Your best guess on what this nMP Configuration will cost?

[slughead]

Coupled with OS X 10.9's "race to sleep" and they also may be able to do "blast, recover, repeat" cycles on a wide variety of workloads that will deliver acceptable user action/response. Non human driven workloads are the only loads that typical drive the engine throttle to 100% power and keep it there for extended periods of time.

There is also slop/safety factor in TDP specs also. If Fred, Barney , Wilma , and Betty all design different components with different coolers there is typically a small factor built in to account from the possibly conflicting designs. If Apple carefully measured and removed the margins they could punt on power/disposal that never gets generated/needed.

Doubtful all of these brings things all the way back to even, but it probably isn't quite as ridiculous as it looks. However, high enough that 6, let alone 7, TFLOPs is going to largely be a mirage on anything be the most contrived codebase+dataset.

Or not running them all at the same time. The common trait where all the CPU and GPU packages are going is that they don't run at a single speed. It is all about dynamic range. The computer should down clock when there is nothing to do.

All great points, I am highly skeptical though. The TBD on the 6/8 core E5 is 130 watts (12core is less). So now we have a TPD of 678 Watts, not including the USB or TB ports. That means Apple has to find a way to shave off 34% on sustained tasks.

What will happen if it hits the wall? My hope is that it'll just down-clock to oblivion, my fear is we're going to see nMPs blinking off.

 

[VirtualRain]

All great points, I am highly skeptical though. The TBD on the 6/8 core E5 is 130 watts (12core is less). So now we have a TPD of 678 Watts, not including the USB or TB ports. That means Apple has to find a way to shave off 34% on sustained tasks.

What will happen if it hits the wall? My hope is that it'll just down-clock to oblivion, my fear is we're going to see nMPs blinking off.

We chatted about this briefly elsewhere, but I believe they ARE down-cocking the GPUs based on the lower TFlop performance specs Apple is quoting vs comparable GPUs.

My core clock calculations are in that table (below), and it looks like they are around 15% lower than the desktop or workstation parts, so this will account for some of the power savings necessary for this to work, although I still don't see how a pair of under-clocked Tahiti GPUs and a 130W TDP CPU can safely fit within a 450W budget.

As you say, hopefully there will be some nasty thermal management that throttles things down before the smoke comes out. These systems will definitely not be suitable for crossfire gaming assuming that might work. It will be interesting to see a brave new owner put this thing to the test.

 

[slughead]

As you say, hopefully there will be some nasty thermal management that throttles things down before the smoke comes out. These systems will definitely not be suitable for crossfire gaming assuming that might work. It will be interesting to see a brave new owner put this thing to the test.

I'm not sure we'll have to worry about the thermals, it'll never have the juice to get that warm!

It'll probably just rev up until it maxes out the power rails and the PSU will shut itself off.

Another question is: If this is a sure thing, and they for sure need to down-clock in a power deficit situation, will that software work in Windows? G_d forbid someone actually gets crossfire working in these things, it'll go black in 2 seconds.

 

[MacVidCards]

I'm not sure we'll have to worry about the thermals, it'll never have the juice to get that warm!

It'll probably just ref up until it maxes out the power rails and the PSU will shut itself off.

Another question is: If this is a sure thing, and they for sure need to down-clock in a power deficit situation, will that software work in Windows? G_d forbid someone actually gets crossfire working in these things, it'll go black in 2 seconds.

It is pretty much guaranteed that the engineers involved will keep these things from going thermonuclear.

That said, there is NOT enough power or cooling to run both GPUs and the CPU.

So one or other is going to have to retreat, HEAVILY.

Here is an experiment you can try at home for $0.

Go to WIndows and temp mod your GPU at 10% down in clocks & voltage.

See how much it buys you in Watts & cooler temps.

Think about what that means for nMP.

The "Ultimate Setup" is going to be setting it on a phase change cooling loop. (On top of an AC vent) This will allow fastest speeds. Who is going to be first to market to finish nMP cooling?

 

[deconstruct60]

We chatted about this briefly elsewhere, but I believe they ARE down-cocking the GPUs based on the lower TFlop performance specs Apple is quoting vs comparable GPUs.

Those are, like the mainstream card numbers, 'up to' (peak) numbers. The compensation techniques may help them actually stay within reasonable distance of those for sustained rates. However, it does point to the core value proposition problem they have in trying to charge "Wx000 Fire Pro" prices when they are bleeding off performance. Not only are this physically segmented away from the 'normal' FirePro card market, they are performance segmented also. There is almost no overlap here at all to warrant highly matching price points.

, although I still don't see how a pair of under-clocked Tahiti GPUs and a 130W TDP CPU can safely fit within a 450W budget.

Apply exactly the same general technique only dynamically. Downclock everything till it all fits. Remember, they are quoting peak performance benchmarks, not sustained. This is an extremely old game in High Performance Computing (HPC). The marketing sheets through out peak this and peak that. Seasoned supercomputer buyers blow right past that and run benchmarks to measure sustained performance and cross section bandwidth to put an actual purchase value on a system.

