AMD Announces First "Supercomputing" Server Graphics Card With 12 GB Memory

[Recus]

Looks like someone doesn't understand what's going on.

The Green500 list, ranks the top 500 supercomputers in the world by energy efficiency. The focus of performance-at-any-cost computer operations has led to the emergence of supercomputers that consume vast amounts of electrical power and produce so much heat that large cooling facilities must be constructed to ensure proper performance. To address this trend, the Green500 list puts a premium on energy-efficient performance for sustainable supercomputing.

While similar performance Beacon and BlueGene/Q using 45 kW and 82 kW, Sanan using 179 kW.

Who even thinking about W when parents paying bills?

 

[librin.so.1]

Recus, dude, the absolute energy consumption level is irrelevant. The only important part is [performance / Watt], in other words - efficiency. And this supercomputer has the 4th best efficiency in the world. Your argument is invalid.

Also, You do show how You can't even do simple math - the Beacon is 110 TFLOPS and the BlueGene/Q is 189 TFLOPS. While the SANAM is 421 TFLOPS. In other words - it outperfroms those two supercomputers combined. Your argument is invalid x2.

 

[NeoXF]

Learn to to understand what Green500 stands for.

Fanboy detected.

Dude, seriously... stop embarrassing yourself, just stop posting.

 

[SIGSEGV]

Looks like someone doesn't understand what's going on.



While similar performance Beacon and BlueGene/Q using 45 kW and 82 kW, Sanan using 179 kW.

Who even thinking about W when parents paying bills?

 

[Recus]

Recus, dude, the absolute energy consumption level is irrelevant. The only important part is [performance / Watt], in other words - efficiency. And this supercomputer has the 4th best efficiency in the world. Your argument is invalid.

Also, You do show how You can't even do simple math - the Beacon is 110 TFLOPS and the BlueGene/Q is 189 TFLOPS. While the SANAM is 421 TFLOPS. In other words - it outperfroms those two supercomputers combined. Your argument is invalid x2.

As I said: "Who even thinking about W when parents paying bills?"

Btw, how did you got 421? Sanam - 532 TFlop/s Rmax, 1098 TFlop/s Rpeak.

Dude, seriously... stop embarrassing yourself, just stop posting.

Go to look at Mantle powerpoint...

 

[Brusfantomet]

At any rate this card already has the tech for fully unified memory with either gpu able to see write and move data anywhere and across busses it's likely its just us consumers waiting to use thesr features , pro grade driver's are a different beast.

since SLI and CF is working on the same data sets (models, textures ++) there will never be a way to get more than one card of RAM without a massive performance impact. The only way would bee if there is invented a new bus with the same low latency AND high banwith as local RAM can give.

The ability to use system RAM does lessen the need for massive frame buffers, given that the prefetching algorithms work.

 

[Nihilus]

Glad to see GPUs are doing amazing things in super computer efficiency (as recus showed us).

 

[TheoneandonlyMrK]

since SLI and CF is working on the same data sets (models, textures ++) there will never be a way to get more than one card of RAM without a massive performance impact. The only way would bee if there is invented a new bus with the same low latency AND high banwith as local RAM can give.

The ability to use system RAM does lessen the need for massive frame buffers, given that the prefetching algorithms work.

these are likely not attached or used in crossfire, why would they be whilst supercomputing and the whole point of that in supercomputing is so that they can work on different bits of the same data set(massive) and update each others memoryIf neccessary, ie the performance drop would be nill if the application efficiently packets the data into workable blocks., im no programmer though

 

[Brusfantomet]

these are likely not attached or used in crossfire, why would they be whilst supercomputing and the whole point of that in supercomputing is so that they can work on different bits of the same data set(massive) and update each others memoryIf neccessary, ie the performance drop would be nill if the application efficiently packets the data into workable blocks., im no programmer though

the answer was directed at consumer cards in CF

 

[TheoneandonlyMrK]

the answer was directed at consumer cards in CF

This isn't one of those, check it's price and its made for a different purpose where cross fire is pointless and driver support and features are not variable they are validated to a spec an ieee spec. They pay lots.

 

[Solaris17]

this card......could play crysis

 

[Brusfantomet]

This isn't one of those, check it's price and its made for a different purpose where cross fire is pointless and driver support and features are not variable they are validated to a spec an ieee spec. They pay lots.

i know, but it seamed theoneandonlymrk was waiting for the possibility to not have the same data in both the cards RAM in CF, like in compute. I merely pointed out that it probably never will happen.

 

[qubit]

this card......could play crysis

Nah, it would be short on memory.

