2016 nMP

[AidenShaw]

AMD stands Polaris Twice on Perf/Watt the preceding gpu generation, ....

Of course, that could be interpreted as "one half the speed, one quarter the power" or "one quarter the speed, one eighth the power". Not too many would like those equations.

Some would like "same speed, half the power", but many more would like "twice the speed, same power". (Especially on a system like the MP6,1 with a seriously constrained power budget.)

"Perf/watt" is pure marketing jargon without a reference to absolute perf. It's important, but only in relation to the actual performance.

 

[Mago]

Of course, that could be interpreted as "one half the speed, one quarter the power" or "one quarter the speed, one eighth the power". Not too many would like those equations.

Some would like "same speed, half the power", but many more would like "twice the speed, same power". (Especially on a system like the MP6,1 with a seriously constrained power budget.)

"Perf/watt" is pure marketing jargon without a reference to absolute perf. It's important, but only in relation to the actual performance.

Do you usually makes clever minded contributions,this its the first one that lacks total sense.

They say (AMD) twice flops/watt that's simple no turn around, pick a reference GPU from AMD and do the math on the expected tdp (50-100-150-175) .

I'm speculating on fp32:fp64 ratio, with wistful thinning but it could be even worse than 1:4 could be 1:32 but still a quantum leap for the Mac Pro.

Of course the most important thing here are all those figures are based on AMD marketing material sadly not the most reliable source for anything, since goebels (propaganda as weapon).

 

[Zarniwoop]

The double perf/watt comes most likely on special circumstances only. It could be only for Polaris 11, that it can achieve with 25W chip same results that former 50W. Similarly like Intel has only made great advantage in low power parts, but desktop parts have gained less.

We don't know where is Polaris's sweet spot, but it could be that Polaris 10 can't do 2:1 ration. One obstacle is going to be memory bandwidth, which doesn't grow at all before GDDR5/x and HBM2.

 

[Roykor]

77 pages and rising. Still no single bit of information from Apple if there will be an update at the nMP. But we still love Apple, right? No-matter-what!

 

[Mago]

The double perf/watt comes most likely on special circumstances only. It could be only for Polaris 11, that it can achieve with 25W chip same results that former 50W. Similarly like Intel has only made great advantage in low power parts, but desktop parts have gained less.

We don't know where is Polaris's sweet spot, but it could be that Polaris 10 can't do 2:1 ration. One obstacle is going to be memory bandwidth, which doesn't grow at all before GDDR5/x and HBM2.

Polaris isn't aimed at servers or compute, the best fp32:64 on consumer gpu from AMD was 3:1 highly unlikely to happen, 4:1 would give Polaris a solid value among professionals but isn't something that cares to gamers, their latest Fiji x2 aimed at gamers/VR shares a dismal 32:1 with infamous products as nVidia Quadro K6000 .

Polaris would give us a glimpse on what to expect from Vega which its aimed at compute and should match Pascal GP100 or at least try to.

 

[AidenShaw]

Polaris isn't aimed at servers or compute, the best fp32:64 on consumer gpu from AMD was 3:1 highly unlikely to happen, 4:1 would give Polaris a solid value among professionals but isn't something that cares to gamers, their latest Fiji x2 aimed at gamers/VR shares a dismal 32:1 with infamous products as nVidia Quadro K6000 .

Polaris would give us a glimpse on what to expect from Vega which its aimed at compute and should match Pascal GP100 or at least try to.

Quadro K6000 is 3:1 FP32:FP64.

http://www.anandtech.com/show/7166/nvidia-announces-quadro-k6000
[doublepost=1461589061][/doublepost]

Do you usually makes clever minded contributions,this its the first one that lacks total sense.

They say (AMD) twice flops/watt that's simple no turn around, pick a reference GPU from AMD and do the math on the expected tdp (50-100-150-175) .

By making assumptions about the wattage, you are bringing in absolute performance.

Which was my point...

 

[t0mat0]

The double perf/watt comes most likely on special circumstances only.

28nm all the way down to 14nm - I'd imagine this explains a majority of power savings. They'd be messing something up badly if the performance per watt wasn't 1.5-2x that at 28nm.

 

[koyoot]

GP100 compared to GM200 has only 35% better efficiency per watt. 26 GLOPs/Watt for GM200 in Titan X, and 35.3 GLOPs/Watt for GP100. So everything here depends how AMD will tackle it.

It can be the way: same performance, much lower power consumption, or lower performance, lower power consumption.

