watt. But that also remains to be confirmed.
P.S. You say that Nano puts a lot of heat, and then say that best of what we have is Maxwell..
My response is: nope.
There is strong word that Vega 10 will be offered in variants: cut down, full die, and... Nano
.If that last bit is true it might be true that Vega 10 part might have 90 GFLOPs/watt. But that also remains to be confirmed.
P.S. You say that Nano puts a lot of heat, and then say that best of what we have is Maxwell.
. My response is: nope
.
Last edited:
This whole thread is about RX480, a card that can't be used in OSX properly at this moment.
This whole thread is about RX480, a card that can't be used in OSX properly at this moment.
And there are some pretty out there comparisons in the last 4 or 5 pages to boot
I know its not exactly related, however...
It appears that on Nvidia Kepler GPUs in Apple computers(iMac 27') for example there is gimped performance with Nvidia web drivers compared to Apple own drivers. It happens in application built with Metal, specifically - with WoW: Legion.
http://us.battle.net/wow/en/forum/topic/18300616773?page=9#172
http://mangohat.net/metal/benchmark.pdf
Can anyone test this with AMD GPUs, how does Metal affect performance?
It appears that on Nvidia Kepler GPUs in Apple computers(iMac 27') for example there is gimped performance with Nvidia web drivers compared to Apple own drivers. It happens in application built with Metal, specifically - with WoW: Legion.
http://us.battle.net/wow/en/forum/topic/18300616773?page=9#172
http://mangohat.net/metal/benchmark.pdf
Can anyone test this with AMD GPUs, how does Metal affect performance?
The GTX 1060 reviews are out and it looks to consistently beat the RX 480 in gaming while using significantly less power. In compute performance its a little more even where they trade blows depending on the test. Until supply evens out with demand it will be sold at a premium compared to the RX 480 ($300) but once AIB start selling at $250 it will be tough to pass up compared to the RX 480.
I gotta say, it will be a shame if Apple doesn't incorporate some of these Nvidia chips into their lineup. Nvidia has a line of efficient GPUs across the performance spectrum while AMD only has the midrange covered with a chip that is only as efficient as the previous generation. If Apple decides to wait on Vega then we may not see an updated mac pro until next year.
I gotta say, it will be a shame if Apple doesn't incorporate some of these Nvidia chips into their lineup. Nvidia has a line of efficient GPUs across the performance spectrum while AMD only has the midrange covered with a chip that is only as efficient as the previous generation. If Apple decides to wait on Vega then we may not see an updated mac pro until next year.