It the money, Lebowski.

If a 4850 sells for under $100, considering the relative die sizes, the 5850 is almost certainly profitable under $150.

AMD is milking a very fat cash cow as Nvidia is cutting profit margins to the bone to stem the market share arterial bleeding.

AMD can kneecap Nvidia at will ... and come the holidays and the initial 6XXX card releases, they probably will.

Bye bye ATI

i don't get it
can someone point out how the crippled dp units are going to affect my usage?
(some low-end video / image editing, 3d games, web browsing)

I can see somebody who's very relaxed about thing forgiving the crippling of DP, but not if they have to pay $700+ for a card, at that point crippling seems a kick in the face.

The test result they have seems... I don't know, awkward?
There is just a hunch that the numbers are not right
Well if it is right that will be the best thing for everyone
If it is "littlebit off" like the old 9800, I think people know what to do

I will wait for testing done by techreport.
Then I will go look at how those NV-bias website test it
and see if this thing is worth to purchase

However, Nvidia has made the decision to limit DP performance in the GeForce versions of the GF100 to 64 FMA ops per clock—one fourth of what the chip can do. This is presumably a product positioning decision intended to encourage serious compute customers to purchase a Tesla version of the GPU instead. Double-precision support doesn't appear to be of any use for real-time graphics, and I doubt many serious GPU-computing customers will want the peak DP rates without the ECC memory that the Tesla cards will provide.

How will this decision affect Folding@home (And other distributed computing) performance? If the 5870 becomes a more attractive folding GPU then NVIDIA just lost a formerly very interested buyer.

Granted the F@H GPU code is in a state of change and will down the road migrate to OpenCL, so I'm not sure how to tie this all together.

I don't care as long as it scales down well and they release lower price derivative chips that are worth a damn - no 8600-style GPUs please. At least that would give us some competition in an area other than the high-end.

It almost seems like NV might need to split their line if they really want to pursue the HPC market with chips that have HPC required functionality which do nothing for graphics.

I see alot of confusion on why ati cards do poorly and STILL do poorly at folding.

One in 5 shaders on ati hardware yes including the 5000 series is FAT as in having all the stuff needed to do everything. All the rest are thin as in having alot cut out to make them much smaller so they could fit alot more shaders on the chip.

Folding needs stuff that simply isnt included with the thin shaders so they cant run it. Thats all folks.

New architecture yeah!!! Stanford is going to have to make a new program to take advantage of the new design for F@H.

I dont game much anymore, the new family keeps me busy, so I dont have no need for this, I will be saving for an SLC SSD, or a 24" IPS monitor 2 years down the road.

Im sure there are a ton of Nvidia fanboi's just waiting for this though, at least the Win7 drivers should be developed well by the time it hits the shelves.

#11 Under $280 for a HD5850 is an ATi ripoff price?
You do recall it smokes all NVidia cards remotely near that price?
Fail wail.

Looks interesting, but it confirms my fears.

It is another FX.

An interesting, albeit overambitious architecture that got hampered by manufacturing difficulties.

IMO, GF100's design philosophy is another "a solution looking for a problem".

Pretty good chip for a software company.

Surely, this will end up as one overpriced desktop product.
ATI/AMD going to pound Nvidia in the price/performance area once again. History repeats itself?
When AMD drops the HD5870 to $300-$350...

I have to agree with Scott. This GF100 is more future than for current crop of games. The industry has to change to take advantage of GF100 which will take a long time.

And what's up with 64TMU when there are tons of games that's texture based? Efficiency? Is that enough? Nvidia says 50% better texture throughput over GT200 with less TMU. I'll believe it when I see it.

I can't wait for performance #'s but I think it's a lot closer to cypress instead of making a mockery. 6 months late at that.

When are we scheduled to get the next marketing brochure and powerpoint presentation from NVidia to make another front page article?

I am personally disappointed by NVidia's repeated delays in bringing DirectX 11 GPUs to market. This product is at least six months late. I planned to upgrade my graphics card this month, since NVidia's astroturf marketing representative in the forums promised that GF100 would be available in quantity at retail in January.

Instead, AMD has introduced several more DirectX 11 cards and there's still zero competition from the green team. Now we're reading speculation about limited availability in March and graphics card prices seven times as high as a quad-core CPU. :(

I'm guessing nVidia is going to be very careful not to repeat GeForce FX situation so I don't doubt that the top-end Fermi will be faster than the HD5870 on release. However, I think the HD4870 proved that raw speed isn't always the most important factor in a GPU, especially to a company's bottom line. If Fermi is big and only available in limited quantities it might be far too expensive to match the value proposition of the HD5800 series once ATI starts their inevitable price slashing. The situation will be even worse if ATI makes a quick turn-around and release of their rumoured Cypress refresh (a la RV790 HD4890) which would cut into Fermi's performance advantage while still being cheaper.

This looks more and more to be exactly the GPU nVidia wanted - which is to say, a GPGPU in sheep's clothing.

The only 'exciting' thing for gamers is the promise of huge geometry power (and it is VERY exciting), but as Scott pointed out - that is at least one game design generation away.

