koyoot
macrumors 603
VR. Gaming Console capabilties. HTPC for living room. Metal shows that they care about their platform, after all 😉.
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.
2016 nMP
- Thread starter ManuelGomes
- Start date
-
- Tags
- nmp
- Sort by reaction score
- Status
If Apple is going to tell their plans about VR at WWDC... One of the GPU makers will be on platform.
I'll don't discard a dual fury nano on the nMP (still fit the 110W/ gpu TDP) but.. will it worth?
The same Apple could drop AMD for nVidia on some products or all not the first time..
What I don't see is Apple using nothing x86 (Zen APU) on something iOS based it will not happen.
Apple could end using a Zen APU on the retina iMac, and eventually a Zen cpu and Vega GPU on the Mac Pro, but this won't happen this year, maybe next year.
(assuming true Zen performance figures it could mean a big win for the Mac Pro over the Xeon, since it's much faster single thread).
Anyway about the Mac Pro nothing it's said, only leaks are about Thunderbolt 3, we don't know anymore, Ahh Apple is collaborating on CUDA 8.
koyoot
macrumors 603
Ive got an answer also why 2560 GCN core GPU with only 256 bit memory bus will be 232mm2 not less. Because the compute units inside the core are bigger due to the new architecture of power gating and management of the CU.
On this matter most informative post so far: http://forums.anandtech.com/showpost.php?p=38159428&postcount=155
I am not convinced we will see an AMD CPU in a mac anytime soon. Sure, lots of nerds have been talking about how great Zen is going to be for the last two years but it remains to be seen if they can actually compete with Intel. It seems like even the most optimistic expect it to be on par with Haswell, which is to say it will be 2-3 generations behind a modern Intel chip. If thats the best AMD can offer than I don't see Apple downgrading from an Intel CPU to an AMD.
Lets assume a Zen/Polaris APU still can't match the performance/watt of an Iris Skylake CPU. What product would Apple offer zen in? Not in any laptop since it would sacrifice battery life, and in the desktops with discrete GPUs there is more than enough space to add in a normal discrete GPU.
Sure, AMD may bring CPUs with more than 4 cores to consumers, but I have yet to see a case that most consumers benefit for more than 4 cores. I bet Apple sells way more machines with dual cores than quad cores.
You are on point here. Apple is unlikely to be jumping on the desktop VR bandwagon anytime soon. Let Oculus and Valve go through the early growing pains. If there is a market there then maybe Apple will enter it. Sure, VR is a cool technology but I still haven't seen anything resembling a killer app. Remember, Apple wasn't the first out with a mp3 player, smart phone, tablet or smart watch, they just jumped in when they felt like they could offer a significant improvement over the rest of the market. Also you had better believe it will be a mobile first effort, since if there is one thing we learned at the last Apple press event, its that Apple still sees lots of growth in their mobile products and isn't afraid of trying to cannibalize the macs to do it.
Lets assume a Zen/Polaris APU still can't match the performance/watt of an Iris Skylake CPU. What product would Apple offer zen in? Not in any laptop since it would sacrifice battery life, and in the desktops with discrete GPUs there is more than enough space to add in a normal discrete GPU.
Sure, AMD may bring CPUs with more than 4 cores to consumers, but I have yet to see a case that most consumers benefit for more than 4 cores. I bet Apple sells way more machines with dual cores than quad cores.
You are on point here. Apple is unlikely to be jumping on the desktop VR bandwagon anytime soon. Let Oculus and Valve go through the early growing pains. If there is a market there then maybe Apple will enter it. Sure, VR is a cool technology but I still haven't seen anything resembling a killer app. Remember, Apple wasn't the first out with a mp3 player, smart phone, tablet or smart watch, they just jumped in when they felt like they could offer a significant improvement over the rest of the market. Also you had better believe it will be a mobile first effort, since if there is one thing we learned at the last Apple press event, its that Apple still sees lots of growth in their mobile products and isn't afraid of trying to cannibalize the macs to do it.
