Will there be a new Mac Pro in 2015?

[filmak]

Please no more Fury battles...

We all wish to have the better available GPUs, but please... there is no need for personal attacks and fights.

 

[koyoot]

MVC, I have no words... You take one situation, where there is no timeframe chart, to prove your point, whereas I have brought many more, not only one situation, where people reduced the TDP of the Hawaii GPUs by 50% to retain 88% from links above while consuming much less power. It also staggers me that you can refuse to acknowledge the most obvious proof which is Mac Pro and iMac. They have 250W TDP cards that consume at max 50% of nominal power, while maintaining 85% of nominal performance. And Fury X from Tomshardware review there is maintaining 90% of nominal performance while consuming almost 100W less. Why? Because the average framerate is at 90% of nominal result. That means average core clock is at 950-1000 MHz. You can refuse to see this. That is your problem. And yes, 12 FPS is not playable. But without timeframe scale there is no evidence to properly judge what is or is not playable. We have average FPS from 4K Thief. That is all what we have. Plus dozens of reports from people who actually owned and tested AMD GPUs in downclocking and reducing the power consumption.


Filmak, it was not me who started the personal attacks. All I wanted to is discussing about downclocking potential of AMD GPUs, not only Fury, and the possibilities for next Mac Pro.

And Im sorry, but I tired of this argument...

 

[rdav]

Well, the new SkyLake Desktop chips are finally out (x2) today.
Does that improve the odds of an nMP(7) in the next few months?

"Intel’s Skylake Core i7-6700K ~ Released & Reviewed."
http://arstechnica.co.uk/gadgets/2015/08/intel-skylake-core-i7-6700k-reviewed/

If only.

 

[Bradleyone]

Please no more Fury battles...

I suspect we're already too far down that road to turn back now!

If we're gonna die, we're gonna die historic on the fury road!

 

[filmak]

MVC,.........

Filmak, it was not me who started the personal attacks. All I wanted to is discussing about downclocking potential of AMD GPUs, not only Fury, and the possibilities for next Mac Pro.

🙂Dear Koyoot, I know that you were not the one who started the attacks.

I really appreciate your thoughts and the whole study you 're doing for our benefit, even if we 're all just speculating, it's very informative and usefull.

As you and everybody(?) else I fully enjoy the calm exchange of thoughts and the polite and constructive discussions.

We 're not nuclear scientists, we 're just trying our best to understand and discuss about some options for the nMP v.2

 

[Bradleyone]

Well, the new SkyLake Desktop chips are finally out (x2) today.
Does that improve the odds of an nMP(7) in the next few months?

Only if Apple does a 270 deg turnaround and positions the nMP as a hard core gamer machine...

(Which I suspect some around here would actually welcome!)

These first Skylakes were chips intended largely for gaming PCs, hence the announcement at Gamescom. More will dribble out over the coming months, culminating in workstation chips (Xeons) a long time from now.

 

[filmak]

I suspect we're already too far down that road to turn back now! ...
If we're gonna die, we're gonna die historic on the fury road!

Very nice. 🙂
I hope that we will avoid the coming explosions.

 

[koyoot]

For Skylake I think the most important are server parts - EP. Which are not due till next or even 2017 year. We now have Haswell EP chips. Which don't bring anything special apart from DDR4 to the table. Broadwell - well, there are reports that it could be cancelled by Intel which in all reality would not be that strange. KitGuru reported that if Broadwell is really canceled we would see Skylake-EP chips a bit faster. The thing is: are new CPUs bringing that jump in single core performance? Disputable. Of course there are tasks that IIRC Skylake brings 29% increase in performance.

What I like however about Skylake are 35W quad core desktop parts. Really, really efficient. But its not related to this thread 😉.

 

[throAU]

Well, the new SkyLake Desktop chips are finally out (x2) today.
Does that improve the odds of an nMP(7) in the next few months?

"Intel’s Skylake Core i7-6700K ~ Released & Reviewed."
http://arstechnica.co.uk/gadgets/2015/08/intel-skylake-core-i7-6700k-reviewed/

If only.

No, the Xeons have not been updated and generally lag the consumer grade CPUs by a year or something. Because they have a huge amount more cache and other stuff that is perhaps harder to get quite so good yields on early on. Or Intel just like to milk the old design a bit more.