As you say, hopefully there will be some nasty thermal management that throttles things down before the smoke comes out.

It not necessarily all that nasty if it is centralized. Power management feedback is sent from the CPU and the two GPU daughtercards to something that takes in the whole picture and adjust accordingly. Since, they are all sharing the same fan and heat sink, they probably should be run this way anyway. A "smart" power management controller will know it only has about a 400W budget to work with and put the breaks on before things get out of hand. ( like anti-lock brakes or traction control. )

If there is more of a kludge were fake power management with local controllers and then just do "fan control" merge on that's probably not so good.

These systems will definitely not be suitable for crossfire gaming assuming that might work.

Was never likely a high priority design objective. In fact, probably almost the opposite. Three somewhat loosely coupled high power consumption computational engines so that can juggle power and thermal dispersion between them.

 

[deconstruct60]

That said, there is NOT enough power or cooling to run both GPUs and the CPU.

There is. Just not at the "max" settings.

Go to WIndows and temp mod your GPU at 10% down in clocks & voltage.

See how much it buys you in Watts & cooler temps.

That test is extremely architecture generational dependent. On old architectures it isn't going to tell a whole lot that is relevant.

The "Ultimate Setup" is going to be setting it on a phase change cooling loop. (On top of an AC vent) This will allow fastest speeds. Who is going to be first to market to finish nMP cooling?

Even if make the central thermal core incrementally more efficient by pumping deeply chilled air into the core (without injecting significant turbulence) so there is a steeper temperature gradient, that only really lets the power available just be used marginally more efficiently. It isn't going to increase the overall budget. It will still have to be allocated in the max everything ( CPU , GPU ,GPU ) configuration.

 

[MacVidCards]

There is. Just not at the "max" settings.

Argh....if I ask nicely will you stop arguing just to argue?

What is your point here? If someone flops down $8-10K for a maxed nMP they are doing it for max hp. When they find out that it benches at 60% of what is components are capable of, how happy are they going to be? Telling them that any individual component RUN ALL BY ITSELF can achieve it's MAX speed JUST FINE for a short duration but since it is all part of a SINGLE UNIFIED SYSTEM , there is no realistic way to do it isn't an answer. It's an excuse.

So you buy the maxed out system but can only run it at a set percentage of theoretical speed due to the limiting factor, the available power. Hence, THERE ISNT ENOUGH POWER FOR EVERYTHING AT ONCE.

People shouldn't buy a computer based on what the parts could do in a vacumn, all that matters is what the total can do. If you jam a 6 liter V12 in a Mini but can only use it for short bursts due to the cooling issues from small radiator, what is the point?

 

[[G5]Hydra]

All the talk about the nMP not having the power to run twin D700(aka w9000's) forgets the most important thing i.e. Apple released the specs for only the quad and hexacore models. The reason the D500 is lower on tflops than you'd expect is they cut down what looks like a W8000's clock to lower the thermals and added back in the wider 384bit memory bus to help boost bandwidth. Peak tflops will likely suffer but sustained tflops will likely do better so its a decent tradeoff to get the power budget where it needs to be. The models with D700's will come out with bigger power supply. If they cut the power budget way back to get the D700(aka W9000) down the necessary limit of 120W per card to work with the 450w power supply it would have been impossible to get anywhere over 3 tflops. The nMP's equipped with D700's are going to be much much more expensive so the cost of throwing in a beefier ps won't even amount to a rounding error on such an expensive machine.

 

[MacVidCards]

All the talk about the nMP not having the power to run twin D700(aka w9000's) forgets the most important thing i.e. Apple released the specs for only the quad and hexacore models. The reason the D500 is lower on tflops than you'd expect is they cut down what looks like a W8000's clock to lower the thermals and added back in the wider 384bit memory bus to help boost bandwidth. Peak tflops will likely suffer but sustained tflops will likely do better so its a decent tradeoff to get the power budget where it needs to be. The models with D700's will come out with bigger power supply. If they cut the power budget way back to get the D700(aka W9000) down the necessary limit of 120W per card to work with the 450w power supply it would have been impossible to get anywhere over 3 tflops. The nMP's equipped with D700's are going to be much much more expensive so the cost of throwing in a beefier ps won't even amount to a rounding error on such an expensive machine.

There is a chance that you are right.

But this would make future upgrades of GPUs even more unlikely/impossible.

You would not only have to buy 2 @ GPUs, you would need to also replace the PSU if upgrading GPUs.

 

[[G5]Hydra]

There is a chance that you are right.

But this would make future upgrades of GPUs even more unlikely/impossible.

You would not only have to buy 2 @ GPUs, you would need to also replace the PSU if upgrading GPUs.

I think that makes it even more likely to be true. Apple basically designed a custom PCI card form factor, even to the point that one of them has a PCI flash card growing out of it, bolted directly to a massive central heat sink that's probably slathered in thermal paste. I don't think upgrading the cards by the customer is even a consideration. They can come out for service but the only thing Apple plans on you touching inside is the RAM and the PCI flash card.