Sidebar Sidebar
Become a MacRumors Supporter for $50/year with no ads, ability to filter front page stories, and private forums.

Will there be a new Mac Pro in 2015?

Status
Not open for further replies.
ManuelGomes said:
But with the Fiji info on El Cap it could in fact be a hint of what's to come.
Click to expand...

Maybe they shoehorn Fiji to 5k iMac for now. They could do it with 65W Skylake and give the +20W saved from CPU for GPU. For nMP there is Tonga Pro and XT chips (D310 and D510 with 2GB and 4GB of vram). If 384-bit Tonga arrives, that could be D710 with 6GB. So nMP would be all Tonga.

Update: Getting rid of GCN 1.0 would give Apple another new feature that they could add to OS X: DSP. In Logic Pro for instance, they could utilize the DSP and save CPU power for other features. Games too would benefit for 3D sound without CPU penalty. DSP comes with GCN 1.1 version and later.
 
Last edited:
When I said Hawaii I meant to say the "new" iteration of it, Grenada. OK, it's been optimized but it's still pretty much Hawaii, when it comes to specs. Optimized process but that's it really.
I meant that Fiji can be molded when it comes to number of cores, but not on the amount of memory. It's a limitation of HBM right now, will not be with HBM2. With any other GPU you can more or less (depending on the actual architecture) choose the amount of memory in the card. Some cards can handle 2, 4 or 8GB, or even 16GB, others can have 3GB or 6GB. But Fiji, due to being an HBM1 card, is locked in at 4GB and you can go no further. That will not allow you to have different SKUs based on the amount of memory in the card, which might be a problem for nMP which, as you know, at the moment very much relies on this differentiation for the Dx00 cards.

Z, that is indeed an option, but it will feel like no real upgrade at all from Tahiti. It's a newer cards based on GCN 1.2 but no real memory upgrade and with cards nowadays with huge amounts of it seems pale. The mid card D510 gets an extra 1GB but comes down in bus width from 384b to 256b.
Looks reasonable? Maybe!!
 
ManuelGomes said:
Z, that is indeed an option, but it will feel like no real upgrade at all from Tahiti. It's a newer cards based on GCN 1.2 but no real memory upgrade and with cards nowadays with huge amounts of it seems pale. The mid card D510 gets an extra 1GB but comes down in bus width from 384b to 256b.
Looks reasonable? Maybe!!
Click to expand...

GCN 1.2 provides optimization for bandwidth utilization. They can do nearly same things with 256bit bus as older architecture did with 384bit. And perf/Watt ratio is better with Tonga. So with the TDP limitations of nMP Tonga can do more.
 
Last edited:
Indeed, that argument might stick as far as marketing goes, and it certainly is a good one, and real.
Still, the number of cores will be lower in the high end cards, which could be a limiting factor.
And the integrated DSP you mentioned is a great feature, I was hoping for support in OS X once newer cards with it came out.
 
Amount of memory in GPU is becoming irrelevant when you get to OpenCL 2.0 and virtual memory/shared memory. The same goes for Metal. More important is bandwidth here in VRAM.
 
Sure, tell that to someone who's gonna buy a new rig with what seems to be an inferior card.
Most people will care for the numbers.
And accessing local VRAM will always be faster than having to go to a remote pool.
It seems the 4GB in Fiji is enough for most workloads but once you run out of local storage the penalty (higher latency) will be high.
Not saying it's a bad decision, on the contrary, but it could be not so peaceful.
What troubles me is the lack of evidence of what's to come in El Cap. Those bits of info available are not exactly definitive about anything. Maybe the final build in a couple of weeks will bring more info.
 
I don't think anyone who does not know anything about computers will buy computer like Mac Pro for 5-6K USD ;).
 
  • Like
Reactions: rdav
If they are not techies, you really think the difference between 4 and 6 will be a problem for them? ;)
 
R9 Fury Nano is using very aggressive power saving features. It scales voltage and clock speed very fast and responds to the demand. If Nano is an option for nMP, this feature could share the 250W pool that is available for two GPU's in a way that when one GPU is idling, another gets more power, even 200W. And when they are working together, they are rarely 100% occupied both at the same time. If handled properly with drivers and HW, double Fiji will work fine in nMP and give optimal performance. That could make nMP also a gaming machine with one GPU / second idle even if there is no CrossFireX support in Metal in the future.. a 5000 € gaming machine.. but anyway.. :p

There's an interesting analysis of the new Nano at Anandtech site. Take a look, if you haven't yet.
http://www.anandtech.com/show/9621/the-amd-radeon-r9-nano-review
 
Last edited:
I somehow got the feeling that the people in this thread know much more about Computers than the Apple engineers and that Apple reads this topic to find Solutions xD

“Look Jeff. #538 said that two tonga gpus will fit if we save 20% power on the cpu“

“but how can we explain that the d710 has less vram??“

“#1337 said that hbm is better. Quick. Send this to our marketing team“
 
Not really. You get 2 GB less but with twice the amount of bandwidth. If you have data that is stored in memory for "future" utilization - that is waste of memory in the moment. If you have smaller amount with gigantic bandwidth that the core will utilize - thats where you don't get bogged up.

