HBM 1st gen allows you to get 1 stack of 4 GB of Memory per chip(GPU). Not more.
Sad, but only HBM 2nd gen allows you to go to 8 GB. And more, and more...
Max clock on 1st gen is 1.4 GHz, 2nd Gen will allow 2 GHz. Yup, 2TB/s with 2nd gen...
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.
nMP wishlist
- Thread starter ManuelGomes
- Start date
- Sort by reaction score
HBM 1st gen allows you to get 1 stack of 4 GB of Memory per chip(GPU). Not more.
Sad, but only HBM 2nd gen allows you to go to 8 GB. And more, and more...
Max clock on 1st gen is 1.4 GHz, 2nd Gen will allow 2 GHz. Yup, 2TB/s with 2nd gen...
Correct, but that's if the next gen is indeed based off Caribbean Islands, which I'm hoping is not, otherwise it will take too long. February doesn't seem real to me.
If, on the other hand, they'll use the existing W7100/8100/9100 then it should be double mem like I said. The retail cards have between 8GB and 16GB, so for the nMP half that seems a logical choice - although full mem would be great, but GDDR5 is quite power hungry so I guess they'll cut it down for power optimization's sake.
Anyway, I find it very hard to believe Apple would release cards with that amount of memory at this point in time. No pro cards with that low amount of memory would be credible in 2015, and that would limit too much people who work with very large frame buffers, the current ones are already on the short side I'd say.
If they wait for HBM 2nd gen then I'm afraid not even in 2015 we get to see an update...
If, on the other hand, they'll use the existing W7100/8100/9100 then it should be double mem like I said. The retail cards have between 8GB and 16GB, so for the nMP half that seems a logical choice - although full mem would be great, but GDDR5 is quite power hungry so I guess they'll cut it down for power optimization's sake.
Anyway, I find it very hard to believe Apple would release cards with that amount of memory at this point in time. No pro cards with that low amount of memory would be credible in 2015, and that would limit too much people who work with very large frame buffers, the current ones are already on the short side I'd say.
If they wait for HBM 2nd gen then I'm afraid not even in 2015 we get to see an update...
If they were going for using Tonga/Hawaii GPUs they've already would been introduced to Mac Pro.
Everything is on market in that scheme. They are waiting for new versions of AMD GPUs.
P.S. If Im counting correctly, going for 16GB alone gives you 80W of power consumption on Memory. That is at least 50W of wasted power.
It's not just a simple case of getting new AMD silicon. Apple has to customise the design and reduce the TDP by almost half so that the GPUs can run fanless alongside the Mac Pro power supply. Maybe they can get silicon early, maybe they will increase the power supply capacity, but whatever happens I think the Mac Pro line is being abandoned and they will promote iMacs for pro usage and then finally tell us to just use cloud computing.
Last edited:
if they did a cut-down 6-8 core with single powerful gaming GPU, i'd be very happy. non-openCL users pay a large overhead so that final cut and other specific app users can enjoy a large benefit as it currently stands.
i also hope they come up with a viable 5+K (why not 8, future proofing once and for all?) cabling solution for monitors on the next gen nMP.
with DP1.3 / TB3 still a way off, maybe these aren't the solution and a new format specifically for this would be better? as i understand it anyway, the problem is not GPU horsepower, but cable bandwidth?
then use TB in a combined cable to carry secondary functions so that users can use the new gen of 4,5 and 8K 'new connector' screens in the same way as they use current thunderbolt displays, using the other TB ports to mass storage throughput ala firewire or lo-res screens. it just seems messy as is.
just a thought, there may well be reasons why this couldn't work, isn't practical (probably down to intel chipsets!)
i also hope they come up with a viable 5+K (why not 8, future proofing once and for all?) cabling solution for monitors on the next gen nMP.
with DP1.3 / TB3 still a way off, maybe these aren't the solution and a new format specifically for this would be better? as i understand it anyway, the problem is not GPU horsepower, but cable bandwidth?
then use TB in a combined cable to carry secondary functions so that users can use the new gen of 4,5 and 8K 'new connector' screens in the same way as they use current thunderbolt displays, using the other TB ports to mass storage throughput ala firewire or lo-res screens. it just seems messy as is.
just a thought, there may well be reasons why this couldn't work, isn't practical (probably down to intel chipsets!)
That was my point exactly, it's not just enough that the GPUs are available for Apple to release cards based on them.
