Oak Ridge National Laboratory gets Kepler K20 "Big K"

[BenSkywalker]

Nine Mega Watts, people still not understand why Tesla has tighter wattage concerns then GeForce?

 

[3DVagabond]

Nine Mega Watts, people still not understand why Tesla has tighter wattage concerns then GeForce?

Yes. All of a sudden 10% (225w/250w) is a sizable feature. Funny how people think that doesn't matter in a HPC environment.

 

[Plimogz]

so you're thinking 14/15 SMX for power purposes? It could just as well be yields, no?

 

[3DVagabond]

so you're thinking 14/15 SMX for power purposes? It could just as well be yields, no?

As someone else figured out, it's 13smx, if the numbers are correct. It could be for either reason. Most likely it's yields. If it were just power they could reduce clocks and use fully functioning chips.

 

[Lonbjerg]

Hardly any surprise:

The GPU decision was just as simple. NVIDIA has been focusing on non-gaming compute applications for its GPUs for years now. The decision to partner with NVIDIA on the Titan project was made around 3 years ago. At the time, AMD didn't have a competitive GPU compute roadmap. If you remember back to our first Fermi architecture article from back in 2009, I wrote the following:
"By adding support for ECC, enabling C++ and easier Visual Studio integration, NVIDIA believes that Fermi will open its Tesla business up to a group of clients that would previously not so much as speak to NVIDIA. ECC is the killer feature there."
At the time I didn't know it, but ORNL was one of those clients. With almost 19,000 GPUs, errors are bound to happen. Having ECC support was a must have for GPU enabled Jaguar and Titan compute nodes. The ORNL folks tell me that CUDA was also a big selling point for NVIDIA.

that is why I laugh when poeple try and use OpenCL or DirectCompute for gauging GPGPU performance....CUDA way more relevant...even if AMD cannot run CUDA.

 

[Plimogz]

As someone else figured out, it's 13smx, if the numbers are correct. It could be for either reason. Most likely it's yields. If it were just power they could reduce clocks and use fully functioning chips.

I do think that it's 14, according to the numbers I've seen. The AT K20 preview said max cores was 15x192=2880 http://www.anandtech.com/show/5840/gtc-2012-part-1-nvidia-announces-gk104-based-tesla-k10-gk110-based-tesla-k20 and the ORNL Titan AT write-up says 2688 cuda cores. So 14x

 

[3DVagabond]

I do think that it's 14, according to the numbers I've seen. The AT K20 preview said max cores was 15x192=2880 http://www.anandtech.com/show/5840/gtc-2012-part-1-nvidia-announces-gk104-based-tesla-k10-gk110-based-tesla-k20 and the ORNL Titan AT write-up says 2688 cuda cores. So 14x

I had only seen BSN's figures. I'm inclined to accept AT's over BSN's though. So, 14 it is.

 

[VulgarDisplay]

Hardly any surprise:



that is why I laugh when poeple try and use OpenCL or DirectCompute for gauging GPGPU performance....CUDA way more relevant...even if AMD cannot run CUDA.

I don't laugh because CUDA is going to become extinct simply because it's not open source.

 

[Jaydip]

dp

 

[Jaydip]

Yeah the same way D3D has become extinct compared to OGL

 

[VulgarDisplay]

Yeah the same way D3D has become extinct compared to OGL

I wasn't aware that D3D couldn't be used on hardware not owned by Microsoft?

I fail to see the comparison.

 

[Jaydip]

I wasn't aware that D3D couldn't be used on hardware not owned by Microsoft?

I fail to see the comparison.

That is not NV's fault.AMD could have licensed it.NV simply has vastly superior ecosystem.

 

[sontin]

I don't laugh because CUDA is going to become extinct simply because it's not open source.

Even if this would be true, nVidia would still be king of the hill.
Look at the Workstation market: OpenGl and nVidia has a marketshare of 85:15.

AMD needs much better hardware to put CUDA in the backseat. In the same power envlope K20 is much faster than w9000. There is no reason to go with AMD from a hardware point.

 

[Lonbjerg]

I don't laugh because CUDA is going to become extinct simply because it's not open source.

You mean like Windows vs Linux?
Like DirectX vs OpenGL?

Wishfull thinking is just that...wishfull thinking.

0/10

 

[BenSkywalker]

I don't laugh because CUDA is going to become extinct simply because it's not open source.