If we are looking at P10 and it will be 100W GPU with 2560 GCN cores, and 1150 MHz, that will bring 5.9 TFLOPs of compute power. That will be roughly 2 times the power per watt, compared to similar GPU from last gen.

 

[Mago]

Quadro K6000 is 3:1 FP32:FP64....

Typo not assumption, I meant Quadro M6000 (m/k where too close in my kbd, no way for autocorrect to save my ass and you know "The night is dark and full of terrors" ..)

 

[Mago]

This post is to introduce a bit the audience on micro electronics:

When you read "that product is made on 28nm process and that other on 14nm process, so the later either its faster and uses less power..."

A 28 nm process means the smallest feature an ASIC (Application Specific Integrated Circuit) have measures 28nm such feature or node either can be a logic gate a nand whatever the foundry defined as its smallest node.

A 14nm feature implies the same asic uses 1/4 the silicone, given power dissipation on circuits its in proportion to the current and voltage switching on an specific area, the same feature on 14nm only allows 1/4 the wattage either working on less current (likely) and ir less voltage (preferably but not always possible).

That means that integrators (as nVidia or AMD) have available either quad the number of logic gates using the same area or use 1/4 the area and save 75% power, or trick in the middle reduce power by 50% and try duplicate performance (not really since a smaller feature also implies use less voltage and indeed less switching speed).

So assume AMD wants an power efficient Fury X, they commission the same gates array on 14nm (ultra scale electronics usually aren't defined on a stencil as circuits boards but using a file describing the logic gates organization, while some times a specific feature maybe commissioned with an special stencil as sensors ), given switching from 28 to 14nm isn't something as easy some features use to be reviewed to correct issues with signalling, heat capacitance etc, so finally this 14nm version of fury x should have an slower clock but will draw 1/4 the power this is good for many purposes but don't sell as well as performance does, so you decide to use more logic gates (or faster ones which are bigger) to scale up processing then using more silicone and power, at the final you have something as Polaris or Pascal delivering twice performance at the time using less power and having better yield since a smaller asic implies less possibility to discard an defective ic.

Depend on how you play with those variables you'll be successful on your new product.

So having a 1/4 the feature size you can improve both performance and efficiency.

Note feature size use to be the longitude but maybe the surface area,so this case switching 28nm to 14nm only allow half power savings and duplicate the gates.

 

[Ph.D.]

A few corrections:

The number in a process node, e.g. "28 nm" or "14 nm" means the minimum length of a transistor's gate (only!) of a single transistor, not the size of a full transistor (gate plus source and drain) or that of a full logic function such as a NAND gate, which traditionally (in CMOS) takes 4 transistors plus some wiring, etc. (though these usually scale down too).

It's "silicon" not "silicone." That's an important distinction when dating as well as when making chips.

Power scaling depends very much on how the scaling (i.e., shrinking from 28 to 14nm) is done: Is everything scaled equally, including the gate oxide and voltage? (That's rare these days.) Is just the gate length scaled? There are many strategies. Naively, power consumption would go down by the square of the scaling factor, meaning that the same design, scaled by a factor of two, would use 1/4 as much power. The chip would also be 4 times smaller. The problem is that scaling is rarely full these days - it's a compromise in many ways - and the quoted power savings in this case is of order 2.5x for the "same" chip, just shrunk.

But, having just scaled things down, people want a bigger, faster chip anyway, and so the power goes right back up. Then, there's an reasonably-arbitrary choice to be made between power and performance. So it's basically true that in 14nm you could have roughly the "same" performance at 2.5x less power, or 2.5x the performance at the same power, or anything in between.

This post is to introduce a bit the audience on micro electronics:

When you read "that product is made on 28nm process and that other on 14nm process, so the later either its faster and uses less power..."

A 28 nm process means the smallest feature an ASIC (Application Specific Integrated Circuit) have measures 28nm such feature or node either can be a logic gate a nand whatever the foundry defined as its smallest node.

A 14nm feature implies the same asic uses 1/4 the silicone, given power dissipation on circuits its in proportion to the current and voltage switching on an specific area, the same feature on 14nm only allows 1/4 the wattage either working on less current (likely) and ir less voltage (preferably but not always possible).
...

 

[Zarniwoop]

77 pages and rising. Still no single bit of information from Apple if there will be an update at the nMP. But we still love Apple, right? No-matter-what!

Don't forget these..

Will there be a new Mac Pro in 2016?
82 pages

Will there be a new Mac Pro in 2015?
122 pages

nMP 2015: where is it?
8 pages

 

[Ph.D.]

