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
Yeah, 'm guessing the GPU parts will come with surprises. This was just show off for now, to keep the interest...
Mmm. I mean if we go with it has to significantly damage Apple, as opposed to simply damage the product line, I guess nothing could qualify. They could produce an iPhone that sells one unit and they'd still survive a decade.
What's weird is it looks like HBM was pointless given the memory limit. You got a card that was overkill for most low resolutions and yet couldn't quite handle huge resolutions due to being a 4GB card. We're going to get GDDR and HBM2 cards soon enough that are better options.
Actually Itanium still produced and sold, but only for certain HP servers I mean. The tech still not dead, it's comeback may occur but for servers (as IBM's PowerPC now Power8).Mago, I know. Itanium was my hope for the future until it wasn't. I don't expect it to come back though, no one seems to care about it anymore, unfortunately.
I'd love to see Itanium based Xeons, on the latest node, they would rock for sure.
Too bad it didn't survive other interests, it was indeed a much better, future proof, arch.
Maybe it will come back, but MS lost interest, Apple won't port OS X for sure, not just for just for a possible nMP based on it. I guess that would make us both happy
And now comes Aiden saying something smart about it...
I think it will comeback as soon Intel realized the x86 architecture can't be patched anymore.
I understood that Jen-Hsun said that GP100 is in production now, has been sampled to close partners (like Baidu), will be available soon to early adopters through the GDX-1 and to hyperscale customers, and available in off-the-shelf systems in Q1 2017.
Rather reasonable way to handle the ramp up of supply to demand.
Also, nothing was said about other variants (other than the one in the Drive PX2). I'd fully expect to see other GPxxx cards coming out over the next few months with consumer prices and power needs (and GDDR5 or a variant).
And to connect to the original topic, a GP2xx with GDDR5 in an MXM card could be a possibility for the nnMP.
Last edited:
Same history as with intel's new Xeon Phi Knights Landing, HPC market is too hungry for these parts and given the money involved they have priority 1AA+ over general consumers.
So anyone have any complaints about the MP architecture? For myself, while I understand the increasing demand for HPC like services increasing the GPU would help, most of my actual workload benefits from shared memory low cache approaches (pointer chasing). AI and large graph algorithms and the like (subgraph isomorphism, neural simulation with large numbers of synapses per cell, recurrent neural nets, etc.). MP is definitely an improvement over running on a MBP, and I can't afford a Cray XMP... (do they even make them anymore?)
Cray (different company, but they bought the name) sells Intel Xeon systems today (their front page touts E5-26xx v4 availability).
From your description, the MP6,1 would probably OK if the current 64 GiB RAM support is enough. (Actually, 128 GiB is known to work with the 12-core, but Apple doesn't sell or support the 32 GiB RDIMMs that are needed.) If Apple moves to E5-26xx v4 CPUs, the four DIMM slots would support 256 GiB.
The nMP actually is an excellent development platform for HPC, not the same as an good platform for running HPC applications.
Cray shouldn't exist, they are just Intel's front end for HPC business, while they don't develope supercomputers (as former cray did), it's like I to build and startup and name it "Turing Systems" just to resell supermicro and mellanox under the umbrella of an iconic name.
Unified Memory Support on Mac OS X
In addition to Pascal support in CUDA 8, CUDA 8 platform support for Unified Memory expands to Mac OS X. Now developers using Macs with NVIDIA GPUs can take advantage of the benefits and convenience of Unified Memory in their applications.
Source: https://devblogs.nvidia.com/parallelforall/cuda-8-features-revealed/
I'm not sure if that means some kind of commitment to deliver Pascal based hardware for Macs.
In addition to Pascal support in CUDA 8, CUDA 8 platform support for Unified Memory expands to Mac OS X. Now developers using Macs with NVIDIA GPUs can take advantage of the benefits and convenience of Unified Memory in their applications.
Source: https://devblogs.nvidia.com/parallelforall/cuda-8-features-revealed/
I'm not sure if that means some kind of commitment to deliver Pascal based hardware for Macs.
Mago, Itanium still exists indeed but way outdated, still on 32nm process node, and there have been no recent developments. Only HP uses them still in some servers. I don't see Intel putting more money into it's further development, but I hope I'm wrong.
