As others have pointed out on these forums, it's not the speed of
Nvidia can't support the 5K resolution yet, hence going AMD. Plus they probably gave Apple a good deal.
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.
The Definitive riMac graphics benchmark thread
- Thread starter fruitpunch.ben
- Start date
- Sort by reaction score
As others have pointed out on these forums, it's not the speed of
Nvidia can't support the 5K resolution yet, hence going AMD. Plus they probably gave Apple a good deal.
This is a complicated question. Warning: rough numbers and lots of assumptions start here.
In logic systems, when work is performed heat is created. You can think of TDP (Thermal Design Point) is a budget for how much heat the particular chassis can dissipate. Therefore we can directly place a cap on the amount of computational work which can be performed in a chassis. Both lithography and architecture changes allow this fundamental to be altered. As the machine grows more efficient you can do more work while producing less heat. Thusly the chassis can do more work inside the same TDP.
Looking back at the last GPU shrink (40nm to 28nm) might allow us to infer some figures. Nvidia went from the GTX 580 to the GTX 680 and AMD went from the HD 6970 to the HD 7970. These are the easiest cards to get data on because they are flagship cards of the era.
On average performance increased by 15% and simultaneously power consumption decreased by 15%. You can't attribute everything here to a shrink. AMD released GCN and the 580 was a particularly inefficient card. Mobile GPUs gain more from shrinks than desktop GPUs. In the desktop you can usually dissipate the additional heat you create, it doesn't slow you down. In mobile the additional heat just causes more throttling.
We can make an educated guess about the iMac's TDP by consulting this article. At 240W max TDP the iMac is dancing around its thermal limits under a gaming load. Assuming 50W idle not including CPU or GPU. Add 50W for CPU load (i7, not full load) and 125W for GPU load (M395X, full load). Add another 30W for interconnects, wireless, drive access, memory, audio, and so forth. That's 255W, beyond the TDPmax by a not insignificant amount. The iMac will be throttling at this time to limit the heat output and power consumption of the machine.
What if your GPU was 10% faster and 10% less power consumptive? Your 110W GPU will no longer cause throttling (hopefully) and so therefore will be 10% faster base. Plus add 12% because the old GPU had to run 15W down and do less work because of throttling. Call the new GPU 20% faster all else being equal.
That's my incredibly rough speculation and I'm sticking to it. Next year's die shrink iMac GPU will be at least 15-20% faster in the same enclosure. 50fps today is 60fps tomorrow.
Make note that this illustrates why PC towers are so much faster than iMacs for gaming. My PC tower has a 250W GPU; that is one card with the same TDP as an entire iMac computer.
Do you have a reference for this information? From my understanding Nvidia supports DP1.2 and MST.
The problem is we don't know when the 20nm or smaller nodes will be available in quantity for high-power (not mobile) GPUs, nor what the real-world performance benefit will be. There are ominous fabrication problems, different from past generations:
"Waiting on 20nm graphics cards? Don't Bother": http://techsoda.com/no-20nm-graphics-amd-nvidia/
"450nm Wafers and Death of Moore's Law": http://www.extremetech.com/computin...y-450mm-wafers-halted-and-no-path-beyond-14nm
Re the new iMac throttling "As soon as you begin to put it under load", the initial tests show this doesn't happen as it previously did:
I agree with everything you are saying here about the lithography. Except all the articles you link to are old. TSMC 16nm and Sam/GloFo 14nm did pose fabrication problems and were considerably delayed. They are also not providing a true shrink as stated. Different parts are shrunk by different amounts, some less and some more.
However, several years on, we are now at the point where both of those processes have entered volume production. We know that Nvidia at least has taped out Pascal on 16nm, I would be surprised if AMD hadn't done the same. These chips are coming, and the performance improvements will be non-trivial.
Regarding the video, which I did watch and I must say it was quite well produced. Unfortunately I don't think what it presents makes total sense. Apple is cramming just as much power as before into the same chassis. Something doesn't add up. I need to know what tools he is using to capture the data. Generally speaking you have to perform these tests under Windows because OS X doesn't give you a good enough look at the hardware metrics.
Correct. DP1.2+MST is exactly what Tonga (M295X) had in the 2014 5K model and somehow they made this work. All evidence at this time points to them breaking the eDP standard and overclocking a 1.2 stream as well as adding some more pins. There is absolutely no reason they cannot do this with Nvidia.