 

[MacVidCards]

MVC, I have no words... You take one situation, where there is no timeframe chart, to prove your point, whereas I have brought many more, not only one situation, where people reduced the TDP of the Hawaii GPUs by 50% to retain 88% from links above while consuming much less power. It also staggers me that you can refuse to acknowledge the most obvious proof which is Mac Pro and iMac. They have 250W TDP cards that consume at max 50% of nominal power, while maintaining 85% of nominal performance. And Fury X from Tomshardware review there is maintaining 90% of nominal performance while consuming almost 100W less. Why? Because the average framerate is at 90% of nominal result. That means average core clock is at 950-1000 MHz. You can refuse to see this. That is your problem. And yes, 12 FPS is not playable. But without timeframe scale there is no evidence to properly judge what is or is not playable. We have average FPS from 4K Thief. That is all what we have. Plus dozens of reports from people who actually owned and tested AMD GPUs in downclocking and reducing the power consumption.


Filmak, it was not me who started the personal attacks. All I wanted to is discussing about downclocking potential of AMD GPUs, not only Fury, and the possibilities for next Mac Pro.

And Im sorry, but I tired of this argument...

You linked to an article:

http://www.tomshardware.com/reviews/amd-radeon-r9-fury-x-power-pump-efficiency,4215.html

You claimed that it found:

"Did you read, that Tomshardware reduced the TDP by 50% and the power consumption went from 267W to 170W? Did you read that link at all? Did you see that at 170W it maintained 95% of nominal performance?"

But the actual conclusion by the Tom's Hardware folks was this:

"Would it be possible to trade a small and expected performance hit for significantly lower power consumption akin to what we see at Full HD? Unfortunately, no."

You keep waving your arms, jumping up and down, but the very article you linked to came to the conclusion that it didn't work. You refuse to acknowledge something that everyone here can go read for themselves. You want to tell everyone that they can have their cake and eat it too. They can't.

The conclusion of the article you linked DISPROVED your theory. Time to admit it.

Or write the Tom's Hardware folks and point out that they mis-read their own graphs and tests.

I do believe Fiji will be in nMP. I believe it will be fully clipped and down clocked. The result will be exactly what your friends at Tom's found:

"Even though the power consumption decreases from 267W to 170W when the power limit is set to -50 percent, the resulting frame rates just aren’t in the playable range any more."

See, those are the EXACT power numbers you keep quoting. They found that while you could lower the power consumption, you lost significant performance. You keep wanting to claim that the article doesn't say that, IT DOES.

If Fiji is in nMP it won't perform anywhere near the 375 Watt versions shipping now. It will need to use 1/3 of the power and will, as Tom's Hardware shows, offer significantly reduced performance. There is no Free Lunch. If you reduce power and heat, you will lose GPU performance.

Your argument is based on articles that come to the opposite conclusion to the one you are "proving" with them.

 

[netkas]

There is one small chance Fury nano could be good.

if it work at say 800 mhz, it means it needs less voltage to operate.

power consumption is function of Voltage squared.

Default Fury x voltage is 1.2 V.

Just speculating now:

if R9 nano is at 800 mhz and 1 V.

-20% chip power from lower clock, - 30% (1-1/((1.2)^2) ) chip power, together -50% power usage, almost at 180 watts now

Example of voltage vs power difference - https://tpucdn.com/reviews/AMD/R9_Fury_X_Overvoltage/images/power.gif

if chip can work at 800 mhz and -200mV is another question. If it can it would be levels around gtx980/r9_390x.

This month we will find out the truth.


My guess is:

This generation of amd card wasn't planned to operate at 1 ghz, for example initial clocks of 7950 - 800 mhz and it works great there.

To get on par with nvidia amd had to rise clocks and voltage and then gcn cards became very power hungty, for example overclocking beyond 1 ghz makes 290x a very power consuming card.

back to topic.

We might not see new mac pro this year, but only next year, with skylake-e, TB3 and 16-nm amd card. Only this 3 things will make new trashcan look better than old one.