It looks like they are counting on the nVidiots to keep them afloat for one more generation until they can finally spin all of the fixed-function hardware out of the GPU. At that point it'll really be an unparalleled (pun intended) accelerator, for which there is likely to be a rapacious market.

I wish nVidia well, because they've done a lot for the enthusiast - but it's sad to see them head in the 'off into the sunset' direction of design. Maybe next generation there will be enough volume in GPGPU for there to be a split in the product lines, and nVidia to produce something that's more sensible for the GPU market?

Do we know anything about possible mid-range models based on GF100? The high-end cards sound interesting, but mid-range is where the game's at ($100-$200).

I see nothing wrong with articles of this sort--indeed, if TR is offered the information, TR is obligated to share it. It's no different from the pre-shipping stuff about Prescott, Barcelona, R600 (which AMD had the good sense to postpone prior to shipping), and, of course, nV30. Often it's possible to read between the lines as much as by what a company doesn't say as by what it does say about its upcoming products.

In this case, nVidia is still saying very little of any specific import, which I find troubling. There's still too much here that's theoretical as opposed to practical. This leads me to believe that GF100 product planning is still very fluid at this time.

With nV30, nVidia screwed up enormously by bragging outrageously long before it even knew what sort of nV30 products it could actually ship, but bear in mind that even that would have not detracted from nV30 as much as it did had it not been for the stellar introduction of R300--which just blew the doors off of all nV30 product development regardless of how grandly it was promoted.

This time, lessons learned, nV is playing its hand a bit differently. My own thought is that things are still so theoretical and low key even at this late date so that, if it has to, nVidia can gracefully back away from Fermi with an absolute minimum of egg on the ol' corporate face.

(or graphics processing clusters, I believe, although I thought that name was taken by Gary Phelps' Choice, as we used to call our Dean of Students' preferred smokes back in college)

Scott, you're a riot. Nobody buys those anymore.

Great article TR... it's more in depth than most earlier article out there... it's a bit late, but it's worth the wait... i just cant help but to suspect something fishy, i've read article about this from a lot of sites, they're all confidently claimed " it;s worth the wait"... i dont know why but i cant help to think something is off since there really is not much prove to support that claim so far...

Nvidia could really be losing market share if it's mainstream part got delayed as rumoured... that'd be bad for them and us... still waiting for the price cut... sigh...

I will say that nVidia has put a lot of engineering resources and processing horsepower into this chip; I applaud them for being bold and going for the proverbial 'gold' in their design. Their competitors have found success in designing more conservative and tailored designs, nVidia really seems to be going all-out to achieve the best performance possible.

That said, I piss on them for doing a 'paper' announcement like this to take away steam from their competitors products. And I'm more than a little disappointed in TR for buying into the hype and running the story. This is another press release; that's all. No product, no bechmarks. Just press. I am fascinated by architecture/design information like this so I am greatful for the update, but shame on TR and even moreso nVidia for trying to protect itself from its massive failure of execution by spounting irrelevant nonsense. No product, no talky. A chip in the hand is worth a hundred billion on the white papers.

nVidia deserves to get burned because they couldn't or wouldn't get the job done on the same timeline as their competitors.

Great article Scott, well-done, and thanks. This definitely answers a lot of the unknown's with GF100.

It's always tough to compare the prime optimised current generation chip with the new, future architecture chip. Certainly GF100 will be undercut by RV8xx series chips in terms of size and cost and perf/$ and probably $/watt, but that's just a side effect of the different generations. We'll see later this year what AMD's future architecture chip looks like.

GF100's main problem is that AMD can release a faster RV870 whenever they want, as 1GHz doesn't seem to be out of the question now that 40nm is becoming mature. That will reduce the performance difference, whilst leaving AMD with a chip that is 1/2 the cost to manufacture.

Never mind the option of a 2400 shader chip, but I think that's unlikely.

I wouldn't be surprised if the GF100 can only keep within PCIe power consumption requirements with 512 shaders by disabling most of the 64-bit support.

If it wasn't for NVIDIA's advantage in development/SDK/etc, AMD would be running away with the compute market as well right now.

"By making the window extend throughout the first quarter of the year, Nvidia has given itself ample leeway."

Certainly, as NVIDIA's financial quarter one ends at the end of April, and it's been confirmed that is what NVIDIA mean by Q1...

However, Nvidia has made the decision to limit DP performance in the GeForce versions of the GF100 to 64 FMA ops per clock—one fourth of what the chip can do

Boo, hiss. Do something in drivers instead NV. 1/4 the speed is a really heavy crippling.

I saw all the graphs and got excited until I realized there are 0 actual benchmarks that reviewers have run themselves.

i'm not an eastern european hacker, but i am SO bummed to hear about the crippled DP support on the gaming cards :(

the whole appeal is that you spend a small amount of money and get a lot of processing power. handing over 10x more money for something you know has been artificially crippled truly stinks, why don't i just buy a few amd gpus? (we're doing ok without exception handling and virtual functions so far.)

this dramatically lowers the gigaflops/$ picture for nvidia: a whole order of magnitude! amd's solution is looking WAY more attractive suddenly...

In reply to #4

Not that there is ANY reason to get DX11 card...

Yep, don't forget to say thanks to consoles for that.

I await benchmarks.