 

[TheoneandonlyMrK]

i know, but it seamed theoneandonlymrk was waiting for the possibility to not have the same data in both the cards RAM in CF, like in compute. I merely pointed out that it probably never will happen.

And I pointed out your wrong

THIS IS a server gfx card it won't be used like a consumer xfire on a stick and would as I said use its whole memory on.

Several different remote virtual servers eg

Or two points of the same or different large data sets stop bringing consumer points up in this thread as its pointless.

No server administrator is buying this without a specific use already planned(xfire wont be one used often) and that purpose is unlikely to be to play crisis sorry.

 

[cdawall]

I'm really confused what happened. The 4th most efficient setup on the list uses s1000000000000000000 cards it just so happens to have more hence the highest power consumption...

 

[Nihilus]

I'm really confused what happened. The 4th most efficient setup on the list uses s1000000000000000000 cards it just so happens to have more hence the highest power consumption...

Most of us get it, one person still does not.

Adding to that, most things lose a little efficiency as you scale up. ALL THINGS EQUAL, an 8 core cpu will never have as good of a performance/watt as a 4 core; an 5.0 liter v-8 will not have the hp/liter as 2.5 4cyl and so on. Being that this supercomputer has very high performance and still at the top of the list speaks volumes.

 

[HumanSmoke]

I'm really confused what happened. The 4th most efficient setup on the list uses s1000000000000000000 cards it just so happens to have more hence the highest power consumption...

Pretty much, although efficiency largely stems from a high ratio of GPU to CPU. A large percentage of supers are either CPU only, or weighted heavily towards the CPU. SANAM has equal numbers of CPUs and GPUs. It's no coincidence that the list is topped by the most efficient CPUs allied with the most efficient GPUs -Eurora and Aurora Tigon are both 1:1, while NICS's Beacon has 2 Xeon Phi GPUs for every CPU.
It is actually a testament to IBM's Power architecture that their CPU-only BlueGene systems still feature highly in the (largely subjective) Linpack synthetic benchmark runs.

 

[Recus]

Most of us get it, one person still does not.

Adding to that, most things lose a little efficiency as you scale up. ALL THINGS EQUAL, an 8 core cpu will never have as good of a performance/watt as a 4 core; an 5.0 liter v-8 will not have the hp/liter as 2.5 4cyl and so on. Being that this supercomputer has very high performance and still at the top of the list speaks volumes.

So it can be most efficient with high power consumption but can't be fastest (780Ti) with high price?

 

[Nihilus]

So it can be most efficient with high power consumption but can't be fastest (780Ti) with high price?

I can not make sense of this at all.

 

[NeoXF]

since SLI and CF is working on the same data sets (models, textures ++) there will never be a way to get more than one card of RAM without a massive performance impact. The only way would bee if there is invented a new bus with the same low latency AND high banwith as local RAM can give.

The ability to use system RAM does lessen the need for massive frame buffers, given that the prefetching algorithms work.

So a APU supercomputer (w/ heterogeneous memory) would have a huge upper-hand in that regard, compared with a CPU/GPU split one? (think 64GB RAM per socket/APU).

 

[qubit]

I'm assuming that picture is some sort of engineering sample and not the finished product made without a cover.

Big ass heatsink though, damn.

It does have a cover and fans. I knew it had to:

Saw it here: http://benchmarkreviews.com/9182/amd-firepro-s10000-12gb-edition-graphics-card-announced/

 

[bencrutz]

It does have a cover and fans. I knew it had to:

http://img.techpowerup.org/131118/AMD-FirePro-S10000-12GB-Edition-Graphics-Card-Announced.png

Saw it here: http://benchmarkreviews.com/9182/amd-firepro-s10000-12gb-edition-graphics-card-announced/

there are passive and active cooling variant.
passive: http://www.sapphirepgs.com/productdetail.asp?IDno=74&lang=eng

active: http://www.sapphirepgs.com/productdetail.asp?IDno=45&lang=eng

*links are for 6GB though

 

[HumanSmoke]

The new Green500 list has just been published. SANAM has dropped from 5th to 11th largely due to the commissioning of TSUBAME 2.0 and Piz Daint - the latter debuted at #6 in the Top500.
Intel Ivy Bridge-E/Sandy Bridge-E + Nvidia GK 110 now fill the top ten spots.

 

[FX-GMC]

Looks like someone doesn't understand what's going on.



While similar performance Beacon and BlueGene/Q using 45 kW and 82 kW, Sanan using 179 kW.

Who even thinking about W when parents paying bills?

Not similar performance, similar efficiency. Take column 4 of your chart and convert it to watts. Then multiply that by column 2 and you have total FLOPS. Ya follow?