I'm surprised we don't have a several-thousand post thread titled "The new Mac Pro is almost certainly coming!"

Don't forget these..

Will there be a new Mac Pro in 2016?
82 pages

Will there be a new Mac Pro in 2015?
122 pages

nMP 2015: where is it?
8 pages

 

[Stacc]

I'm surprised we don't have a several-thousand post thread titled "The new Mac Pro is almost certainly coming!"

You must be new around here. All we talk about is how "APPLE DOESN'T CARE ABOUT PROFESSIONALS!!!!!" and "THE MAC PRO IS NOT A REAL WORKSTATION!!!!" or "BRING BACK THE CLASSIC MAC PRO!!!!"

 

[Roykor]

You must be new around here. All we talk about is how "APPLE DOESN'T CARE ABOUT PROFESSIONALS!!!!!" and "THE MAC PRO IS NOT A REAL WORKSTATION!!!!" or "BRING BACK THE CLASSIC MAC PRO!!!!"

If you sum it up like this... its even worse that I think it was. No information about an update, not even strong rumors, and a lot of people feeling not being back-uped by there favorite brand (who is married with the creative world but is crashing into a divorce)

What do we have? We have news and rumors about the Apple Car, the iPhone 6s touch id and 4k video (when will there be a iPhone "pro" ?), new macbook 12" with more fancy colors, Apple Watch 2 rumors, iphone 7 rumors, Apple car stuff (tech for inside a car), more iOS beta news, Intel maybe for their iPhones...

I just summed up the last two weeks of news on this frontend website. If i continue it will not be very different. The name 'mac'rumors is from the past. I even don't know why visit this website. Its more a bit of my personal system by visiting the same websites every day. But havent read anything interesting the last 2 years really. And thats Apple for me in a nutshell. Not really interesting anymore. And its getting worse every day. Typing this on my iMac with ElCaptain, the slowest OS i have had on my 2011 system with tons of annoying issues.

Still using it.. i have to. I have spend serious cash on my ecosystem.

 

[koyoot]

http://www.game-debate.com/news/?ne...ly Offers Near 980 Ti Performance For 300 USD

AMD reportedly hosted an event designed to showcase its upcoming Polaris GPUs and the Radeon Pro Duo to journalists behind closed doors in Taiwan recently, ahead of an expected official unveiling in May. The big noise coming out of the event is that the switch to the 14nm FinFET fabrication process means the Polaris 10 GPU performs extremely close to the GeForce GTX 980 Ti, but for a drastically cheaper price point.

 

[Stacc]

If you sum it up like this... its even worse that I think it was. No information about an update, not even strong rumors, and a lot of people feeling not being back-uped by there favorite brand (who is married with the creative world but is crashing into a divorce)

What do we have? We have news and rumors about the Apple Car, the iPhone 6s touch id and 4k video (when will there be a iPhone "pro" ?), new macbook 12" with more fancy colors, Apple Watch 2 rumors, iphone 7 rumors, Apple car stuff (tech for inside a car), more iOS beta news, Intel maybe for their iPhones...

I just summed up the last two weeks of news on this frontend website. If i continue it will not be very different. The name 'mac'rumors is from the past. I even don't know why visit this website. Its more a bit of my personal system by visiting the same websites every day. But havent read anything interesting the last 2 years really. And thats Apple for me in a nutshell. Not really interesting anymore. And its getting worse every day. Typing this on my iMac with ElCaptain, the slowest OS i have had on my 2011 system with tons of annoying issues.

Still using it.. i have to. I have spend serious cash on my ecosystem.

There was a good deal of sarcasm in my post. If I were to try and summarize most of the posters here, most are frustrated with the fact that the mac pro hasn't been updated in 2.5 years, especially given that there have been suitable components that could be used to update the machine. Apple doesn't communicate its plans, we know the game.

I think most of us are aware of the general trends in computing. Moore's law has ended, computing performance has somewhat plateaued and significant jumps in performance are hard to come by these days. Intel struggled significantly to shrink its processors to 14 nm and graphics card manufacturers are just about to release cards at this size. It may appear that Apple is neglecting its macs, but its only representative of the same trend across all computer manufacturers.

There are many who post here who are frustrated by the design of the mac pro. Apple has simultaneously reduced the updates to the machine and made the machine harder/impossible to upgrade yourself with standard PC components. External expansion via thunderbolt is a nice idea, but it adds cost and limits the ability to upgrade components like graphics cards.