M395X is still Tonga, so it doesn't support eDP1.3 as far as we know. This means Apple is once again performing some kind of hack on their AMD GPUs to get the 5K iMac to work. No reason they can't do this with Nvidia if they want to.
So t
You forget that Nvidia already has a 28nm GTX 980M that's about 50% faster than the m395x in gaming and uses about the same 100-125 watt TDP. So by going to 16nm and adding the speed gains on top we stand to get a huge boost in performance of around 60+% which is much better than the crumbs we've been getting for 3 years. Of course if AMD may swing back and make a much faster HBM equipped 16nm "m495x" next year too!
Thanks, I chose the m395x for the 4gb of ram. That's the main reason. I realized the 5k display is using 1gb just for the display. Hopefully the imac will last 5 years or so.
Nvidia card of course can drive 5k display, it works in the Mac Pro. So, I am sure they can work in the iMac if Apple want to do it.
http://barefeats.com/imac5k17.html
Last edited:
Hi Huffy, that is indeed interesting and relevant info, could you paste a screen shot with that info please. Thanks
Also if possible a screen of FCPX or other similar app with a 1080p file on the timeline so that we can check how much vram is being used by the screen + a video editing app
I was basing my comment on this post
https://forums.macrumors.com/threads/m380-m390-m395-m395x-thread.1928278/page-10#post-22141559
I run final cut pro x with a 1080p video playing while another 4k video was running with quicktime at the same time.
I run final cut pro x with a 1080p video playing while another 4k video was running with quicktime at the same time.
![]()
That is great info
. Do you remember how much was used without the 4K video playback. If possible a screen-shot (hope not asking too much) Useful for me and those that are between the choice of the 395 and 395x
ThanksThat is great info
Useful for me and those that are between the choice of the 395 and 395x
yes of course
fcpx with a video 1080p playing
fcpx with a video 1080p at stop
yes of course
fcpx with a video 1080p playing
fcpx with a video 1080p at stop
![]()
Thanks!! hehe!! That is great. I dont use istats but it shows that is using about 1GB of vram with FCPX which is great!!
Nice
I quit fcpx and it doesn't move that muchThanks!! hehe!! That is great. I dont use istats but it shows that is using about 1GB of vram with FCPX which is great!!
Nice
I don't use Final Cut Pro, so please excuse my naiveté, but have you tried running some sort of effect which is known to use opencl, on a 4k stream?
such as, say, gaussian blur, or rain, or underwater
such as, say, gaussian blur, or rain, or underwater
Looks like FCPX doesnt use that much vram, at least on 1080p which is what i use so looks goodI quit fcpx and it doesn't move that much
![]()
Thanks again
It's not merely playing video in FCP X that uses VRAM, it is applying and rendering effects.
A much better GPU test within FCP X is the BruceX benchmark. It is extremely simple to run. You just import the XML, double-click in the event browser to put it on the timeline and hit the share button to export a 5k video file. Share>Master File>Settings
Format: Video & Audio
Codec: ProRes 422
Resolution: 5120x2700
Instructions and test file are here:
http://blog.alex4d.com/2013/10/30/brucex-a-new-fcpx-benchmark/
I did the test 3 times in a row and I got an average of 31 seconds with the ProRes 422. Doesn't look too bad to be honest.
Last edited:
Hi Joema, thanks for the input, im aware of the BruceX benchmark. What i was trying to figure is the actual vram usage on the 5K display with a app like FCPX and Huffy was of a great help.
Although Brucex is a great representation how the GPU will handle high demanding tasks, my concern at this moment is real work flow and brucex does not represent what i do in real life.
Just to give some perceptive i work with a triple display setup 3x 1080p and ill be upgrading to an imac as soon as i have all my questions solved which at this moment is just the GPU
It looks like the 395 is the way to go in terms of price / performance, and the concern is running besides the 5K plus 2x 1080p (I would have no problems in paying 300€ more on the 395X if it had a huge difference from the 395 but so far i haven't seen anything that justifies the price, i may be wrong)After Huffy tests i calculate that 5K with FCPX is using roughly 800mb of vram and plus 2 displays will get up to 1GB of vram so it looks promising which is what im using +- at this moment having Premiere Timeline on Main display, Full 1080p preview on the left monitor and just some random aps on the right side (mail etc)
Now BruceX could be interesting if someone with the 395 and the 395X could run and compare the results (which probably someone already did
and i didnt catch it )My 395x is coming maybe next friday or early the other week. I'll definitely do the same rendering again