DP is 2:1 in Pascal, that's great. Finally.
I wonder what the ratio will be in Polaris and Vega. Let's hope it's 2:1 too.
I believe FP16 will also come in at double the rate.
The architectures of both camps are coming closer, it will be interesting to see how HS will work on Pascal.
I wonder what the ratio will be in Polaris and Vega. Let's hope it's 2:1 too.
I believe FP16 will also come in at double the rate.
The architectures of both camps are coming closer, it will be interesting to see how HS will work on Pascal.
Reading Cuda8 and Pascal it's unified memory model put these GPGPU at par to Xeon Phi Knights Landing about the freedom your algorithm can access memory, it's very important for Graph analysis and finance, also as they target for deep learning.
I need to remember you that's AMD FX line still on 28nm until they launch Zen FX
Itanium obviously is not a priority for Intel but they still keep it staff.
PD I'm tired on all these nomenclature "Polaris pascal, Zen, Phi c'mon...
Unfortunately I think it s highly unlikely
. So far what has been seen P10 has 1/16 ratio of DP.What is interesting about Pascal architecture is that it is much more GCN-like than any Nvidia architecture before. It will be very optimized for any console port in gaming. What is more interesting is the fact that GP102 looks like in architecture it will have 128 ROPs. It will also have very high core clock. This thing will be fast in gaming. Also it has pretty unusual approach to Asynchronous Compute because each SMM can execute it differently depending on the tasks. It will adapt itself very well.
Now we have to wait for red team answer.
Polaris as far I know isn't targeted at compute but I think it may include something to trick fp64 thru fp32 and get a descent 4:1 execution ratio (as nVidia did via driver on maxwell), 1:16 corresponds to fiji.Unfortunately I think it s highly unlikely
What is interesting about Pascal architecture is that it is much more GCN-like than any Nvidia architecture before. It will be very optimized for any console port in gaming. What is more interesting is the fact that GP102 looks like in architecture it will have 128 ROPs. It will also have very high core clock. This thing will be fast in gaming. Also it has pretty unusual approach to Asynchronous Compute because each SMM can execute it differently depending on the tasks. It will adapt itself very well.
Now we have to wait for red team answer.
BTW nVidia aimed directly to Xeon Phi when redesigned it's compute core, while we need to wait for real testing on both platform on the paper (or on power point to be precise) nVidia seems ahead Xeon Phi w/o none it's previous restrictions that prevented some HPC deployment due algorithm incompatibility which favored Xeon Phi and Mic ( nVidia pascal actually looks more as a MIC solution than SMM).
Although a MIC based on ARM hardly will be competitive against Pascal (at least on the paper).
Unfortunately I think it s highly unlikely
What is interesting about Pascal architecture is that it is much more GCN-like than any Nvidia architecture before. It will be very optimized for any console port in gaming. What is more interesting is the fact that GP102 looks like in architecture it will have 128 ROPs. It will also have very high core clock. This thing will be fast in gaming. Also it has pretty unusual approach to Asynchronous Compute because each SMM can execute it differently depending on the tasks. It will adapt itself very well.
Now we have to wait for red team answer.
While I can't find where I read it, I believe I have seen a 1/3 double precision ratio either rumored or speculated. There is no way it would be 1/16 like Fiji. AMD needs a compute focused chip since they haven't had one since Hawaii. Just like Nvidia went compute heavy with Pascal after going graphics heavy with Maxwell, AMD will mirror this by going compute heavy with Polaris after going graphics heavy with Fiji. Remember that Tahiti, the first AMD GPU on 28 nm, was also compute focused with a 1/4 ratio.
What is so surprising to me is that Nvidia is shipping a 600 mm2 GPU on 16 nm in June. Before this announcement I think people were wondering if they would make it by the end of the year. I wonder if AMD is regretting going with Global foundaries if they are going to be stuck with a ~230 mm2 GPU until Vega comes out in 2017.
I wonder if there is any chance we would see Pascal in a Mac Pro. Some of those technologies like NVLink and GPU/CPU memory sharing would be pretty slick. Those Pascal boards look very much like the GPUs already in the Mac Pro.
- Status