 

[Demigod Mac]

MVC, I have no words... You take one situation, where there is no timeframe chart, to prove your point, whereas I have brought many more, not only one situation, where people reduced the TDP of the Hawaii GPUs by 50% to retain 88% from links above while consuming much less power. It also staggers me that you can refuse to acknowledge the most obvious proof which is Mac Pro and iMac. They have 250W TDP cards that consume at max 50% of nominal power, while maintaining 85% of nominal performance. And Fury X from Tomshardware review there is maintaining 90% of nominal performance while consuming almost 100W less. Why? Because the average framerate is at 90% of nominal result. That means average core clock is at 950-1000 MHz. You can refuse to see this. That is your problem. And yes, 12 FPS is not playable. But without timeframe scale there is no evidence to properly judge what is or is not playable. We have average FPS from 4K Thief. That is all what we have. Plus dozens of reports from people who actually owned and tested AMD GPUs in downclocking and reducing the power consumption.

I took only a cursory glance at Tom's numbers but it looks like the reduced power cards stumble badly when stressed.
Not really a surprising result because most games don't put full stress on the GPU 100% of the time, so they don't need to drink maximum power on average. But the card does a total faceplant whenever it needs that extra power and can't get it, resulting in very choppy framerates during intensive rendering scenes. A very uneven experience most gamers wouldn't find acceptable (that's what they mean by 'unplayable'). I bet if Tom used a 100% stress situation the difference would be even more massive (encoding/rendering) - exactly what the Mac Pro is supposed to be for...

So yes. Deeply compromised performance. Like I said before, you can't beat the laws of physics.

 

[Bradleyone]

We might not see new mac pro this year, but only next year, with skylake-e, TB3 and 16-nm amd card. Only this 3 things will make new trashcan look better than old one.

-E, not -EP?

The Broadwell-E it would replace tops out at 6 or 8 cores. I can't find what Skylake-E is, but if the same wouldn't it seem a definite step back on the max core count? Especially as they could go 18-core today, with a Haswell-EP Xeon E5-2699v3 2.3 18c.

(Low clock speed but turbos up to 3.6 GHz, and there are 18 of the suckers!)

I understand that Thunderbolt 3 could be done with Broadwell-EP and, rumors aside, if that is still due Q1 2016, isn't that a strong possibility? Might be wrong, but the Broadwell roadmap seems to be intact.

As for the video boards, the recently announced AMD Fury/nano tech, though hotly debated, would probably still be a significant step up from the current D300-D700 tech. It might let Apple trumpet "2x increase" or something like that.

Just something to throw the poor trashcan a bone!

 

[MacVidCards]

I took only a cursory glance at Tom's numbers but it looks like the reduced power cards stumble badly when stressed.
Not really a surprising result because most games don't put full stress on the GPU 100% of the time, so they don't need to drink maximum power on average. But the card does a total faceplant whenever it needs that extra power and can't get it, resulting in very choppy framerates during intensive rendering scenes. A very uneven experience most gamers wouldn't find acceptable (that's what they mean by 'unplayable'). I bet if Tom used a 100% stress situation the difference would be even more massive (encoding/rendering) - exactly what the Mac Pro is supposed to be for...

So yes. Deeply compromised performance. Like I said before, you can't beat the laws of physics.

Thank you for words of wisdom.

To Netkas as well.

But I think most people have common sense and hopefully know that there is no free lunch.

AMD isn't running Fury cards at 375 Watts because they want to keep Coal plants burning coal.

Whatever compromises they make for Fiji we will know soon enough. And everyone is able to read that test and form their own conclusions.

My apologies to gentle readers who I may have ruffled. I have felt that the "Pizza & Beer Diet" promo was getting a little thick. I will walk away.

 

[koyoot]

There is one small chance Fury nano could be good.

if it work at say 800 mhz, it means it needs less voltage to operate.

power consumption is function of Voltage squared.

Im Sorry for continuing this but this is last bit from me.

Also keep in mind that if Apple will stay with 125W TDP power envelope for their GPUs, and Full Fiji will be able to be squeezed to it at that level to 850 MHz - that will mean from 250W TDP you will get 14 TFlops of performance on two GPUs.

From my perspective, and Mac Pro perspective, as I said higher in post - best way would be getting the core to 900 MHz, and increasing the bandwidth to 640 GB/s by running the HBM on 625 MHz.

So now, you all agreed with me. All of the data, the Review from TomsHardware, posts from Anandtech people, comments from Internet I provided was to backup my theory of squeezing Full Fiji chip to 125 WITH 900 MHz on core clock.