Just ordered a Dell Precision 7910 with dual liquid cooled 2687w v4 (in total 24 physical cores @ 3 Ghz). 32G ram, 512ssd and nvidia quadro k2200 gpu. I also have a nvidia tesla K20 that I will plug in when it arrives. So it is a pure number-crunching machine with no big need for memory or storage. For the same price I would have had to settle with one 2,7GHz 12core, 16Gb ram and dual D500 gpu if buying a nMP. I will not leave OS X however, in addition to the Precision I have an iMac at office and a MBP at home and when traveling. I am however slightly annoyed that apple cannot sell a decent workstation. Would have bought a classic MP with up-to-date hardware in a heartbeat (with realistic pricing).
Last edited:
fuchsdh
macrumors 68020
What is the benefit of Apple switching to AMD processors? There's no evidence they're going to avoid the massive production bottlenecks Intel faces.
Bubba Satori
Suspended
What massive production bottlenecks is Intel facing? Source?
All their fabs are up and running.
They are getting more cpus per wafer with the new smaller process.
So supply is up.
Chip demand has been down the last few years.
koyoot
macrumors 603
I will tell you this, but supposedly nobody will believe me. I have written something similar 1.5 years ago on this forum.
AMD is currently working on their own CUDA-like API for professional, compute applications. It will be based on Mantle v2. What is more important it will be very similar to Metal. It will most probably be Open Source. That is very reason why AMD launched Boltzmann Initiative, and CUDA compilers for OpenCL.
AMD is currently working on their own CUDA-like API for professional, compute applications. It will be based on Mantle v2. What is more important it will be very similar to Metal. It will most probably be Open Source. That is very reason why AMD launched Boltzmann Initiative, and CUDA compilers for OpenCL.
fuchsdh
macrumors 68020
Production bottlenecks as in roadmaps and Moore's law breaking down.
koyoot
macrumors 603
Nvidia sent invites for presentation of new GPUs. GTX X80 with 8 GB of GDDR5 - 30% faster than GTX 980, with the same thermal envelope.
Boost clock is over 1400 MHz for core.
All in all: 2048 CUDA core GPU with 1400 MHz boost clock will be 30% faster than GTX 980.
New node, that provides 70% lower power consumption, or 30% higher performance at the same power target. And I was called AMD fanboy when I said that Pascal will not be as efficient as Maxwell was.
What is becoming also apparent that Pascal is only improved Maxwell. At least on lower end that the highest models.
Boost clock is over 1400 MHz for core.
All in all: 2048 CUDA core GPU with 1400 MHz boost clock will be 30% faster than GTX 980.
New node, that provides 70% lower power consumption, or 30% higher performance at the same power target. And I was called AMD fanboy when I said that Pascal will not be as efficient as Maxwell was.
What is becoming also apparent that Pascal is only improved Maxwell. At least on lower end that the highest models.
koyoot
macrumors 603
There is nothing in Pascal architecture that differentiates it from Maxwell, apart from splitting 128 CUDA core SMM's into two bits that share the same amount of Cache, both with 64 Cores. Nvidia did this to allow better asynchronous compute performance in their hardware. On lower end, the GPUs that will have 2048 GCN cores, GDDR5 and lower ROP count than Maxwell GPUs they will be different from Maxwell only through higher core clocks.
Everything has been debunked tuxon, already. Volta was supposed to be after Maxwell, but TSMC was not able to deliver 20 nm process that was supposed to be for Maxwell. Thats why Nvidia took out FP64 from Maxwell, ported it to 28 nm, and sold as GTX 9XX series. But with that, they added one another lineup between Kepler and Volta, based on 16 nm process, added FP64, and 16 and called Pascal.
EVERYTHING falls in line perfectly. Only people who want to reject reality do not see it.
Yes, there is, and it's not trivial, pascal unified memory it's radical new and allow each sm to address the full address range inside the card(s) and the system also, maxwell only for a traditional "encapsulated" model where sm where restricted on memory access, requiring special steeps just to read data from other smm, now it's direct this means a lot for CUDA programming model, and actually unleash your code to do everything you can do on a conventional cpu.
To me this is a big deal on pascal and worth it over amd is Intel Xeon phi.
koyoot
macrumors 603
Unified Memory is HSA feature... And it is done through CUDA... For crying out loud, differentiate Hardware Features from Software ones...
Not the way as on pascal, check again it's a main differentiator and nothing trivial to implement
AidenShaw
macrumors P6
And there's nothing in the Vega architecture that distinguishes it from a 3D Rage, apart from.... 😉
Almost all changes are incremental when comparing adjacent generations, but over time significant progress is made.