The most crucial part is how it will be utilized from technical point of view. From marketing it is a bit of a problem, but always you can turn to the bandwidth. You also have to take into account how many times a second you can refresh the data stored in both cases(6GB GDDR5 vs 4 GB HBM) every second. And with HBM it is massively bigger amount. Just because of the extreme bandwidth you play with.

But I will give you something technical to think about. GCN is out-of-order architecture. And with that, think what and how it can work with virtual memory/shared memory in OpenCL 2.0/Metal.
 
koyoot said:
Not really. You get 2 GB less but with twice the amount of bandwidth. If you have data that is stored in memory for "future" utilization - that is waste of memory in the moment. If you have smaller amount with gigantic bandwidth that the core will utilize - thats where you don't get bogged up.

The most crucial part is how it will be utilized from technical point of view. From marketing it is a bit of a problem, but always you can turn to the bandwidth. You also have to take into account how many times a second you can refresh the data stored in both cases(6GB GDDR5 vs 4 GB HBM) every second. And with HBM it is massively bigger amount. Just because of the extreme bandwidth you play with.

But I will give you something technical to think about. GCN is out-of-order architecture. And with that, think what and how it can work with virtual memory/shared memory in OpenCL 2.0/Metal.
Click to expand...

To render scene in many 3D renderer, the whole scene must fit in the vram. If your scene is 5gig worth and you only have 4 then the rendering will crash. For OpenCL and GPGPU application, yes, HBM as the edge, but not when it comes to graphical works.
 
PhiLLoW said:
I somehow got the feeling that the people in this thread know much more about Computers than the Apple engineers and that Apple reads this topic to find Solutions xD

“Look Jeff. #538 said that two tonga gpus will fit if we save 20% power on the cpu“

“but how can we explain that the d710 has less vram??“

“#1337 said that hbm is better. Quick. Send this to our marketing team“
Click to expand...
Well, you know, when the topic is "Will there be a new Mac Pro in 2015?" and Apple wont tell... our next option is to go through all the available options and throw (more or less) sophisticated guesses.. Apple has made it's choices long time ago, but we try to predict them. Just for fun.
 
  • Like
Reactions: rdav
Yeah, sometimes people tend to forget that this is supposed to be fun and come guns blazing...
I wonder when the nMP really comes out what will happen to this thread? For a while we'll go something like "I told you it would be like that!" or "I was right" or even "We were all completely wrong!".
Or there will be no nMP anytime soon and we keep at it :)
Or we start a new thread "Will there be a new Mac Pro in 2017?", which is by the way unlikely - unless of course there isn't one till then :-(

koyoot, I was also more concerned with the renders. The rest is not really an issue.
Also, bandwidth is good and all but I wouldn't consider a waste having data in local mem even if it isn't being used at the moment, sooner or later you'll need it and fetching it from elsewhere will cost you a butt load of cycles.
I'm not just picking on it, don't take me wrong. For me those 4GB would be just fine. But I reckon there are those that need more, possibly the ones that got the D700 exactly because of that.
I would settle just fine with an 8 core and a couple of those Nanos in there :)
Just what the doctor ordered :)
 
SoyCapitanSoyCapitan said:
AMD has just spun off the graphics business into a new division called Radeon Technologies Group.
Click to expand...

AMD spun off nothing; a division isn't a spin off. Folks who work on GPUs all report to the same person. That wasn't the case before ( several groups do GPU work embedded inside of CPU groups ). There are still GPU teams working to integrate onto the same die as the CPU folks.
 
Broadwell -E ( Core i7 59xx ) doesn't have to be perfectly coupled to the Xeon E5 v4 ( -EP ) but they share related technology and cihipsets.

http://www.cpu-world.com/news_2015/...PUs_slightly_delayed_planned_for_Q1_2016.html

Early Q1 2016 appears to be out (if this info is correct) . Looks like mid-end of Q1 ramp and volume production.
Sliding Broadwell isn't going to help pull these product classes of Skylake iteration forward in time.

Previous launches the -E models tended to make it out the door first. Maybe Intel switched the priority order and hence small slide for -E models. I still wouldn't expect it in 2015 though. Ramped and volume for Xeon E5 v4 could be early Q1. But if the samples are being delayed due a defect that is common to both derivations, then E5 v4 would probably slide too. Another round of info should trickle at Supercomputing '15 in november about E5 v4 if Apple says nothing in October.
 
Status
Not open for further replies.
Register on MacRumors! This sidebar will go away, and you'll see fewer ads.
Top Bottom
Back