Live data from Hacker News

Resistor hack turns a Nvidia GTX690 into a Quadro K5000 or Tesla K10

eevblog.com

141–150 of 196 posts

Re: Resistor hack turns a Nvidia GTX690 into a Quadro K5000 or Tesla K10

#141
post #23

The corresponding reddit thread is quite insightful: "It is far cheaper to make one very good chip for the highest market, and modify it slightly for lower end markets... Now, this is the part you hate: this is the only viable alternative. There's no point demanding different designs for different market segments as that would significantly increase the cost you, the consumer, pays. You can demand chips be sold at lo…

It's also worth noting that the electrical performance of your cheaper card's chip may not be as good as the higher grade chip, so may overheat more or have a failure earlier in it's life. On chip lines like these, there's usually a process called "Speed Binning" where basically when you test the performance of each chip, and you put the higher performing ones in a different "Bin" than the slightly less efficient one…

I had no idea that the manufacturing variation among chips is large enough to create performance tiers. Pass/fail I get, but this sounds like more extreme. Like, the difference between C and fail may be such that some non-negligible percentage of chips that test into the C bin should really be in the fail bin.

Re: Resistor hack turns a Nvidia GTX690 into a Quadro K5000 or Tesla K10

#143

Earlier quoted context omitted.

It's just a situation where capitalism doesn't lead to the most economically efficient solution. The company and the consumers are rational, yet for no additional cost, everyone could be benefiting from better hardware. Unfortunately I don't have a better system to propose :)

Why is that? The major cost is the one-time cost of designing the chip, not producing it. Software is even more extreme: the distribution costs are essentially zero. That doesn't mean the most economically efficient solution is for the software to sell for nothing.

Say you have 2 audiences: Gamers and professionals.

For gamers, a low performance value yields a value of $10, and a high performance value yields a value of $12.

For professionals, a low performance value yields a value of $5, and a high performance value yields a value of $20.

If nVidia sold a single design to all of them, with a high performance value, then gamers would be willing to pay less than $12, and professionals would pay less than $20.

If they want to sell to the entire market, they have to price it less than $12, and lose around $8 from each professional buyer.

However, if they sell 2 designs (e.g: one a cripple of the other), they can derive around $10 of value from each gamer, and $20 from each professional.

The reason this is inefficient, is because for each gamer buying the crippled design, there is a loss of $2 of real value. This loss is a real value loss that could be avoided if there was some other way to extract the actual value for gamers and professionals without crippling the product.

In short, the inefficiency is not that the price is not near the production cost, but that the intentional crippling of the product (necessary to extract maximal value via market segmentation) is causing an actual value loss in the economy.

Re: Resistor hack turns a Nvidia GTX690 into a Quadro K5000 or Tesla K10

#144
post #23

The corresponding reddit thread is quite insightful: "It is far cheaper to make one very good chip for the highest market, and modify it slightly for lower end markets... Now, this is the part you hate: this is the only viable alternative. There's no point demanding different designs for different market segments as that would significantly increase the cost you, the consumer, pays. You can demand chips be sold at lo…

Exactly, binning is good for the business AND for the consumer.

It's good in the sense that it's not clear there's a better alternative, but there is a painful point with a significant part of the market using a crippled design and losing value -- just to make the market segmentation possible.

If there was some other way to do the market segmentation: those who derive $X of value from the chip pay $X, and those who derive $Y pay $Y, without the crippling, then more economic value would be derived.

An example alternative for this approach could possibly be funding such developments with income tax. Then everyone can gain from the benefits of an uncrippled product, and the income tax already approximates how much value you gain from the R&D. This introduces a whole host of other problems, of course, but it does solve the crippling problem.

Re: Resistor hack turns a Nvidia GTX690 into a Quadro K5000 or Tesla K10

#145
post #141

Earlier quoted context omitted.

It's also worth noting that the electrical performance of your cheaper card's chip may not be as good as the higher grade chip, so may overheat more or have a failure earlier in it's life. On chip lines like these, there's usually a process called "Speed Binning" where basically when you test the performance of each chip, and you put the higher performing ones in a different "Bin" than the slightly less efficient one…

I had no idea that the manufacturing variation among chips is large enough to create performance tiers. Pass/fail I get, but this sounds like more extreme. Like, the difference between C and fail may be such that some non-negligible percentage of chips that test into the C bin should really be in the fail bin.

It's many, many different things that can make the quality vary so much. From the purity and quality of the raw silicon itself, to the quality of the design of the chip itself, where some critical part of the chip architecture is incredibly hard to fabricate to the highest standard.

Remember that the widths of the oxide tracks within the silicon is on average 40nm these days (that's only ~400 atoms across!) or even smaller. with hundreds of process steps. One big molecule from some tiny error in the production process on the wrong part of the chip may not cripple it, but may impede performance, it's just probability at the end of the day.

With regards to potential fails going into production, it does happen, there are several test phases during production to catch as many as you can, but at the end of the day you won't get them all.

