Ryan Smith's review, http://www.anandtech.com/show/3809/nvidias-geforce-gtx-460-the-200-king

I know what "I'd" do with an 8 SM enabled GTX460! I'd launch a GTX465(384) LOL!

---

Anyone knows how much faster 800 Mhz GTX 460 is compared to gtx 260?

Theoretically that's 25% increase from gpu clocks and 50% more sp.

The GTX 460 design is somewhat different than the previous Fermi cards, so we don't really have an accurate reference to go by. Compared to a GTX260, my GTX470 @ 707MHz (WinXP) can get about 65K credits per day, while my GTX260 (WinXP) could in theory only get about 40K per day, overclocked (core 625MHz, Shaders 1525MHz, RAM 1100MHz).
My guess is that a GTX460 @ 800MHz could reach about 55K per day, all being well. That is smack in the middle of the 25% to 50% you mentioned. However no-one has tested one here yet, so we dont even know for sure that it will work! In theory it should work, but there could be issues; the shaders are controlled in three groups, so it might work more like a card with 256 shaders than a card with 336 shaders (under crunching conditions), or not?
It is also a bit early to be bench marking Fermi card here. They will eventually benefit from CUDA updates, and there is another Beta CUDA version for developers released for assessment (3.2). Perhaps in a month or two we will see driver and app improvements that will make the Fermi cards look more attractive.

I'm expecting a GTX 460-768-OC in a couple of days.

They were out of the 1GB version so the baby 768MB version is on the way.

The 1GB version has a wider memory bus, more RAM and more ROPs (whatever they are), so it is said to be worth the extra cost.

Reviews of the card have been very positive.

I'm torn between GPUGRID and Folding...

The 9800GT cards are Folding because they are better at Folding.

The ATI card are Folding because you can't GRID with ATI cards.

The GT240 cards are GRIDing because they are better at GRIDing.

The GTX460 is going to replace an ATI HD4850 which is currently Folding. If the 460 is used for Folding more CPU time will be available for Rosetta (since nVidia cards use less CPU time than ATI cards). Folding and Rosetta scores would both go up.

Suggestions?

Wait till the 1GB version is available and buy one for GRIDing and one for Folding you say?

Good idea!

I would suggest GPUGrid all the way! If you could at least run GPUGrid for a few days so we can get a good idea of how they perform for real that would be great.

---

Thanks - Steve

I can run Folding and GPUGrid tasks at the same time on my GTX470.

Perhaps you will be able to do the same on your GTX460

Definitely try both.. or take a look around where they're more efficient. Depending on the code GF104 can perform like 336 shaders or like 224 shaders.

MrS

---

Scanning for our furry friends since Jan 2002

Actually I hope that as 4/7 execution units can be used at once instead of 2/6 of GF100, maybe GF104 also has better sustained performance compared to a GTX480 for a well optimized code.

gdf

Getting a sustained 2-way superscalar execution shouldn't be too hard, especially factoring in loads & stores. GPU-Grid would use the 3 blocks of 16 shaders, one block of 16 load/store units and sometimes the one block of 8 SFUs, right?

Edit: so for every load/store op you'd need to find at least (3 shader ops) or (2 shader ops plus 1 special function) and not more than (3 shader ops plus one special function). If there are no pipeline stalls (all data fetched into caches early enough) this could lead to sustained 4 ops/clock operation :)
All of these ops taken from 2 different warps, of course.

MrS

---

Scanning for our furry friends since Jan 2002

Getting a sustained 2-way superscalar execution shouldn't be too hard, especially factoring in loads & stores. GPU-Grid would use the 3 blocks of 16 shaders, one block of 16 load/store units and sometimes the one block of 8 SFUs, right?

Yes.


gdf

So it looks like Revenarius could not get GPUGRID to work with a 460:

http://www.gpugrid.net/forum_thread.php?id=2227#18002

Has anyone else tried or is it known that the 460 is not going to be able to GRID at this time?

Yep, tried 3 projects. See the GTX 460 thread for more info...

A Fermi with all 512 shaders?

http://en.expreview.com/2010/08/04/exclusive-the-design-card-and-testing-of-geforce-gtx-480-preview/8878.html

This is supposed to be a GTX480 with 512shaders. Although this report could be a fake, which would discredit the website, its as likely to be a sample card, and GPUZ just doesn’t have a card to reference it from its database. Although a 512shader GF100 is expected, it's more likely to be called a GTX485, and it's worth noting that not all sample cards make it to production, and when they eventually do get mass produced there can be differences. If this was destined for the more expensive version of the Tesla, then its lacking some GDDR, which would also need to be ECC.

While there has been nothing official, I can't see NVidia wanting to make a song and dance about a GF100 with 512shaders - the words late and hot would once again make headlines; better to slip it out the back door as a golden sample or limited ed. possibly along with some other release.

Fake or not it makes a lot of sense: if the cards are different, nVidia will give them the same number (GTX260, GTX460 etc.). They spare the different numbers for similar cards ;)

MrS

---

Scanning for our furry friends since Jan 2002

I've heard this pcb might be msi lightning version of 480.

Now it would be really cool if nvidia allowed some partners to make limited runs of 512 sp cards without doing official release.

So it looks like Revenarius could not get GPUGRID to work with a 460:

http://www.gpugrid.net/forum_thread.php?id=2227#18002

Has anyone else tried or is it known that the 460 is not going to be able to GRID at this time?

Works alright for me, if a little slow.

---

BOINC blog

I've heard this pcb might be msi lightning version of 480.

That would make sense since 6 chips on the front and 6 on the back match well with a 384 bit bus.

@MarkJ: that post was a little old. So far GTX460 runs but not very efficiently. Possible improvements are expected in September, if everthing goes well.

MrS

---

Scanning for our furry friends since Jan 2002

No thanks to a 3month driver gap the GTX460 still runs inefficiently on Windows, and along with the GTS450 not at all on Linux. However this will hopefully change next week with the new software, and the release of 2 drivers, a week apart. A bit of skulduggery perhaps but these drivers don’t work well with the 200series cards, so avoid unless you have a Fermi. Whatever the motive it looks like there is a push to get things to work properly for the Fermi cards.

To elude to a possible +ve motive, the 261.00 driver seems to have been pulled for revealing a series name jump in the form of a GTX580, along with other discerning naming/sequencing info; apparently the GTS455 will have less cuda cores (shaders) than the GTS450 !?! We’ll see.

Let’s hope these were typo’s, but despite the names another couple of non-OEM Fermi’s might be useful here, especially if the GF110 revision brings any performance improvements. As always I would suggest people wait until these are confirmed as being well supported and their crunching performances reported before buying one. Release dates are supposed to be in Nov.

LOL .. that naming mess would be so nVidia!

MrS

---

Scanning for our furry friends since Jan 2002