In the SuperComputer space I haven't ever seen this become a serious issue. Almost all apps are custom built to run on the exacting machine available to them. While CUDA lead time helps nVidia out a lot in terms of building a strong development community, if AMD came down with the hardware and development software to back it up there is nothing stopping them from replacing nVidia is this exacting system. The issue is, nVidia has *had* the software side done for years, and it is relatively recently that they have made major inroads in the segment. A large part of that is you have to go through the having a viable part *with* the development tools available so proofs that your setup can work before being selected in the planning stages for such a device. In other words, if AMD had it all right now, which they don't, it could be another couple of years before we actually see those results.

 

[3DVagabond]

You mean like Windows vs Linux?
Like DirectX vs OpenGL?

Wishfull thinking is just that...wishfull thinking.

0/10

While I don't agree that CUDA is going to go away, I don't think your comparisons really prove your point for HPC. Who uses Windows and Dx?

 

[BenSkywalker]

Yeah, posting right after I just did, hate it when I do that but I found the numbers fast enough that noone replied

The average for US households concerning electrical usage is 940kWh/month. In other words, every ~nine and a half minutes this machine uses as much electricity as an average house consumes in a month.

25 watts x 18.7K is another ~467kW. Honestly I'm not sure what the perfromance increase would be if we gave nV another 25 watts to play with so it may or may not have been *worth* it, but I think they had an energy target one way or the other regardless.

 

[Lonbjerg]

While I don't agree that CUDA is going to go away, I don't think your comparisons really prove your point for HPC. Who uses Windows and Dx?

The majority of PC users and gamers.
Just like CUDA is the most used GPGPU language

the open source crowd are a boring bunch...always with the "wishfull" arguments...to heck with facts or reality.

Bottom line is still:
CUDA (and it's eco-system) are lightyears ahead of the competition....like it or not.

 

[BenSkywalker]

CUDA (and it's eco-system) are lightyears ahead of the competition....like it or not.

XeonPhi would like to argue that point with you

Not saying how it will turn out, but at the current time it seems like the HPC space is going to be owned by the two of them.

 

[3DVagabond]

The majority of PC users and gamers.
Just like CUDA is the most used GPGPU language

the open source crowd are a boring bunch...always with the "wishfull" arguments...to heck with facts or reality.

Bottom line is still:
CUDA (and it's eco-system) are lightyears ahead of the competition....like it or not.

I'm not talking about liking anything. Stop making it personal. I'm only pointing out that we aren't talking about PC's or gamers. In just about everything else, and particularly HPC, Linux rules. Specifically because it's open source and you can do whatever you need to to it to fit it to your needs.

 

[Lonbjerg]

I'm not talking about liking anything. Stop making it personal. I'm only pointing out that we aren't talking about PC's or gamers. In just about everything else, and particularly HPC, Linux rules. Specifically because it's open source and you can do whatever you need to to it to fit it to your needs.

Point me to something similar to CUDA.
Not something WISHING it were CUDA...but something that actually is a real competitor?

If not...you point is nothing but a red herring.

 

[sontin]

HPC is not open source.

 

[3DVagabond]

Point me to something similar to CUDA.
Not something WISHING it were CUDA...but something that actually is a real competitor?

If not...you point is nothing but a red herring.

I don't have to point at anything. I'm telling you that citing Windows and Dx doesn't make your point because we aren't talking about gaming or PC's. Linux is the OS of choice and it's open source.

 

[Jaydip]

XeonPhi would like to argue that point with you

Not saying how it will turn out, but at the current time it seems like the HPC space is going to be owned by the two of them.

It is still x86 and if you pair it with gpu it can still be accelerated via openacc. I don't see x86 rivaling gpu in parallel workload anytime soon. Moreover all the existing apps need to be redesigned not just recompiled just by using the mmic flag

 

[ShintaiDK]

It is still x86 and if you pair it with gpu it can still be accelerated via openacc. I don't see x86 rivaling gpu in parallel workload anytime soon. Moreover all the existing apps need to be redesigned not just recompiled just by using the mmic flag

Xeon Phi is doing above 1Tflop in DP. And thats x86. Thats the same ballpark as GK110s DP. Xeon Phi is just much easier to code for and dont have performance penalities in alot of cases like AMD/nVIdia cards will.

Just see on AMD vs nVidia when 1 single instruction is missing in bitcoin mining.