That said, this will be a good year for the mac. Thunderbolt 3, skylake/broadwell-ep should finish spreading across the lineup and some macs should get updated graphics cards with a big jump in performance. Things that would make it a great year for the mac would be things like an external retina display, support for external graphics cards or a headless mac that fits between the mac mini and the mac pro. Going forward though, these slower update cycles are likely to become the norm.

http://www.game-debate.com/news/?news=20006&graphics=Radeon R9 490X 8GB&title=AMD Polaris 10 GPU Reportedly Offers Near 980 Ti Performance For 300 USD

The rumors coming out about Polaris have been pretty juicy. I am psyching myself up for a WWDC mac pro announcement.

 

[ManuelGomes]

Yep, expecting a huge performance improvement with the new nMP will get you disappointed, mostly small bumps nowadays. Still, at least having the latest (or close) tech inside is surely a nice welcome.
The days of performance jumps are gone.
I'm hoping we finally get some good stuff but let's wait and see.
Polaris looks good, let's hope it lands on nMP.
I'd like to see the new rTBD along with it, 10b HDR would look sweet. Too bad it's not yet possible at 5K.

 

[koyoot]

I personally expect that single D710 or whatever it will be called will offer around 7 TFLOPs of compute performance.

 

[tralfaz]

A few corrections:

It's "silicon" not "silicone." That's an important distinction when dating as well as when making chips.

HEHE How to invalidate a post with a single word.

 

[Hank Carter]

Yep, expecting a huge performance improvement with the new nMP will get you disappointed, mostly small bumps nowadays. Still, at least having the latest (or close) tech inside is surely a nice welcome.
The days of performance jumps are gone.
I'm hoping we finally get some good stuff but let's wait and see.
Polaris looks good, let's hope it lands on nMP.
I'd like to see the new rTBD along with it, 10b HDR would look sweet. Too bad it's not yet possible at 5K.

If they went back to a dual CPU machine we would see a substantial boost in performance, but that will probably not happen since Apple is prioritizing design over functionality and performance. Kind of a silly thing to do in a workstation. It's supposed to be a tool, not a fashion accessory.

 

[AidenShaw]

The rumors coming out about Polaris have been pretty juicy. I am psyching myself up for a WWDC mac pro announcement.

In the last day or two, there was a comment about how underwhelming and disappointing some of the Apple hardware announcements are. (Just check the comments across MacRumors after many of the recent ones.)

Look at the calendar. Computex is 31 May to 4 June. 9 days later is MacWorld SF 2016.

Nvidia is expected to announce at least part of their range of consumer Pascal (GP104 series) in Taipei. If Nvidia "knocks one out of the park" with the Pascal announcement (like they did with the GP100 announcement), Polaris would be "a day late, a dollar short" at MacWorld.

On the other hand, if Apple announces that the GP104 will be in the next cylinder - they could look very good!

 

[fuchsdh]

In the last day or two, there was a comment about how underwhelming and disappointing some of the Apple hardware announcements are. (Just check the comments across MacRumors after many of the recent ones.)

Look at the calendar. Computex is 31 May to 4 June. 9 days later is MacWorld SF 2016.

Nvidia is expected to announce at least part of their range of consumer Pascal (GP104 series) in Taipei. If Nvidia "knocks one out of the park" with the Pascal announcement (like they did with the GP100 announcement), Polaris would be "a day late, a dollar short" at MacWorld.

On the other hand, if Apple announces that the GP104 will be in the next cylinder - they could look very good!

What is the relevance of MacWorld? Do you mean WWDC?

 

[AidenShaw]

What is the relevance of MacWorld? Do you mean WWDC?

Yes, MacWorld SF 2016, 13 June at Moscone.

The "relevance" is that back in the reign of the Lord God Jobs, Apple announced that they didn't need user conferences, and pulled out of the east coast and west coast MacWorld conferences.

Yet, not too long after, Apple realized that *they did need* user conferences - and co-opted the keynote of their developer conference for public announcements many of which are at most vaguely related to the developer tracks.

 

[fuchsdh]

Yes, MacWorld SF 2016, 13 June at Moscone.

The "relevance" is that back in the reign of the Lord God Jobs, Apple announced that they didn't need user conferences, and pulled out of the east coast and west coast MacWorld conferences.

Yet, not too long after, Apple realized that *they did need* user conferences - and co-opted the keynote of their developer conference for public announcements many of which are at most vaguely related to the developer tracks.

What is the point of being so back-handed about everything? You must be a hoot at parties.