If It was possible to maintain 1035MHz on core clock at 225W of TDP, if it was possible to maintain 950-1000 MHz of Core clock at 170W, then it was possible to maintain 900 MHz at 125W with downvolting and setting Power Cap. All this argument started because MVC was trying to prove me wrong, without even considering what I meant in my posts, at the beginning. It is from technical point of view quite risky to set that low power envelope for GPU with that core clock. More realistic is 850 MHz, simple doubling of performance from D700.

There is a difference in setting power cap, and undervolting. As we could see in all of this argument, -48 mV on core allows even 80 MHz OC, with 40W reduction in power consumption. What would do reducing -48 mV on core with downclocking it? But that is not what I wanted to discuss. TDP limit allows GPU to not down volt itself. It will simply jump between different settings of core clock it can be 900 it can be 1000 depending on the task. The Voltage there stays the same. And that is exact definition of Fury Nano which is supposed to be full Fiji chip with 175W power limit. Core clock is limited by power draw. The same way as the Fury X from TomsHardware review, Hawaii chips from Anandtech forum and from site article comments, that I can't recall its name right now, but I have provided a link with complete quote.

The other way we can look at this is this. Voltage and core clock defines TDP. We can lock voltage really low, with quite high core clock for that power envelope. What that gives is steady performance without exceeding power limits, just like we see in FirePro DXXX from Mac Pro.

And with this, lets end this ongoing argument.

 

[flat five]

I bet if Tom used a 100% stress situation the difference would be even more massive (encoding/rendering) - exactly what the Mac Pro is supposed to be for...

i don't think that's true.. there are (comparatively) humongous limitations which happen well before a gpu can fire all cores at max speed for rendering.. as well as limitations which arise downstream once the data has been through the gpu.

if we follow the logic that #of processors, speed of processors, wattage, and whatnot--

a d700 would be 50x faster for parallelized tasks than a 12core cpu @ 3GHz.
(2048cores X .85GHz = 1740GHz --vs-- 12core x 3GHz = 36GHz)

but in reality, that speed increase you'll likely experience in a well coded application is 10-20x faster instead of 50x faster.


i think a problem is delivering usable data to a gpu at a rate near what it can devour.. apparently, the amount of video ram is critical in this delivery and we'll possibly see vram sizes leapfrogging typical ram size in the next decade.

another bottleneck is that you can't simply divide any given problem up into enough smaller problems to saturate the gpu resources.. say you are building a house with 1000 blocks.. you can't distribute those 1000 blocks to 5000 workers.. further, just delivering the 1000 individual blocks takes considerable time when compared to delivering one or two (non-parallel) blocks..
and even further-- to build the house, you can't just place all 1000 bricks at once.. the second row can't happen until the first row has been laid.. if the bricks are to be placed 100 rows tall with 10 per layer, the most efficient way to lay them be something like 5 workers (processors) handling 2 bricks per layer.

then even further still... say you solved all of the above and can now process 100,000 requests on 100,000cores in the same amount of time it takes to process one request on one core.

all of that data still needs to be combined into a unified whole once it comes out of the gpu.. and will likely need to be combined A+B = C then C + D = E.. etc instead of (A+B+C+D+E).. combining the data is a string of calculations that have to happen in order instead of one single calculation.. this takes time and in this scenario, the single core on the cpu would still be causing the holdup while the gpu is creating a backlog of data.

anyway, there are definite improvements happening regarding parallelization of large tasks.. new algorithms.. advanced mathematics etc.. next level stuff.. but still, we're not to the point where you can just read a spec and expect it to work X amount faster/slower than that other gpu.. 100MHz faster? cool.. but can the process make use of that?

---
[edit] bad maths above.. i'll sort it a little later..
edit2 - sort of sorted.. meh.

 

[flat five]

And with this, lets end this ongoing argument.

you say this in every one of your posts 😛

 

[rdav]

No, the Xeons have not been updated and generally lag the consumer grade CPUs by a year or something. Because they have a huge amount more cache and other stuff that is perhaps harder to get quite so good yields on early on.

Apple currently supply Refurbished nMP(6) models @15% off list price. That may have been a reasonable deal in mid-2014, but it certainly isn't now. Suspect it would need to be about -33% (or more) to make any sense today.

