Cray Announces AMD Bulldozer CPU and NVIDIA Tesla GPU Supercomputer Capable of 50 Pet

[dawp]

http://www.pcper.com/news/Systems/C...A-Tesla-GPU-Supercomputer-Capable-50-Petaflop

wasn't sure where to post this so I decided to start a thread here.

 

[DirkGently1]

That's a number crunching monster, but will it be able to calculate the Question to the Answer 42?

 

[sm625]

The first customer is Swiss National Supercomputing Centre. I was sure it was going to be Goldman Sachs.Why didnt they use AMD gpus? There can only be one reason....

 

[PlasmaBomb]

300 cabinets at up to 50 kW per cabinet is a lot of juice...

From the article -

With a claimed capability to utilize up to 300 cabinets full of XK6 blades, customers are looking at approximately 44 petaflops of computing horsepower, with GPUs delivering 19.14 petaflops, and the CPUs estimated to provide 25.2 petaflops of floating point computational power.

How is ~44 PFlops = 50 PFlops?

 

[Tuna-Fish]

The first customer is Swiss National Supercomputing Centre. I was sure it was going to be Goldman Sachs.Why didnt they use AMD gpus? There can only be one reason....

At the moment, the nVidia "accelerators" are better for double floating point loads, while AMD cards are better for integer (codebreaking).

 

[Topweasel]

300 cabinets at up to 50 kW per cabinet is a lot of juice...

From the article -


How is ~44 PFlops = 50 PFlops?

Capable might mean drop in upgrades?

 

[nonameo]

Capable might mean drop in upgrades?

it means you can overclock it, yo

 

[Phynaz]

300 cabinets at up to 50 kW per cabinet is a lot of juice...

Plus you need to cool all that - more juice.

Cooling a 50kw rack is very difficult. The traditional hot isle / cold isle doesn't work anymore in the 25kw - 30kw range. At 50kw you have to get fancy, such as local liquid cooling, which in turn is cooled by your main cooling system.

 

[itsmydamnation]

Plus you need to cool all that - more juice.

Cooling a 50kw rack is very difficult. The traditional hot isle / cold isle doesn't work anymore in the 25kw - 30kw range. At 50kw you have to get fancy, such as local liquid cooling, which in turn is cooled by your main cooling system.

contained hot/cold isle will do above 25kw quite happily, 50 is getting a little on the extreme side of things.

 

[gevorg]

I guess China will soon have e-peen insecurities again.

 

[Phynaz]

contained hot/cold isle will do above 25kw quite happily, 50 is getting a little on the extreme side of things.

"Row-oriented offers the highest usable capacity across the broadest power density range. Due to the close coupling of the CRAC units to the load, all of the capacity can be delivered to the load up to power densities on the order of 25 kW"

http://www.datacentres.com/papers/papers/apcWP130.pdf

 

[skipsneeky2]

put that bad boy on the folding table and it would make any other farm we have look like a celeron email machine lol

 

[pw38]

We are the Cray, resistance is futile. I can imagine it achieving consciousness and assimilating the people that walk by.

 

[wahdangun]

wow ddr4 !! I can't wait for it,

 

[itsmydamnation]

"Row-oriented offers the highest usable capacity across the broadest power density range. Due to the close coupling of the CRAC units to the load, all of the capacity can be delivered to the load up to power densities on the order of 25 kW"

http://www.datacentres.com/papers/papers/apcWP130.pdf

http://www.apcmedia.com/salestools/DBOY-7EDLE8_R0_EN.pdf

we run 3 HP C class blade chassis to a rack, thats max of around 45kW per rack. During commisioning of our first "POD" we did several tests including:
running severial jet burners, before servers where in
having every CPU in a POD (HP C class enclosures) running at 100% in which we where averaging around 35kW of cooling per rack.

changing the configuration of Hot isle operating temprature, room temprature and ratio of racks to cooling units can increase cooling per rack far higher then what we require.

 

[Topweasel]

it means you can overclock it, yo

Is there a switch on the side and does hit this OEM provided switch void the warranty?

 

[Cerb]

The first customer is Swiss National Supercomputing Centre. I was sure it was going to be Goldman Sachs.Why didnt they use AMD gpus? There can only be one reason....

ECC? AMD hasn't tried to compete with nVidia, here. They are likely planning on not doing so, until Opterons start coming with on-die GPUs.

 

[JFAMD]

The first customer is Swiss National Supercomputing Centre. I was sure it was going to be Goldman Sachs.Why didnt they use AMD gpus? There can only be one reason....

I have been to the swiss center (CSCS) and it is impressive. They run a variety of different workloads.

