http://www.anandtech.com/show/6446/nvidia-launches-tesla-k20-k20x-gk110-arrives-at-last

---

Reno, NV
Team: SETI.USA

OK so you don't need a new car... and have a spare $3199 lying around for the little one,
but where are you going to plug your monitor in?

;)

---

FAQ's

HOW TO:
- Opt out of Beta Tests
- Ask for Help

Not into a Tesla, that's for sure ;)

MrS

---

Scanning for our furry friends since Jan 2002

So who's got a spare 35k lying around for a nice SR-2 build?

There is no mention of releasing the K20 (aka BigKepler) as a GeForce card in this article. Just as I've expected. This is bad news for the common crunchers like me. I was wrong about the number of cores (my guess was 1024), in reality it has 960 DP and 1536 SP cores. So it's not that different from what I've expected. With these specs this card wouldn't be better here than a GTX680 (GPUGrid doesn't use DP).

Unless there is a second release wave/series, akin to the GF400 and GF500 series (1024 rather than 960, similar to the 512shaders up from 480), I don't think crunchers will be looking to these cards. Even then it could be 9 months or more. With top supercomputers looking more and more to GPU's, they just won't come our way. All it takes is a few super computers to be interested and every Tesla is gone. Anyway, at $3500 and up they are probably not financially viable outside the supercomputer market. Even if one could do 3times the work of a GTX690 it still costs more than 3times a GTX690 and needs to be supported by more expensive hardware.

To me it seems more likely that a Quadro product will turn up with a design more similar to the Tesla's than GF600'S. Maybe wishful thinking but the Quadro 6000 and 7000 are basically beefed up GTX470's - a generation or two old. So the line up needs a refresh.

---

FAQ's

HOW TO:
- Opt out of Beta Tests
- Ask for Help

Obviously buying Teslas for BOINC won't be an option.

However, saying that GK110 is not any better than GTX680 at GPU-Grid is simply wrong. It's got 15/8 the amount of shaders, i.e. in the largest configuration 87.5% the performance at similar clock speeds. Stock clock speeds will be lower, but so will be the voltage -> higher efficiency than a stock GK104 using Turbo Mode.

MrS

---

Scanning for our furry friends since Jan 2002

However, saying that GK110 is not any better than GTX680 at GPU-Grid is simply wrong. It's got 15/8 the amount of shaders, i.e. in the largest configuration 87.5% the performance at similar clock speeds.

Ohh, it's my mistake. Let's blame it on that I've used my phone to multiply....
It has 2880 SP cores (at least on the figure).
In reality one or two SMX are disabled (obviously because manufacturing reasons similar to the GF100), so the real world K20 and the K20X has 13/8 and 14/8 the amount of cores of a GTX680, respectively.

Stock clock speeds will be lower, but so will be the voltage -> higher efficiency than a stock GK104 using Turbo Mode.

Sure, but I think that a chip of this size (and transistor count) couldn't run at the clock speed of a GK104 even at the voltages of a GK104.
Just like a GTX 580 runs at lower voltages and higher (shader) frequency than a GTX 680.

BTW, hte GPUGrid staff should apply for computing time at OAK Ridge :)

While I expect we're going to see a consumer card based on GK110 I would not bet on it being the fastest possible configuration with 15 SMX. So we may very well see 14 or 13 SMX and the reduced performance this results in.

And you're right, clock speeds of a huge chip are generally lower than for a smaller chip. Some signals have to travel further and the possibility of exceptionally slow transistors being somewhere where it hurts overall clock speed increases.

However, the example you've choosen is not good to show this off :p
Fermi was, like previous chips since G80, a high frequency design. This allowed the shaders to be clocked significantly higher than the rest of the chip. Kepler drops this in favor for more power-efficient transistors and hence can't reach quite the clock speeds of the predecessors, even using a more advanced process technology. And GF110 was actually the physically huge chip (~550 mm²), whereas GK104 has a comparably moderate size (~350 mm² if I remember correctly).

A better example would be G92b and GT200: same process node (55 nm), rather similar architecture, but 1.8 GHz versus 1.4 GHz stock speeds. Part of this difference was a power-limitation of the big chip, though.

I'd agree if we speculated that GK110 might reach ~100 MHz less than GK104 at the same voltage, but I wouldn't expect a more dramatic difference.

MrS

---

Scanning for our furry friends since Jan 2002

Is it possible to use the Nvida Tesla K20 and K20X with GPUGRID?
Are they working as well on the other Boinc projects?
Actually bring benefits or are they working as the GTX690?
It makes sense to buy a PC with 4 Nvidia Tesla K20X for better crunching?

It makes sense to buy a PC with 4 Nvidia Tesla K20X for better crunching?

No, it certainly doesn't. The strength of K20 and K20X compared to the other Kepler chips) is mainly strong performance in double precision - which GPU-Grid doesn't use. It's also got potential if the new features alias HyperQ and *Something* are used, but this requires a (moderate) code modification. And it's only useful if you can't make good use of a GPU with a single thread (e.g. POEM, Einstein, SETI etc.), which doesn't apply to GPU-Grid.

I suspect these Teslas would actually be slower than a GTX690 here.

MrS

---

Scanning for our furry friends since Jan 2002

It makes sense to buy a PC with 4 Nvidia Tesla K20X for better crunching?

No, it certainly doesn't. The strength of K20 and K20X compared to the other Kepler chips) is mainly strong performance in double precision - which GPU-Grid doesn't use. It's also got potential if the new features alias HyperQ and *Something* are used, but this requires a (moderate) code modification. And it's only useful if you can't make good use of a GPU with a single thread (e.g. POEM, Einstein, SETI etc.), which doesn't apply to GPU-Grid.

I suspect these Teslas would actually be slower than a GTX690 here.

MrS

I always thought simulating a molecule atom by atom would require at least more decimals than the floats provide...

I mean, why would Milkyway require double precision and not this?

GTX680 and GTX690 are enough for good crunching and they are not so "expensive" like Tesla GPUs.

We have tested on a K20 and it is on par with a GTX680, a bit less.
The K20X should be a bit faster.

I am not considering the price component of course.

gdf

We have tested on a K20 and it is on par with a GTX680, a bit less.
The K20X should be a bit faster.

I am not considering the price component of course.

gdf

Thank you for this results!

---

Member of Boinc Italy.