It's been said so many times before, but they should just go ahead and provide an nMP(7) with the best of what's available now. Sorta like they did with the latest (Haswell) MacBookPro(Retina)15" models.

Two years is already a very loooong time.

 

[SoyCapitanSoyCapitan]

No point talking about retail OEM cards. Apple orders their own spec. If nMP continues to exist for another generation and has an AMD option it will be modded Fiji with slower but more RAM (not HBM) than AMD offers on Fury branded cards.

 

[ManuelGomes]

OK, a tool is out to check and enable locked cores in Fiji, as well as Hawaii and Tonga. Cool!
It checks if there are functional working units and if so, you can enable them at your own risk.
I'm sure it will not work in possible Fiji based cards for nMP, but still for the rest of people it's great.

 

[Macsonic]

back to topic.

We might not see new mac pro this year, but only next year, with skylake-e, TB3 and 16-nm amd card. Only this 3 things will make new trashcan look better than old one.

I think you're right, my guess is the same as yours, we may not see a new Mac Pro this 2015. Though now the discussion has been focused on newer hardware, newer GPUs to be coming out, we may want to also consider how software responds to the new hardware with the coming new Mac Pro. The current new Mac Pro users has been having problems with Adobe Premiere Pro and After Effects. Users are getting horizontal artefact lines while rendering video. And Apple needs to fix this before releasing the next new Mac Pro.

https://forums.adobe.com/thread/1422947?start=640&tstart=0

The Adobe thread is 16 months old and no definite solution for the line artefacts problems so far. Some say it's open CL that's causing the problem or it was a GPU heat issue. Some members suggested to switch to software mode while rendering but makes the process slower. I only skimmed thru the thread so I may have miss other comments. The design of the new Mac Pro itself with the GPUs being proprietary makes it difficult to diagnose the source of the problem. There is no option to easily swap out the GPU to test with other models of GPUs for compatibility as the user's choice is limited to the D300s to D700s GPU for the next 5+ years of usage.

 

[edanuff]

It's been said so many times before, but they should just go ahead and provide an nMP(7) with the best of what's available now. Sorta like they did with the latest (Haswell) MacBookPro(Retina)15" models.

There's a very high likelihood that that's what's going to happen, except it won't likely happen until October or January. Whenever that new model comes out, it's a pretty safe bet that it's going to be last-gen (at least by the standards of people on this forum) on almost every component except for one or two big differentiators, such as perhaps the inclusion of an Alpine Ridge TB3 controller, which Intel has said doesn't require Skylake. Apple tends to be very conservative on their part selection these days, and with all the criticism over recent quality issues, is likely to err on the side of being even more conservative.

 

[Bradleyone]

So... It looks like a Skylake Xeon is coming out sooner rather than later.

The catch is, it's an E3 "laptop" chip, and will probably max out at 4-cores, possibly 6, 8 at a stretch.

It also brings Thunderbolt 3 to the party, making it very desirable to Apple.

Given Apple's propensity to put laptop parts in desktops (has the iMac ever had anything but an "M" GPU?) I could see them using this in a Mac Pro 7,1.

Especially if AMD Fury video boards focused attention away from the CPU core count to the GPU compute potential.

 

[edanuff]

So... It looks like a Skylake Xeon is coming out sooner rather than later.

The catch is, it's an E3 "laptop" chip, and will probably max out at 4-cores, possibly 6, 8 at a stretch.

I saw saw this as well and it has Apple written all over it. Too soon to tell what it means, but it's possible Apple goes in a very different way with this, taking the Pro line even further prosumer.

 

[flat five]

Especially if AMD Fury video boards focused attention away from the CPU core count to the GPU compute potential.

they make a lot of money off of 12cores. (and expected higher counts).
many (most) applicable applications currently need the cpu cores instead of gpu.

not saying apple might not make a move along these lines at some point in the future.. just not right now.

there aren't enough programs which are ready to be used as examples.. lots of theoretical improvements to be made out there but not so many solid examples.

it's too soon for apple to take focus off of 8+ core cpus.

to me, it's highly unlikely apple will put one of these xeons in a mac pro.

---

taking the Pro line even further prosumer.

isn't prosumer describing a person? not a computer.
prosumer being an amateur (for lack of better word) buying equipment suitable for professional use.
(paraphrasing dictionary.app)