Semiconductor fabrication is fantastically expensive, new Fab plants cost several billion to build, so if you want to guarantee quality you have to pay for it.

Re: Resistor hack turns a Nvidia GTX690 into a Quadro K5000 or Tesla K10

#146
post #23

The corresponding reddit thread is quite insightful: "It is far cheaper to make one very good chip for the highest market, and modify it slightly for lower end markets... Now, this is the part you hate: this is the only viable alternative. There's no point demanding different designs for different market segments as that would significantly increase the cost you, the consumer, pays. You can demand chips be sold at lo…

It's just a situation where capitalism doesn't lead to the most economically efficient solution. The company and the consumers are rational, yet for no additional cost, everyone could be benefiting from better hardware. Unfortunately I don't have a better system to propose :)

It is a form of price discrimination. In a nutshell, price discrimination is always better for the producer and results are mixed for consumers -- sometimes it leaves them better off, and sometimes worse off.

The claim that this is not economically efficient isn't so straightforward. In particular, your counterfactual isn't "at no additional cost." It would reduce revenue to the card-maker as no one would pay the higher price for the more expensive cards. If the producer can only charge one price, then it would be somewhere in the middle, which would hurt the consumers buying the low end card. Worst case it would make the entire card unprofitable and it would never get produced in the first place.

Re: Resistor hack turns a Nvidia GTX690 into a Quadro K5000 or Tesla K10

#147
post #143

Earlier quoted context omitted.

Why is that? The major cost is the one-time cost of designing the chip, not producing it. Software is even more extreme: the distribution costs are essentially zero. That doesn't mean the most economically efficient solution is for the software to sell for nothing.

Say you have 2 audiences: Gamers and professionals. For gamers, a low performance value yields a value of $10, and a high performance value yields a value of $12. For professionals, a low performance value yields a value of $5, and a high performance value yields a value of $20. If nVidia sold a single design to all of them, with a high performance value, then gamers would be willing to pay less than $12, and profess…

However in your latter scenario, the net loss is only 2$ versus your original scenario where the net loss is 8$. (lose 8$ per card to the perfessional, versus gamer losing 2$ of performance)

So its socially optimum to continue to do what Nvidia is doing ;).

Re: Resistor hack turns a Nvidia GTX690 into a Quadro K5000 or Tesla K10

#148
post #23

The corresponding reddit thread is quite insightful: "It is far cheaper to make one very good chip for the highest market, and modify it slightly for lower end markets... Now, this is the part you hate: this is the only viable alternative. There's no point demanding different designs for different market segments as that would significantly increase the cost you, the consumer, pays. You can demand chips be sold at lo…

Here's a key insight on this issue that the general public seems to be unaware about: the most expensive thing in workstation cards like Quadro and Tesla is warranty and support. This is also the reason that they are clocked down, they have to be more reliable than a gaming gpu.

Gaming GPUs have some parts fused off. These parts are not essential for gaming performance, and powering them off with fuses gives better thermal properties, which allows more gaming performance. These things run HOT.

Finally, this hack does not really make your card into a Tesla or a Quadro, it just fools the driver with a bogus PCI ID and enables some software features that are disabled. It does not enable hardware that has been fused out.

Re: Resistor hack turns a Nvidia GTX690 into a Quadro K5000 or Tesla K10

#149
post #141

Earlier quoted context omitted.

I had no idea that the manufacturing variation among chips is large enough to create performance tiers. Pass/fail I get, but this sounds like more extreme. Like, the difference between C and fail may be such that some non-negligible percentage of chips that test into the C bin should really be in the fail bin.

It's many, many different things that can make the quality vary so much. From the purity and quality of the raw silicon itself, to the quality of the design of the chip itself, where some critical part of the chip architecture is incredibly hard to fabricate to the highest standard. Remember that the widths of the oxide tracks within the silicon is on average 40nm these days (that's only ~400 atoms across!) or even s…

It does make sense in that they are always pushing the edge as hard as they can, and it sounds like this process allows them to do just that. They get to sell the 'lucky' results for a premium, their 'average' for their bread and butter, and even the sub-par chips provide income.

It seems like a really good idea. It also seems like it must be pretty hard to do in a way that gives you a reliable (say) 10% emerging as 'lucky.'

Re: Resistor hack turns a Nvidia GTX690 into a Quadro K5000 or Tesla K10

#150
post #113

Earlier quoted context omitted.

Ow, my eyes. Between the microscopic font and jpeg artifacts. HTML source of the print friendly version: https://gist.github.com/anonymous/5193769 Or less artifacty screenshot: http://i4.minus.com/itHdo0GR7zXw1.png (Sorry to steal any thunder!) Not affiliated, but minus.com doesn't compress high-res PNGs into JPEGs like imgur does (it's only at a certain size that imgur does it, but it's annoying for screenshot threa…

Here is an optimized version of that image: http://i4.minus.com/ivpBgQ8dNf5Ge.png (smaller filesize, loads quicker). Why did you save the image with transparency?

Because that's what scrot did?
Post reply on HN