You are just making up numbers here. This is all speculation.
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
You are just making up numbers here. This is all speculation.
That's not what a "massive production bottleneck" is.
A massive production bottleneck would be Intel Hillsboro being nuked by Kim Ding Dong.
Your guess for Polaris is probably in the right ballpark but you are underestimating Nvidia's chip. You simply half the single precision compute performance of GP100 to land at 5.7 TFLOPS. However, GP104 will likely not maintain the 1:2 SP to DP ratio and shift the balance much more in favor of SP. I imagine GP104 will exceed the performance of the 6 TFLOP Titan X.
I am not talking about full GP104 die. Full GP104 die will have GDDR5X memory, and will be Ti model. GDDR5 models will be cut down from 2560 CUDA cores. 30% higher performance over GTX 980 has 2048 CUDA core GPU with 1400 MHz turbo boost state.
P.S. GP100 has 6 GPCx10x64 Cu's = 3840 CUDA cores.
GP104 has 4 GPC'sX10x64 = 2560 CUDA cores.
GP107 has 2 GPC'sX10x64 CU's = 1280 CUDA cores.
Everything comes from this setups.
P.S. Cut down, GDDR5 1400 MHz GPU die will have 165W TDP. Full die will have slightly under 200W TDP.
Last edited:
Ive always wondered are you on contract or is this a full time gig?
Unfortunately, I do not understand the question...?
What contract? What are you talking about?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
Show me when AMD's hUMA (beyond APU where its obvious) support this efficiency on 8x2 setups (8gpu X 2 cpu) plus enhanced memory management/mapping...
I'm surprised that you're acting dumb on this one.Unfortunately, I do not understand the question...?
The real question behind that is "does someone pay you to tout AMD products here, or do you have a real job?".
And before you defend yourself by pointing out a few pro-Nvidia posts - of course when Nvidia redefines the playing field with a GPU like the GP100 you have to acknowledge that. You'd look stupid and lose all value to your overlords if you claimed "Polaris 12 will be even better". With the GP100 Nvidia has raised the bar so much that even you admit that AMD has nothing in that space.
So, now - realizing that the high end is a lost cause - you focus on spreading FUD on the unseen Pascal consumer chips - right down to quote die size, wattage, and number of CUDA cores for the unknown GP104. Even citing such an absurd metric as to how many power pins it will need.
And every time you imply that you know something that you can't cite sources on, we hear:
- my overlords told me to say this - since I'm an approved company leak
- or, I'm seriously violating NDA agreements that I've signed
[doublepost=1460512502][/doublepost]
Irrelevant.
Do they have some hot guys doing booth duty?
(Your comment was sexist, mine is parody.)
Last edited:
They do not have to. Exactly the same thing is part of HSA 2.0 and OpenCL 2.1. Which is hilarious in context of what we were discussing. Open Source initiative is sold as a magical miracle by Nvidia. You have proven me right in two things.
Show me when AMD's hUMA (beyond APU where its obvious) support this efficiency on 8x2 setups (8gpu X 2 cpu) plus enhanced memory management/mapping...
First: That Unified Memory is only software feature, which also gets backed up by CUDA 8 on OS X.
Secondly, that Nvidia's Reality Distortion Field is really strong and it is easy to fall in it.
We discussed that Pascal is updated Maxwell architecture, and that it does not bring anything new to the table(my argument), you said that it brings Unified Memory. Well I argued that it is software feature. Nothing, on the hardware, apart from arrangement of memory controllers would allow it to be magical hardware feature.
Also, people do not understand one thing. Even if SM's of Pascal are split into two parts, they do not have higher performance than SMM's of Maxwell. Clock to clock 64 CU's from Single SM from Pascal Arch. have exactly 50% of performance of Maxwell 128 CU's from SMM.
Do I spread FUD?
https://www.chiphell.com/thread-1562478-1-1.html
http://www.bitsandchips.it/english/...ill-have-2-pcbs-base-gddr5-and-premium-gddr5x Personally I tend more believe Bitsandchips sources, than Chiphell.
Secondly: ChipHell also leaked Die shot of GP104 with 300mm2. And technical specs.
4 GPCx10x64 = 2560 CUDA cores - That is the highest end GPU.
4 GPC x 8 x 64 = 2048 CUDA cores.
About Pascal GPUs Everything I post is based on facts, rumors, and TECHNICAL analysis of architecture. I may know few things about AMD's future plans. But Pascal is ONLY my technical understanding of the architecture.
And even If I am Nvidia fan, I will ALWAYS point out every single LIE that they try to sell as miracle, and revolution.
And I do not criticise Pascal that much, as you think. It is only what you expect from me. Nowhere I did said anything about Pascal being good or bad. It is as it is, Updated Maxwell architecture, that will even behave exactly the same way, with better asynchronous compute capabilities. Show me post where I deliberately pinched it into Polaris, to compare, or even any other AMD architecture in a bad way.
