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
LOL
So a gaming, Geforce Titan card with castrated FP64 performance is a bad thing. I suppose Nvidia's state on this forum is that you cannot say anything, even factual to not be recognized as negative.
That was only your choice to understand it as negative. Not what I have written.
You probably won't see two GP100 GPUs in a tube - it would stress the power supply.
Although there have been no announcements, you can expect that over the course of the summer and fall that Nvidia put out more consumer-oriented Pascal chips (cheaper/lower wattage). Those would have potential for a home in a tube.
[doublepost=1459960356][/doublepost]
It's really hard to interpret the adjective "castrated" as anything but negative.
Aiden first of all, I do not think Apple would want compute only monster, even if it has enormous amount of power. They would want that GPU to work in any environment, gaming, compute, etc. General purpose.
Secondly If you would read properly how I speak about this GPU you would see that I am pretty astonished. I genuinely am. What is more, after analyzing this GPU a bit I think putting out the FP64(to not use "castrate" for you), and lowering the core clock to lets say... 1050 MHz, we end up with still powerful FP32 GPU, that can be put into 125W of power thermal envelope! Most of that 300W TDP comes from EXTREMELY high clocks(1480 MHz boost!). It will be efficient, just not as efficient as Maxwell was, due to the differences in architecture, that requires a bit more of oomph despite the smaller node and HBM2 memory. I just do not see any faults in this architecture, apart from lack of hardware scheduling and it is still pretty narrow architecture. 3840 Cores. GCN like, to add, at this level. So SIMD vs SIMD performance will be pretty similar. And it looks like AMD will have wider architecture still, that might be more power efficient.
P.S. One more pun into Nvidia. Paper launch - accomplished
.Not really. Nobody uses the Mac Pro for gaming as it's primary usage.
You are probably right, but its not impossible they could use a downclocked GP100, a la the AMD R9 Nano, and put two of them in the tube. I shudder to think about the cost though...
When has Apple showed they cared at all about gaming on the mac? A compute monster is exactly what they want.
Of course any Pascal on the nMP couldn't have more that 2/3 this power due ps/TDP in the trashcan, but a couple scaled down Pascal GPUs plus the announced unified memory model for osx would enable a lot of things really pros on the nMP, only if Apple ditches AMD again at least for the nMP trashcan.
It's intriguing or a sort of cue Cuda 8 supporting unified memory with osx this requires some degree of cooperation with Apple since isn't something that would enable a simple Kext it requires deep integration and kernel level access, so I see some active involvement from Apple on Cuda 8 (available curiously at moreless same time as the WWDC), would be really great to see pascal (and nVidia) coming back to Apple (the coming back specialist)
Edit: some typos
It's intriguing or a sort of cue Cuda 8 supporting unified memory with osx this requires some degree of cooperation with Apple since isn't something that would enable a simple Kext it requires deep integration and kernel level access, so I see some active involvement from Apple on Cuda 8 (available curiously at moreless same time as the WWDC), would be really great to see pascal (and nVidia) coming back to Apple (the coming back specialist)
Edit: some typos
Last edited:
Why do you take example as a factor what Apple wants?
.It was example of GPU that can be used in much more cases not only for compute.
A trash can downgraded pascal gpu ( 2x110w) could theoretically deliver about 3000 TFlop fp64 or DP as you like, this is an big jump further about 12 gb total HBM2 memory Unified with upto 256 GB and upto 18 cores Xeon cpu will represent a solid and surely well accepted upgrade (plus NVMe and tb3)
Note that none of the Pascal GPUs shown so far have any display outputs. No Display Port, no HDMI, no....
Possibly the logic is on the GP100 die, possibly not.
It's a lot of guessing until Nvidia shows a card with DP outputs.
ps: I think that you meant 3 TFLOP or 3000 GFLOP of FP64...
Display output not necessarily must be part of the gpu it maybe added on an external chipset or on a gpu revision, BTW it requires an unannounced pascal gpu "derated or downgraded" to 1/3 if developed logically these details should be properly addressed.
This is all great but let's just wait for the desktop parts, and see if this was just not a carrot to get people talking.
And please stop calling SMM to what is now only SM
SMM was with Maxwell only.
And please stop calling SMM to what is now only SM
SMM was with Maxwell only.
Often with integrated circuits external outputs (e.g. graphics or networking) are buffered with "driver circuits" to drive the (relatively) high voltage and high current external connections.
You also don't some static electricity on the Cat7 cable to discharge directly into the heart of the PCH. The driver circuits add protection.
Even if the display logic is on the Pascal GPU, there will be external driver chips.
In any event, I wouldn't be surprised if no GP100 chip is ever connected to displays.
In a Mezzanine form factor it is highly unlikely that it will be directly connected to any display. GP100 that will come Q1 2017 will have "normal" form factor that we all know and love, most likely as a Quadro GPU. GP102 is GP100 but without FP64 and also in normal form factor with Titan branding as a consumer product. And the last two most likely will see a display .
.You don't think the term "castrated" is a negative??? How would you feel if the procedure was done to you?
I doubt nVidia not selling a gpu/graphics workstation quadro on pascal this is another big business, also now VR is something gaining hype those things have market.
What I still taking with a bit of Salt is the rumours on nVidia comeback to the nMP, it should be ideal for everyone also Apple by I don't know the extent of Apple exclusive supplier deal with AMD further those pascal gpu will not come at discount prices
Frankly I was more interested by all the improvement to existing and new API/solution that NVidia presented at that conference. The new iRay, the material editor, the unified memory for CUDA enabled apps... Those matter more to me. But I expect my company to get their hands on a few of those card for our research lab.
I don't think Pascal is a requirement for unified memory. Metal already has unified memory on Nvidia GPUs (and AMD and Intel as well.)
I'm not suggesting that. The mysterious GP102 would be the top graphics chip for VR and gaming. It will have FP64, but possibly not at 1:2.
I'm suggesting that GP100 may only appear in Tesla-branded compute cards. (Most Teslas, and perhaps all current ones, do not have display hardware.)
Note that Maxwell is mediocre at FP64 - but great at gaming, VR and ML/DL. A Pascal with less robust FP64 (1:3 or 1:4) would also be great at gaming, VR and ML/DL.
Last edited:
- Status