When they redid the blogs the picture of me in front of the Monte Rosa cluster was dropped, don't know why:

http://blogs.amd.com/work/2009/09/23/amd-opteron-processors-scale-the-alps/

As for the GPUs, nvidia has ECC, we will have it soon, I am guessing that is the reason.

Capable might mean drop in upgrades?

Yes, Cray will be doing drop in upgrades.

ECC? AMD hasn't tried to compete with nVidia, here. They are likely planning on not doing so, until Opterons start coming with on-die GPUs.

Opterons will not have on-die GPU, at least not in the near future.

 

[Idontcare]

Opterons will not have on-die GPU, at least not in the near future.

John this comment catches me by surprise.

I thought AMD was, has been, pushing for a future-is-fusion vision in which their CPU's will incorporate APU's?

Your statement implies that APU's and the vision behind their creation are not applicable in the workstation/server/HPC environments which is confusing to me as Nvidia has shown that these are the very environments where APU's have significant utility.

 

[Cerb]

As for the GPUs, nvidia has ECC, we will have it soon, I am guessing that is the reason.

That I have not read, before.

Will it be done using additional interface width?

 

[AnandThenMan]

Your statement implies that APU's and the vision behind their creation are not applicable in the workstation/server/HPC environments which is confusing to me as Nvidia has shown that these are the very environments where APU's have significant utility.

I would guess that the rationale behind this is simple, the performance from the on-die graphics would not be significant enough to justify the added expense and thermals. A big part of this is memory bandwidth, a workstation card has bandwidth and memory capacity packed to the rafters.

 

[bryanW1995]

That's a number crunching monster, but will it be able to calculate the Question to the Answer 42?

Haha, that's my favorite answer. I told my wife to remember that a few years ago when she was going into day-surgery. She didn't realize she had been put out b/c she was so focused on "42", so when she woke up she just continued her sentence to the Dr, telling him that I had told her to remember that number. Don't know if said Dr was a douglas adams fan, unfortunately...

How is ~44 PFlops = 50 PFlops?

I think it's the same fuzzy math used to show that my 80gb hd really only has 74gb of useable space.

 

[Topweasel]

Haha, that's my favorite answer. I told my wife to remember that a few years ago when she was going into day-surgery. She didn't realize she had been put out b/c she was so focused on "42", so when she woke up she just continued her sentence to the Dr, telling him that I had told her to remember that number. Don't know if said Dr was a douglas adams fan, unfortunately...



I think it's the same fuzzy math used to show that my 80gb hd really only has 74gb of useable space.

Its not Fuzzy math. To HDD manufacturers, 1000MB is 1GB, to Software guys its 1024MB's. Since its software guys doing the reporting they will always report less space then the Manufacturers say they have.

 

[Dribble]

John this comment catches me by surprise.

I thought AMD was, has been, pushing for a future-is-fusion vision in which their CPU's will incorporate APU's?

Your statement implies that APU's and the vision behind their creation are not applicable in the workstation/server/HPC environments which is confusing to me as Nvidia has shown that these are the very environments where APU's have significant utility.

Perhaps that's because the APU part is just marketing and in fact all it is right now is a low end gpu stuck on chip. If AMD were serious about gpu compute in super computers they'd be doing their own project denver by combining an opteron with some gpu compute version of a radeon.

As for the GPUs, nvidia has ECC, we will have it soon, I am guessing that is the reason.

Well it's a reason. There's all the others like gpu architecture and software support. In particular providing a basic opencl driver won't cut it when the opposition has comprehensive sdk's, debuggers, C++ support, and all the rest of it. AMD previously has relied on others to produce the software - e.g. MS, but MS isn't really interested in DX compute right now - they are fully focused on windows 8 + arm + tablets. Hence this is a pretty huge stumbling block.

Imo AMD are hardly any more focused on compute then MS right now - they have basically let nvidia have the market, choosing to believe that they can just step in any time - all you need is an opencl driver and EEC support. Clearly there's a whole lot more to it then that, and nvidia have spent a lot of money developing it and are very unlikely to share.

 

[AnandThenMan]

Perhaps that's because the APU part is just marketing and in fact all it is right now is a low end gpu stuck on chip. If AMD were serious about gpu compute in super computers they'd be doing their own project denver by combining an opteron with some gpu compute version of a radeon.

You think the power envelope and packaging would be feasible for such a chip?

BTW, AMD would have to be pretty much out of their mind to stuff a GPU into Bulldozer at this stage, it's an all new architecture on a new process. Unless you would not mind it being delayed for another year.