Last edited by a moderator:
Hsa and nvum both do offer the same thing, with few significant differences: Hsa only implementation yet is on APU, I've read no paper confirmation it will be included on Polaris at least, while amd promises to include on a future gpu maybe it will be included on compute oriented Vega.
2nd thing you elude to admit nvum on pascal it's done at hardware level as not by nvLink (this it's only it's fabric) this implies sm cores having specific logic to integrate cache board memory and nvLink and pci buses, on maxwell thus it's done by software and lack of fabric simplifies the smm logic by an order of magnitude. (take your time to analyze nVidia papers on pascal UM). it's not an trivial improvement.
I do support open-source and open standard, I had fights I'm this thread backing opencl specially 2.0, but there is nothing opencl has to do to keep up with cuda 8, it's like to compare programming on objective-c vs swift, and beyond that not only the programmers are more productive on cuda (it's well known beyond its associated utilities, how it's easier to port applications to cuda) but the hardware itself it's more effective (not only efficient which is closely related), having less delay thread synchronization it's less an bottleneck on nVidia pascal than any other platforms, it's something you can't argue against period.
Currently I don't use symmetric GPGPU because thread synchronization it's an huge bottleneck for my specific purpose but pascal quits this bottleneck, at the point I'm considering by first time to build an Linux based gpu server in case Apple still insist on amd only Mac Pro, at least for production Pascal seems what I was waiting for (at the moment I only considered Xeon phi or high core count Xeon ), while still using the Mac as development platform.
Closest competitor to Pascal it's Xeon Phi knights landing (on Intel's omni path).
Last edited:
Why do we even compare Polaris to GP100? Why this always has to be AMD vs Nvidia? And secondly, why do you think that I discard this feature as bad? I only say that it is only done through software. Because Nvidia has both software and hardware, they were able to optimize both of them. Nothing on hardware level, apart from Memory Controller arrangements in GPClusters would imply that. Only thing that comes to my mind is that MegaThread Engine has been updated to handle 49 bit packages of data. Small thing, but changes a lot. But without Page correction it will be useless, unless Pascal has it. I don't know that for a fact, try to confirm this, if this is apparent on Pascal - then we know all we needed.
And yes, all of this is part of specification of HSA 2.0, and features that OpenCL 2.1 uses. And I am not saying that one solution is better than the other. Keep in mind that.
P.S. Like I have said: Pascal is updated Maxwell. What you do not know is that it is very similar to GCN architecture currently. Im wondering what future holds, for Apple...
And yes, all of this is part of specification of HSA 2.0, and features that OpenCL 2.1 uses. And I am not saying that one solution is better than the other. Keep in mind that.
P.S. Like I have said: Pascal is updated Maxwell. What you do not know is that it is very similar to GCN architecture currently. Im wondering what future holds, for Apple...
Last edited:
But you never do the same for AMD which has way worst record on the book in that matter... Hence why we honestly consider you a shill...
You lack honesty and integrity.
If I will see anything that AMD lied about anything in recent months I will pinpoint it(Overclockers dream, anyone?). Hardware is only thing what for me is interesting. Regardless of brand. If AMD would be dishonest, recently about anything, I would point it. So far, everything comes to OpenSource software, and hardware, that I analyzed. Where do I lack honesty and integrity? Pinpoint exactly. It always is beyond me, when someone brings message is attacked.
Why not attack then that? Why none of you prove me wrong? Why not focus on hardware, and educating ourselves, and attack a person?
Most Mac's are old; mini, pro, air, MBP 13" and 15"
- Pro runs with E5 v2 and AMD GCN 1.0
- Mini with Haswell
- Air without retina
- MBP 13" is actually quite slow to be a "Pro", and if Air ever gets retina, one of them is eating others markets.
- MBP 15" is still in Haswell era
Thunderbolt display is old. But it is still available.
- no retina
- five year old
- TB 1.0
Software is lacking;
- openGL seems to be on a legacy state, only receives GPU driver updates, but no new features for three years
- openCL has stayed at 1.2 and although it has received a lot of bug fixes, v.2.0 or v.2.1 are not here
- Pro apps have been on a silent mode. No big updates for a long time.
They all seem to be on a wait state.
There has to be coming a new vastly improved Mac line very soon. Or big changes in the OS. It is like Apple has been waiting for something. Is it
- complete Metal (with full GPGPU,GPU support)
- TB3/DP 1.3
- New GPU's?
- ARM A10?
- New displays?
- A totally new idea to use a computer?
Apple has invested a lot of money on OS X even lately and just brought Metal to it, so there has to be a future for Mac.