Remember that Apple's cards are usually not the same as retail cards, different PCB of course but even the specs are sometimes customized.
That takes time to refine, to validate and to market.
And that's why I said that even though retail cards will have 8 and 16GB, the Apple nMP cards will have half that because GDDR5 is very power hungry, reducing the mem amount will help fit the power envelope.
Or Apple will beef up the power supply which I believe to be very unlikely.
Let's hope they don't kill off the MacPro line, I'm not sure iMac would be a good solution that handles all.
TB3 will handle 5K, not sure about 8K though. Math required to see if it fits the available bandwidth.
But TB already carries multiple data, that's the beauty of it really.
Remember that Apple's cards are usually not the same as retail cards, different PCB of course but even the specs are sometimes customized.
That takes time to refine, to validate and to market.
And that's why I said that even though retail cards will have 8 and 16GB, the Apple nMP cards will have half that because GDDR5 is very power hungry, reducing the mem amount will help fit the power envelope.
Or Apple will beef up the power supply which I believe to be very unlikely.
Let's hope they don't kill off the MacPro line, I'm not sure iMac would be a good solution that handles all.
TB3 will handle 5K, not sure about 8K though. Math required to see if it fits the available bandwidth.
But TB already carries multiple data, that's the beauty of it really.
No need for faster than 1gbps Ethernet when nothing else in your setup can profit from it. The nMP is a workstation, not a server farm.
Yes, but this isn't Ibgb was wishing for. Every pc I built in the last 5 years have 2 1gb Ethernet port...
What tangential line of thought are you on? server farms? How does that connect ???
Who does not want IO faster than 120 MB/s ? Why no need for more than 120 MB/s? Ethernet is the most standard data interconnect; why not 1000 MB/s ?
a) Internet/thought noise
b) lack of caffeine
c) posting while distracted
?
(Sometime I do a+b+c--but I really try not to!)
Do you own any hardware in your home that need more than a 1gbs Ethernet connection? You do know that you already have a 20gbs connection on you nMp via TB?
The only place presently commonly using more than 1gbs Ethernet are data centre aka server farm.
Nas
Virtually all current spinning disks are faster than 1 GbE - so file sharing in all forms is bottle-necked by 1 GbE.
Virtually all current spinning disks are faster than 1 GbE - so file sharing in all forms is bottle-necked by 1 GbE.
Most people, almost everywhere I have worked too, when we had media files and large images shared on the network we worked directly off the server without a need to copy the files locally.
On The Needs for Many Ports
In short: My NAS* does not have a Thunderbolt port, but it does have dual GB ethernet ports for trunking. The USB port is useless because you can't use the NAS* as direct-attached storage.
PS: I won't be so brazen to say that my NAS* has two useless eSATA ports because I'm sure someone out there uses eSATA.
*I won't mention any names, but the initials are QNAP.
In short: My NAS* does not have a Thunderbolt port, but it does have dual GB ethernet ports for trunking. The USB port is useless because you can't use the NAS* as direct-attached storage.
PS: I won't be so brazen to say that my NAS* has two useless eSATA ports because I'm sure someone out there uses eSATA.
*I won't mention any names, but the initials are QNAP.
LACP/Trunking does not actually yield 2 gigabit data transfer rates for a single computer. Your max transfer is still limited to the max throughput of a single gigabit ethernet connection. It makes sense on the NAS side because you will conceivably have multiple computers all simultaneously trying to transfer data. For a single computer connected to a single NAS, though, it's rare to be able to exploit the aggregate speed.
Any single socket connection (file transfer) between the devices will be assigned to one of the two ethernet channels. Generally this is done based on a hash of the source and destination IP address so it will always be the same ethernet channel between the same devices. Some switches let you factor in port number also, which can mix things up a bit, but only if your use can be split across different source ports at the same time). AFP or SMB won't make use of the bonded channels for a single filesystem mount.
Any single socket connection (file transfer) between the devices will be assigned to one of the two ethernet channels. Generally this is done based on a hash of the source and destination IP address so it will always be the same ethernet channel between the same devices. Some switches let you factor in port number also, which can mix things up a bit, but only if your use can be split across different source ports at the same time). AFP or SMB won't make use of the bonded channels for a single filesystem mount.
