AMD Introduces Industry’s Most Powerful Server Graphics Card

[tweakboy]

Good news for some AMD people...

http://www.amd.com/us/press-releases/Pages/amd-most-powerful-server-graphics-2012nov12.aspx

:whiste:

 

[ShintaiDK]

Most powerhungry?

The S10000 is 3.94Mflops per W.
Telsa K20X is 5.57Mflops per W.

Anyway, in the 225W segment thats its all about.
FirePro S9000, 900Mhz, 0.806 peak DP Tflops, 6GB@264GB/sec. (3.58Mflops/W.)
Xeon Phi 5110P, 1053Mhz, 1.01 peak DP Tflops, 8GB@320GB/sec. (4.49Mflops/W.)
Tesla K20, 706Mhz, 1.17 peak DP Tflops, 5GB@208GB/sec. (5.2Mflops/W.)

 

[OCGuy]

I was looking for something that actually backed up the headline, all I got was a press release.

 

[3DVagabond]

The S10000 is a graphics card, though. It's not meant as a single purpose compute monster. The S9000 (also a graphics card) and S10000 are a generation behind the GK110. We'll have to see what Sea Islands top single GPU can do and if it can match the K20 or not. "In theory" GCN can do twice the DP it does now (1/2 compared to 1/4).

 

[SKORPI0]

http://www.anandtech.com/show/6445/amd-announces-firepro-s10000

Launch price - $3599.

:'(

 

[boxleitnerb]

Perf/$ is quite good. K20X will be much more expensive probably.

 

[Keysplayr]

The S10000 is 3.94Mflops per W. 2xGPUs
Telsa K20X is 5.57Mflops per W. 1xGPU

 

[boxleitnerb]

I know they are not really comparable spec wise, but where you can fit those 375W and where power consumption is of little consequence, the S10000 might be appealing. $3599 vs. likely $4000-5000 for K20X which has less of everything, especially bandwidth.

I can understand why AMD would release such a card given the fact it cannot compete with S9000 against K20. Brute force approach, but good perf/$. Where have we seen that before...Bulldozer anyone?

 

[ShintaiDK]

Not to mention x86 beating it in efficiency (and speed at 225W)

The S10000 and S9000 are statements of why the HPC segment is beyond AMD.

Its nVidia vs Intel there. Both with their own benefits.

S9000, 2499$ 0.806Tflop DP. Thats 0.32Mflop per $.
K20, 3199$ 1.17Tflop DP. Thats 0.37Mflop per $
5110P, 2649$ 1.01Tflop DP. Thats 0.38Mflop per $.

Just looks worse and worse for AMD.

 

[sontin]

Yeah, AMD is dead in the server market. No OEM is supporting their FirePro cards. There are no software tools, there is no infrastructure. The cards are much slower than the competition or need to much power.

Ironic: You can't beat nVidia with a Fermi copy.

 

[FalseChristian]

ATI was so much better when they were a standalone Canadian GPU maker. I wish Intel would level the playing field buy buying nVidia.

 

[AnandThenMan]

I wish Intel would level the playing field buy buying nVidia.

Huh?

 

[OVerLoRDI]

ATI was so much better when they were a standalone Canadian GPU maker. I wish Intel would level the playing field buy buying nVidia.

You are all over the place tonight man.

 

[RussianSensation]

The S10000 is 3.94Mflops per W. 2xGPUs
Telsa K20X is 5.57Mflops per W. 1xGPU

It's not Mflops/w, but Gflops/w.

For example, K20X => 732mhz x 2688 SP x 2 Ops/clock / 1000 = 3935 Gflops
Then to get DP divide by 3 => 1311 Gflops
Then to get per Watt divide by 235W => 5.58 Gflops / Watt

Anyway your and ShintaiDK's comparison mysteriously focuses ONLY on double precision performance. Not all apps use double precision - hence the reason for K10's existence.

SP
S10000 = 15.8 Gflops/w
K10 = 20.4 Gflops/w
K20 = 15.7 Gflops/w
K20X = 16.8 Gflops/w

Outside of superior Glops/watt, other reasons NV is dominating HPC are:

1) Software integration - CUDA
2) Customer relationships - NV created GPGPU industry
3) Technical expertise / support
4) AMD is only now entering this space - can't expect Cadillac ATS to overthrow a BMW 3 series that has established itself for a long time in a particular segment.

It is true that Kepler has superior performance/watt no doubt, but it took a 7B transistor chip to barely beat a 4.3B chip in compute (although NV gimps its consumer GPUs, which means GTX680 is hopelessly outclassed in Compute in the consumer space for games/distributed computing projects):

A 4.3B HD7970 Ghz @ 1050mhz on a 365mm^2 die, with 250W TDP:
SP = 4.3 Tflops
DP = 1.08 Tflops

Asus Matrix $499 @ 1300mhz:
SP = 5.3 Tflops
DP = 1.33 Tflops

K20X > $3,200 (seen estimates of $4-5K):
SP = 3.94 Tflops
DP = 1.31 Tflops

If you just compare it on paper, for us consumers, we need to spend $3000-5000 on an NV card that offers comparable compute performance to a $500 Asus Matrix (of course I realize professional compute differs from consumer compute but for us distributed computing guys, ECC is irrelevant). The hardware power is not that impressive given the size of the chip. What's impressive is that they built a 7B chip and their DP performance is 1/3 of SP, not 1/4th like it is for GCN.

