GF100 reviews are out

[nucleon]

Here's one at Anand:

http://www.anandtech.com/video/showdoc.aspx?i=3783&p=1

Anyone able to get the cuda prime floating around to run on it yet.

I'm curious to see how fast it is. Vs GTX285, it's 3.6x faster in folding@home, 8x faster in other CUDA benchmarks.

-- Craig

[msft]

http://forums.nvidia.com/lofiversion...p?t164417.html

In the GeForce family, double-precision throughput has been reduced to 25% of the full design.

I can save my $500.

[ATH]

Another review:

http://www.tomshardware.com/reviews/...-480,2585.html


About GPGPU:
http://www.tomshardware.com/reviews/...0,2585-20.html

"But with GF100, Nvidia completely reworked the architecture, and the dedicated MAD unit has been done away with. Now, the same units handle single- and double-precision calculation, and performance is reduced only by half with double-precision."

[nucleon]

Originally Posted by msft

http://forums.nvidia.com/lofiversion...p?t164417.html
I can save my $500.

MSFT - that's a reduction of the _planned_ performance. It's still waaay faster than previous generations.

I'd be curious if anyone is going to give it a burl. I'm thinking of getting one - but they aren't out here for another couple of weeks yet (at least).

-- Craig

[joblack]

It seems that the graphic card gpu between home and professional card is the same.

If that's the case it could be easy to get the handbrake out of the 'cheaper version' by patching the graphic card firmware (removing delays or noops) ... :).

[ldesnogu]

That'd be extremely stupid from nVidia not to use fuses to disable hardware to allow segmentation.

[joblack]

Originally Posted by ldesnogu

That'd be extremely stupid from nVidia not to use fuses to disable hardware to allow segmentation.

They would have to designed a different chip to slow down double precision. The easier thing would be to just slow it down with a software solution.

[ldesnogu]

Originally Posted by joblack

They would have to designed a different chip to slow down double precision. The easier thing would be to just slow it down with a software solution.

Not at all, they could just physically prevent some units from being used, just like some cores can be disabled.

I'm not saying they're doing so, that would just make sense.

[nucleon]

The professional version might have ECC memory. The chip supposedly supports it, but I haven't seen any consumer cards with ECC.

-- Craig

[S485122]

Originally Posted by ldesnogu

Not at all, they could just physically prevent some units from being used, just like some cores can be disabled.

I'm not saying they're doing so, that would just make sense.

That would make commercial sense. Fabricating something and impairing part of it while doing that does not make sense. Economy does not make sense. It is a general problem with tools becoming goals. Economy has been a tool for a society, now society is a tool for economy.

Sorry about the of topic ranting.

Jacob

[cheesehead]

Originally Posted by S485122

That would make commercial sense. Fabricating something and impairing part of it while doing that does not make sense.

Rant status noted, here's a counterargument:

Suppose (it's an economic argument -- sorry):

Company A offers a 3 MHz widget for $89 and a 4 MHz widget for $99.

Company A's 3 MHz widgets have a manufacturing cost of $87. Their 4 MHz widgets have a manufacturing cost of $86. (This is explained later.)

The competitors' average market prices are $92 for 3 MHz widgets, $99 for 4MHz widgets.

The competitors' average manufacturing costs are $83 for 3 Mhz widgets and $93 for 4Mhz widgets.

Company A has found a way to use certain rejects on its 4 MHz assembly line which work perfectly at 3 MHz but not at 4 MHz, and manufacture all these 4 MHz widgets, whether passing the 4 MHz test or passing only the 3 MHz test, for $86. That's above their competitors' cost for 3 MHz widgets but below their competitors' cost for 4 MHz widgets. They use the net cost savings on their widgets to sell 3 MHz widgets for less than their competitors, while matching their prices on 4 MHz widgets.

Their manufacturing cost savings on (all-speed) widgets arises solely from their innovative method of being able to market all manufactured widgets from their 4 MHz assembly line, instead of discarding all those that could pass the 3 MHz test, but not the 4 MHz test. They make a simple change to the less-than-4-MHz-capable widgets that restricts their speed to 3 MHz and costs just $1.

A while after their introduction of their 3 MHz widgets at $89, consumer demand exceeds the number that come from less-than-4-MHz-capable rejects, so they divert a small additional proportion of 4-MHz-capable widgets to the line that installs the $1-cost 3-MHz speed limitation.

Isn't this a win-win for both consumers and company A, made possible by company A's innovative use of less-than-4-MHz-capable widgets from their 4 MHz manufacturing line? Either way, the consumer who pays for a 3-MHz-capable widget gets a 3-MHz-capable widget, and the consumer who pays for a 4-MHz-capable widget gets a 4-MHz-capable widget.

- -

Now suppose the situation is that company B can manufacture double-precision-capable chips, and it's cheaper to disable DP on a portion of them than it would be to have their chips manufactured on two separate lines (one for DP, one for non-DP), because they get greater economy-of-scale from the combined line (with a low-cost extra step to disable DP on a portion of them) than on two separate lines. Part of the savings comes from their flexibility in meeting separate demands for DP and non-DP chips without having to have a separate manufacturing line for non-DP chips.

The consumer who pays for a non-DP-capable chip gets a non-DP-capable chip. The consumer who pays for a DP-capable chip gets a DP-capable chip.