Fake Double-Ethernet ports on Mac Pros
Wait, so, there's no purpose in connecting two ethernet ports from a mac to two ethernet ports on an NAS that supports dual gigabit ethernet trunking and jumbo frames, etc? It's all fake? All those TB and USB to En0/En1 adapters are useless and the "Dual Gigabit Ethernet" is really just two jacks spliced together?
There outta be a law... http://www.apple.com/mac-pro/specs/
Wait, so, there's no purpose in connecting two ethernet ports from a mac to two ethernet ports on an NAS that supports dual gigabit ethernet trunking and jumbo frames, etc? It's all fake? All those TB and USB to En0/En1 adapters are useless and the "Dual Gigabit Ethernet" is really just two jacks spliced together?
There outta be a law... http://www.apple.com/mac-pro/specs/
I have no idea why you are as frustrated as you seem to be here. No, it's not "fake." It just doesn't work the way you thought it did. If there are only two devices (a single computer and a single NAS, for example) it's probably not going to offer any benefit. That doesn't mean that there's never a benefit. If you have ten computers and a single NAS, it's a great win (on the NAS side). It just depends on what you're hoping to accomplish.
Just to be clear, because I'm having a tough time understanding your post. The two gigabit ethernet jacks on a Mac Pro are absolutely two gigabit ethernet jacks. The theoretical bandwidth is 2 gigabits. There's just does not exist a way to spread a single file transfer or a single mounted filesystem's activity across both ports. When you copy a file across from your NAS it's only going to traverse one of the two bonded ports. You will only ever see a benefit from LACP/bonded ethernet if your network activity is concurrent and spread among multiple host IPs (or, possibly IP port combinations).
I agree, there should be a law against people that don't understand common computer standards that have existed for many years and then act like it is someone else fault when their misunderstanding doesn't work the way that they thought it should.
Link aggregation is very commonly used with servers and storage arrays.
Nugget has done a good job of explaining it so I won't bother to repeat what he has said. If you don't understand what he has said I would recommend reading the link above and then follow the links within for additional reading.
Abraham Lincoln said:
Yes, nMP. A 1 gigE link limits it to 120MB/s IO. I have to move projects to/from storage.
You can buy motherboards with good 10 gigE ports for a few hundred $ more. I have an empty pci slot, waiting for a 10 gigE card, but it may be a while.
It'a like dual processors. One task can only be one on one processor at a time. One transfer can only exist on one ethernet port at a time. If you are doing two transfers, it can load balance the transfers though.
I don't know what that means for individual network file protocols though. If stuff like Samba or AFP opens a new socket for each file operation, you might be able to do two copies to the same NAS at once with double speed, but that could also not be the way it works at all.
Two connections to two different hosts should definitely run over two different ports.
(So yes, for most people channel bonding is totally useless. Can be useful for a server though.)
Based on my last post in the USB 3 discussion, I predict there will be USB 3.1 on the 7,1 and Apple is currently customising the new GPUs. I would hope they are moving to Nvidia's Maxwell architecture as AMD's Caribbean Islands specs have leaked and are not as power efficient as Maxwell. Rumour is that NVidia are also skipping 20nm process and going directly to 16nm.
AMD's Architecture is FAR MORE power efficient than any Nvidia architecture.
Doesn't matter how power efficient is Maxwell architecture, Apple will always cut TDP of the GPUs to around 125W at max.
Don't know where you get this from. The leaked specs put Carribean Islands lowest spec TDP at 195-205w. Power efficiency, performance per watt, looks to be about the same (however, NVidia will move to a 16nm process first)so please refrain from exaggerated forum broisms like "FAR MORE"
But then there is Open CL performance that AMD might still be ahead on as NVidia prefer developers to use CUDA. But the GTX 980 improved CL performance so much over the 780 that maybe NVidia will finally give great performance in both APIs.
Last edited:
Yes, and R9 380X is around 20% faster using 15% more power. That is power efficiency. Secondly, getting 3.5 Tflops from 129W of power is power efficiency bigger than Nvidia is capable. GCN architecture is the most power efficient architecture in the world, right now, but mainstream does not admit it.
Secondly, if you believe that marketing gimmick from Nvidia than yes, Nvidia GPUs are more power efficient. In games.
At whatever work GTX980 draws around 280W, or... it is really slow. Sustained power envelope for Maxwell cards is a problem.
Yes, it will be the same, for Caribbean Islands. Nonetheless - they will be still faster in similar power envelope.
Edit: 195W is for 4096 GCN core GPU. Almost the highest end GPU that will become R9 380X.