I belive that Metal is in a central part of OS X / Mac's future. GPGPU and GPU will be both done 100% with Metal. CPU's role will become less relevant for most users. (Even on Windows most games run around as fast with Core i3 as with i5... 2 core, 4 threads is enough for most. It's the GPU that counts.) Just Pro users will care that they can get the most out of their CPU's. CPU's have also hit a wall.. only in power efficiency we've sen a lot of improvement. Sure core count went up too, but outside Pro use there's little benefit for them.
My guess is that more and more software start to utilize GPGPU in their processes. Apple just needs to create a good programming toolset that make it happen. I presume Apple will create a competitor for CUDA. With a catchy name of course. It will be a programming library for Metal that anyone can use easily.
"The whole is a sum of its parts". Apple didn't want to release just parts. Small updates. They want it to be a whole package. That's why I believe, "there's a great big beautiful tomorrow" for Mac's. It has been missing key components, so they haven't released any updates recently. All their resources have been used for the new, upcoming things.
Last edited:
I may have to retract this statement. 64 CU's from Pascal hardware have the same number of threads and wavefronts as 128 Maxwell CU's. So clock-to clock it looks like 64 CU's have 90%(because lower amount of L1 cache available to each CU in Pascal, but overall amount of cache if we compare 128 SMM - 128 SMP is bigger for newest architecture(128 vs 96) of 128 CUDA cores from Maxwell.
VERY interesting, indeed. And those CU's will have much higher core clocks. Interesting times, are coming.
Or they are content to do let the entire lineup wither on the vine as they have been since 2010. Do a drive-by update now and then, before moving their engineering resources on to their main growth businesses.
Apple should take note on Tesla launches, they don't have "model year" they just sell the updated products ASAP are ready for production.
Apple it's so sicologically engaged on "launch events" they both miss sales opportunities and mess their users.
Apple it's so sicologically engaged on "launch events" they both miss sales opportunities and mess their users.
At last you realize pascal isn't just a simple maxwell rebuild on 14nm.I may have to retract this statement. 64 CU's from Pascal hardware have the same number of threads and wavefronts as 128 Maxwell CU's. So clock-to clock it looks like 64 CU's have 90%(because lower amount of L1 cache available to each CU in Pascal, but overall amount of cache if we compare 128 SMM - 128 SMP is bigger for newest architecture(128 vs 96) of 128 CUDA cores from Maxwell.
VERY interesting, indeed. And those CU's will have much higher core clocks. Interesting times, are coming.
Yep. This may be the worst shape the Mac lineup has been in since Jobs returned to the company. Outdated, intentionally crippled and with little or no upgradeability, I'm having flashbacks to the Performa days. It's almost difficult to believe that aside from a few speed bumps the Macbook Air is essentially the exact same machine that Jobs introduced several years ago and even still uses the original screen technology.
I guess the message is that we're all supposed to just use our iPhone or iPad and not do anything more complicated than read email or browse iTunes for mediocre top 40 music.
Last edited:
Well said.
Unfortunately I can't press the like button more than once.
Not that I haven't tried to do so...But really, this complete silence / lack of updates on the Mac front in addition to the Performa syndrome, is a little bit worrisome.
Last edited:
Right now the mac pro does not have the pci-e lanes to do TB3 unless they.
Add 1 more cpu
or
Cut the video cards down to x8 x8 or x16 x16 feed with an x16 to dual x16 pci-e 3.0 switch.
going to DMI 3 will not get them to where they need to be.
Well said.
Unfortunately I can't press the like button more than once.
But really, this complete silence / lack of updates on the Mac front in addition to the Performa syndrome, is a little bit worrisome.
And yet, we see articles like this Worldwide Mac Sales Hold Steady as PC Market Sees Shipments Decline 9.6% in Q1 2016. Not saying we should be celebrating the current lineup, but Apple probably feels the data supports their choices.
Guess, we have no updated MacBook Pro, no Macbook Air, no Mac Pro, no Thunderbolt Display, no Airport Express router neither timecapsule, no updated Mac mini, no Mac server specific hardware, no shi7.
BUT WE HAVE APPLE MUSIC , it cost just 1.2 Billions and Apple earns about 2-3 millions at month on ever less subscribers, Ahh and those over hyped and actually out earphones...
With 1/100 that money all and every one of the cited outdated system should have been updated to the best available, and Apple Mac sales will be steady solid, capitalizing much more that it lost on Apple Music.
Jobs would have been fired cook just 3 months before Apple music flop.
BUT WE HAVE APPLE MUSIC , it cost just 1.2 Billions and Apple earns about 2-3 millions at month on ever less subscribers, Ahh and those over hyped and actually out earphones...
With 1/100 that money all and every one of the cited outdated system should have been updated to the best available, and Apple Mac sales will be steady solid, capitalizing much more that it lost on Apple Music.
Jobs would have been fired cook just 3 months before Apple music flop.
I hear you. Maybe it's time people to go back windows till they resolve their identity crisis.
[doublepost=1460580052][/doublepost]
And this is why Cook needs to go back to COO.
- Status