NVIDIA Fermi: Architecture discussion

[Jawed]

That would be odd, whats the point of a firestream card if it can't do DP? The whole thing doesn't make much sense to me, the power/perf and perf/cost of 57XX compared to 58XX doesn't really make it a much, if any better option. In fact, in power/perf is lower to use 57XX.

Stream computing is about more than double-precision. Also, R700 doesn't support OpenCL Local Memory on die, so there's a serious performance cost for algorithms based upon it. So, erm, take your pick.

Trinibwoy, suggesting that GPGPU DP is no use is pretty blinkered. The Milky Way @ Home folks like it an awful lot.

Response from AMD here:

http://milkyway.cs.rpi.edu/milkyway/forum_thread.php?id=1505&nowrap=true#35863

Double precision floating point operations are not required for any known or anticipated GPU compute applications for consumers (i.e. gaming, multimedia, and productivity), where single precision processing is sufficient. Implementing high performance DP FP processing in hardware carries significant costs in terms of die size and power consumption, and wouldn’t provide any value to the vast majority of our customers. This is the main reason why all of the companies who participated in the definition of the new industry standard OpenCL and DirectCompute APIs agreed to make this an optional feature.

That being said, there are some high performance computing applications in the scientific or industrial fields where double precision can be beneficial, so we chose to include that capability in our most powerful GPUs to provide an attractive solution for those markets as well.
- Antal Tungler

Though you might argue that leaving the computing to a few keen individuals:

http://milkyway.cs.rpi.edu/milkyway/forum_thread.php?id=1508&nowrap=true#35947

The Pro version is the oil cooled one, which will end up with the four 5970 GPU's oil cooled and a Fermi GPU also oil cooled. According to my supplier the Fermis are buggers for heat. Even the single entry level GPU will run 200 watts.

I am about to install the other 5970 now, then on Tuesday we will do the conversion to oil cooling as the blocks for that have also, finally, arrived.

BTW: When I was cooling one of the the twin 295's I had in the box, with air, they ran in the high 80's to max at 91 degrees. When I converted just one card the temp dropped to a steady 38 degrees on that card and its neighbour dropped to low 80's. An eye popping difference I thought.

takes the strain out of it

Jawed

 

[trinibwoy]

Trinibwoy, suggesting that GPGPU DP is no use is pretty blinkered. The Milky Way @ Home folks like it an awful lot.

Sorry, what does MilkWay have to do with lower end Quadros and Firestreams lacking DP? Do companies buy those cards to run that particular application?

 

[Ninjaprime]

Stream computing is about more than double-precision. Also, R700 doesn't support OpenCL Local Memory on die, so there's a serious performance cost for algorithms based upon it. So, erm, take your pick.

Jawed

I just mean, why would you need a firestream card at all if you weren't using it for some scientific purpose that required DP. A consumer 58XX card would work just fine. The only scenario I can think of is if firestream cards come with more memory and you needed more memory for a SP purpose...

 

[GZ007]

The question is how much Tesla customers buyed the card for DP ? 933SP Gflops versus 78DP Gflops is quite a difference. Multi CPU configurations could reach more than 78DP Gflops with much more flexibility.
Also want to note that there is a great dead space betwen SP and DP. For SP the maximum representable value is (2−2−23) × 2127 ≈ 3.4 × 10^38 and for DP the maximum representable value is ≈ 1.79769 × 10^308 (both from wikipedia).
SP calculations can be still reasonably accurate for plenty of scientific work. Science = DP is quite a myth.

 

[GZ007]

But than again if your DP is like 1/4 or 1/2 of SP rate than there is no a single reason to use SP when u can wait a litle longer and get 10 times higher accuracy.

 

[Jawed]

What does RV770 have to do with lower end Quadros and FireStreams lacking DP?

Considering usable DP support didn't exist at all on professional cards a few years ago (and it could be argued that RV770 and GT200 implementations weren't useful either)

If it's useful in MW@H (which is a science application), why would it not be useful in the form of a professional card?

Just curious what the basis of that argument is, since it seems baseless to me.

When would you deem GPGPU DP to become useful in Quadros (perhaps you really mean Teslas) and FireStreams?

Jawed

 

[Jawed]

I just mean, why would you need a firestream card at all if you weren't using it for some scientific purpose that required DP. A consumer 58XX card would work just fine. The only scenario I can think of is if firestream cards come with more memory and you needed more memory for a SP purpose...

Memory + that Stella Tortoise "reassuringly expensive" warm glow inside

Supposedly you get more support if you buy a Tesla.

Jawed

 

[trinibwoy]

What does RV770 have to do with lower end Quadros and FireStreams lacking DP?


If it's useful in MW@H (which is a science application), why would it not be useful in the form of a professional card?

Just curious what the basis of that argument is, since it seems baseless to me.

When would you deem GPGPU DP to become useful in Quadros (perhaps you really mean Teslas) and FireStreams?

Jawed

No, Im referring to Quadros and Firestream cards as that's what Ninjaprime asked about. Those products have never had DP so why would a lack of DP in the lower end parts be a problem? Do the professional graphics applications and APIs they run even have support for GPU DP?

 

[Jawed]

No, Im referring to Quadros and Firestream cards as that's what Ninjaprime asked about. Those products have never had DP so why would a lack of DP in the lower end parts be a problem? Do the professional graphics applications and APIs they run even have support for GPU DP?