While we have no facts about mid-range Pascal chips, the GP100 has clearly blown away other high end compute chips. Even you said that ATI has nothing to compete with the GP100. ( https://forums.macrumors.com/posts/22780361/ )
koyoot
macrumors 603
Do you know why? Nvlink. And software. On hardware level of the GPUs there is nothing that would stop Kepler or Maxwell to have the same feature. But Nvlink, and CUDA software blocks it. Why? Marketing to sell new GPUs. But that is absolutely understandable. What is not understandable however is that why people think that Pascal is new architecture, when it is only improved Maxwell with gigantic boost clocks? Why NVlink makes it different? Nvidia's Reality Distortion Field 😉.
Did I said that, or it did confirmed your bias? Polaris does not compete with GP100. 232 mm2 GPU vs 600mm2 monster. It would be funny if it would compete however...
S9300X2 competes with it in SP performance, and beats it. 13.9 TFLOP in 300W of TDP vs 10.6 300W GPU. However there is absolutely NOTHING, yet, in the market that can compete with GP100 in DP.
Not only nvLink you also need to figure how to drive bits from the sm to other sm and the system this requires silicon not present in maxwell architecture.
koyoot
macrumors 603
Shared caches between SM's and High Bandwidth Memory Controllers... There is nothing in Pascal scheme that shows what you say.
SM's share L1 cache. Cores are connected to HBM2 memory controllers, that share everything between the cores. It is improved Maxwell, it will behave exactly like fast, robust Maxwell. With slightly improved Asynchronous Compute capabilities.
Unified Memory comes from Software and Nvlink.
Last edited:
Maxwell unified memory is on software, on pascal is managed on hardware, this is from anandtech.comShared caches between SM's and High Bandwidth Memory Controllers... There is nothing in Pascal scheme that shows what you say.
SM's share L1 cache. Cores are connected to HBM2 memory controllers, that share everything between the cores. It is improved Maxwell, it will behave exactly like fast, robust Maxwell. With slightly improved Asynchronous Compute capabilities.
Unified Memory comes from Software and Nvlink.
. http://www.anandtech.com/show/7900/nvidia-updates-gpu-roadmap-unveils-pascal-architecture-for-2016
On nVidia forums there are more details, this actually an feature that enables a lot of algorithms to efficiently handle the memory ops as on large data array now you can define an array on the whole 128gb served by the 8 gpu interconnected by nvLink plus all what you can give thru system ram.
koyoot
macrumors 603
That is correct. And NOWHERE contradicts to what I have written. Unified Memory came from NVlink, and Software. Thats why Maxwell lost it, when It was ported back to 28 nm from failed 20 nm. On architecture level it is exactly the same. But NVlink allows to see whole pool of memory on GPUs to be seen as one big pool - by the software, and it comes from... CUDA. There is nothing on hardware of the chips, apart from HBM2 and shared L1 cache.
Unified Memory was before, but it was restricted to memory pool of GPU. Now, because there is direct connection to GPU through Nvlink - it is possible to expand beyond that limitation. There is no magic here 😉.
Unified Memory was before, but it was restricted to memory pool of GPU. Now, because there is direct connection to GPU through Nvlink - it is possible to expand beyond that limitation. There is no magic here 😉.
These two things contradict each other. If Pascal is faster in the same thermal envelope then by definition it is more efficient.
Maxwell was a huge step up in efficiency. If Pascal is a compute oriented Maxwell with the improvements brought about by a smaller node then it will be a fantastic chip. AMD has a lot to worry about on the compute front especially if they can't bring a bigger chip to market quickly.
koyoot
macrumors 603
No they are not. You compare 300mm2 GPU to 400mm2 GPUwith the same thermal envelope, but the smaller one is 30% faster... on a node that brings 70% power consumption reduction, or 30% increased performance. Scale of efficiency. Thats what I meant.
P.S. 2048 CUDA core GPU will be slower and less powerful than 2560 GCN4 core GPU, and will consume more power, than AMD counterpart.
2560 GCN cores, clocked at 1152 MHz. 125W TDP. 5.9 TFLOPs. 232 mm2.
2048 CUDA cores with 1400 MHz. 165W TDP. 5.7 TFLOPs. 300 mm2.
And graphical behavior of GCN4 will be... surprising for most people here.
- Status