What really happened is NV's engineers took a better approach and outsmarted AMD's by going with more shaders @ lower clocks vs. less shaders @ high clocks. That was a good engineering move by them after Fermi. Also, NV once again had the balls to build a large monolithic die - a 7B transistor chip, with a die size of 520+ I bet. Still, on a compute performance/mm^2, AMD's 10-months old chip offers a lot more compute performance. With HD8970 series, if it has 2560 SPs, it will easily overthrow K20X in both SP and possibly even DP.

Again to me, what's most impressive here is that NV actually built a 7B chip. That's :thumbsup::thumbsup: because it increases the possibility for GK110-based GTX780. As far as compute performance goes, given the size of the chip, it's not impressive per say unless we take power consumption into account. Even then at 235W, this chip has lower SP performance than a 250W TDP HD7970 Ghz. And we all know that HD7970 Ghz doesn't use 250W of power at load.

Comparing AMD / Intel and NV on theoretical FLOPS also misses the point since that doesn't tell us the real world throughput in any particular app. With features like Hyper-Q and Dynamic Parallelism, NV could be 3-5x faster if a particular app takes advantage of these features and runs way faster with CUDA support (we've seen this scenario play out with Quadro vs. anything AMD throws at them). Really, the more important aspects for professional compute are not just theoretical #s but software/application support, ability to meet specific client's needs, technical support/expertise, etc. That's where NV excels and they should since they have invested millions of dollars into HPC compute while AMD is only now taking this space seriously.

If NV can improve yields and 28nm node matures even more by next year, if they can launch an 850-875mhz GK110 2688 SP with a 250W TDP, it would be a monster for sure. Of course we know NV is going to cripple its DP performance for us consumers anyway just like it has done every single generation in its existence.

Seeing how NV met its target of 3x increase in DP Gflops / watt with Kepler over Fermi, things are looking exciting for Maxwell. In 2-2.5 years, K20X is going to look like a toy based on NV's roadmap.

It's definitely exciting to see GPUs become far more important in the world outside of games.

 

[Jaydip]

RS I believe it will be 5.58 Gflops / KWatt

 

[truckerCLOCK]

Most powerhungry?

The S10000 is 3.94Mflops per W.
Telsa K20X is 5.57Mflops per W.

Anyway, in the 225W segment thats its all about.
FirePro S9000, 900Mhz, 0.806 peak DP Tflops, 6GB@264GB/sec. (3.58Mflops/W.)
Xeon Phi 5110P, 1053Mhz, 1.01 peak DP Tflops, 8GB@320GB/sec. (4.49Mflops/W.)
Tesla K20, 706Mhz, 1.17 peak DP Tflops, 5GB@208GB/sec. (5.2Mflops/W.)

If there is an AMD thread.......can count on you to bash them. LOL

 

[Atreidin]

If there is an AMD thread.......can count on you to bash them. LOL

I usually skip his posts and instead replace them mentally with the phrase "AMD is awful" and move on.

 

[Lonbjerg]

If there is an AMD thread.......can count on you to bash them. LOL

I usually skip his posts and instead replace them mentally with the phrase "AMD is awful" and move on.

I suggest people do that with your posts...:whiste:

 

[NIGELG]

I suggest people do that with MY posts...:whiste:

Fixed that for you.

 

[Vesku]

AMD now has impressive raw compute performance, but imo they'll need to implement shared memory space via their HSA initiative in order to have a compelling feature over Nvidia's offerings.

Side note:

If you just compare it on paper, for us consumers, we need to spend $3000-5000 on an NV card that offers comparable compute performance to a $500 Asus Matrix (of course I realize professional compute differs from consumer compute but for us distributed computing guys, ECC is irrelevant).

I thought the 7900 series had ECC?

 

[Jaydip]

No it doesn't

 

[f1sherman]

Some funny stuff. I suppose it goes here.

http://forum.beyond3d.com/showpost.php?p=1679174&postcount=439

 

[Dribble]

It's the software support that's the real problem - you need to do a lot more then just produce a set of drivers with some basic opencl support in. Look at the amount of time, money and effort nvidia puts into the software side of it's HPC offering. AMD puts in hardly any effort and then expects all the software infrastructure to magically appear -it's just not going to happen. It's not like x86 - there is no Intel or MS putting all the effort in which AMD can piggyback on.

 

[tviceman]

Seeing how NV met its target of 3x increase in DP Gflops / watt with Kepler over Fermi, things are looking exciting for Maxwell. In 2-2.5 years, K20X is going to look like a toy based on NV's roadmap.

It's definitely exciting to see GPUs become far more important in the world outside of games.

Everyone's favorite satire writer seems to think Maxwell will be dead in the water and Nvidia is doomed (again).

 

[boxleitnerb]

Care to quote? I want something to laugh