Do you think there might be a hint in "Firestream" (hint: that isn't FirePro)

http://ati.amd.com/products/streamprocessor/specs.html

that Ninjaprime used the wrong product name for NVidia? Or, even, in the topic of discussion which is compute cards not all having DP?

Jawed

 

[pcchen]

No, Im referring to Quadros and Firestream cards as that's what Ninjaprime asked about. Those products have never had DP so why would a lack of DP in the lower end parts be a problem? Do the professional graphics applications and APIs they run even have support for GPU DP?

Firestream is not a high end professional graphics card, it's a GPGPU card, like Tesla. So having DP is probably more important for them. On the other hand, for some signal processing works, DP is not that important, and those works are pretty well suited for GPGPU. So it may make sense to have some lower end products which don't support DP.

 

[trinibwoy]

Do you think there might be a hint in "Firestream" (hint: that isn't FirePro)

http://ati.amd.com/products/streamprocessor/specs.html

that Ninjaprime used the wrong product name for NVidia? Or, even, in the topic of discussion which is compute cards not all having DP?

Jawed

Ah, I had no idea AMD even had a separate line of products for compute. Original point still stands though, if companies need DP they'll pay for it. The MilkyWay folks should be happy with 5850's @ $300. Where's the market for cheap DP capable professional compute cards?

 

[GZ007]

Chasing DP in GPUs is not very efective. I would rather say that both nvidia/ati could make pure compute chips with full DP and without the GPU pipeline bloat.
For the same transistor budget they could be monster DP chips. And the costs would come back very fast with profesional hardware margins.

 

[green.pixel]

Oh dear, well I certainly hope developers target something less than a 5760x1200 resolution for their games because all we'll end up with is a lot of ugly. Do you really look forward to running the best that the Xbox and PS3 can do at the end of their lifetimes blown up in all its unrefined glory?

What do you exaclty mean by that? That when you up the res, you can see more "cutting corners"?

 

[Jawed]

Ah, I had no idea AMD even had a separate line of products for compute. Original point still stands though, if companies need DP they'll pay for it. The MilkyWay folks should be happy with 5850's @ $300. Where's the market for cheap DP capable professional compute cards?

Do you think these guys with their "Milky Way One", 5th fastest supercomputer:

http://forum.beyond3d.com/showthread.php?t=55519

care about cheap double precision GPGPU?

Jawed

 

[Jawed]

Chasing DP in GPUs is not very efective. I would rather say that both nvidia/ati could make pure compute chips with full DP and without the GPU pipeline bloat.
For the same transistor budget they could be monster DP chips. And the costs would come back very fast with profesional hardware margins.

These chips have to be GPUs first, because consumers subsidise the development of the GPGPU stuff - otherwise you end up with Larrabee, which even with the billions and process technology that Intel has, is deemed not fit for the market for the time being.

Jawed

 

[Groo The Wanderer]

Keep an eye on section 4.0:

http://developer.amd.com/gpu/ATIStreamSDK/assets/ATI_Stream_SDK_Getting_Started_Guide_v2.0.pdf

when the Evergreen FireStream cards arrive. I don't expect Juniper based FireStream to support double precision.

Given the gestation of Juniper and Cypress, and the work involved with removing the units, I don't think that will be the case. We will see.

-Charlie

 

[GZ007]

These chips have to be GPUs first, because consumers subsidise the development of the GPGPU stuff - otherwise you end up with Larrabee, which even with the billions and process technology that Intel has, is deemed not fit for the market for the time being.

Jawed

I didnt say that it wouldnt have a gpu counterpart on which the GPGPU architecture would reside. I was thinking about real profesional supercomputers where DP is a must. But anyway amd-s fusion would kill this thing anyway. A opteron+on die gpu with cpu/gpu OpenCL code would be much more efective and work on all kind of tasks (nvidias mighty fermi super computer still needs cpu-s and lots of them, but thats usualy not in the news). AMD could even push more DP on the opteron cpu/gpu than the mainstream cpu/gpu.

 

[trinibwoy]

Do you think these guys with their "Milky Way One", 5th fastest supercomputer:

http://forum.beyond3d.com/showthread.php?t=55519

care about cheap double precision GPGPU?

Those guys certainly care but a market they do not make. The MilkyWay project alone isn't going to subsidize the added cost of including DP in mainstream chips where the vast majority of consumers won't perceive any added value. The "waste" was bad enough with GT200, it'll only get progressively worse with more price sensitive parts.

 

[trinibwoy]

What do you exaclty mean by that? That when you up the res, you can see more "cutting corners"?

No I mean that those bigger resolutions become less practical from a performance perspective as each pixel gets more expensive with more advanced rendering techniques. So while you can run Quake 4 at that resolution, you can't run Stalker:COP. Although, if most people are like me and are still playing a backlog of 3 year old games that might not be much of a problem.

 

[FrameBuffer]

No I mean that those bigger resolutions become less practical from a performance perspective as each pixel gets more expensive with more advanced rendering techniques. So while you can run Quake 4 at that resolution, you can't run Stalker:COP. Although, if most people are like me and are still playing a backlog of 3 year old games that might not be much of a problem.

Though it is highly subjective, it would also mean that as resolutions increase dramatically (5760x1200) and pixels (pitch) remains the same, the need for higher levels of AA decreases.