Free the video card!

[dakels]

One thing that astounds me is the so called "binning" of chips. Is manufacturing that inconsistent?

 

[BenSkywalker]

Anyways, I've always wondered why we can't have video "motherboards".

There are actually a lot of problems with this idea, mainly it would result in seriously slowed advancement in the GPU space as numerous compromises would have to be made for each generation. You look at one small selection of GPUs, say from the NV6600 up to the 6800U, you have different bit widths for the data bus, different types of RAM supported and different voltage/wattage issues. In order to have a 'vobo' the complexity of the PCB would absolutely skyrocket to the point that you would be paying more for a 6600 then you would now for a 6800 now. Besides that, when the next gen hits with different types/speeds of RAM, different power requirements etc. you would be limited by the 'vobo' you had at the moment. I have spent considerable time trying to think of reasonable way that IHVs could implement something like a 'vobo' and it really isn't feasible until the rate of progress in the GPU market slows down significantly(likely after we see radiosity in hardware- a decade or so from now).

What i dont like about this is...(assuming no imperfections in the cpu/gpu core) why they sell underclocked versions. I mean, it costs the same to manufacture them all...they all come off the same assembly line.

Intel and AMD do the same with CPUs. Say AMD works through 100 wafers and every single one of them bins to 3800+ speeds but they only have orders for two wafers worth of 3800+ processors- what do they do with the rest? Since they are viable parts at 3800+ there is no reason not to sell them at say 3400+ speeds- and for enthusiasts that tends to give us quite a bit of OCing room.

One thing that astounds me is the so called "binning" of chips. Is manufacturing that inconsistent?

If you think about the complexity of the chips and all of them pretty much run with ~15%-20% of each other I wouldn't say it is that inconsistent. 60Million+ transistors on something smaller then a dime working at a few billion clocks per cycles and getting them that close isn't too d@mn shabby

 

[sao123]

i am still of the old school thinking that...
right now with binning you have 2 chips are both identical, flawless, and therefore cost the same to make.
one sells @ 3.0 Ghz for price X, and the other sells at 3.4Ghz for x+y.

My belief is that they should both be sold at 3.4Ghz at price x.

It doesnt cost any more to make them...so binning to me feels like price scamming.

now lowering speeds to bypass defects (a legitimate concern) is another story all together.

 

[fuzzynavel]

interesting discussion.....

If GPUs are so much more advanced(or specialised) then why cant' some of the technology be ported to CPU advances??(don't want to hear things such as patent and business reasons)

Wouldn't it be nice having processors that run 16 instructions per cycle rather than 2 or 3??

 

[itachi]

Originally posted by: sao123
i am still of the old school thinking that...
right now with binning you have 2 chips are both identical, flawless, and therefore cost the same to make.
one sells @ 3.0 Ghz for price X, and the other sells at 3.4Ghz for x+y.

My belief is that they should both be sold at 3.4Ghz at price x.

It doesnt cost any more to make them...so binning to me feels like price scamming.

now lowering speeds to bypass defects (a legitimate concern) is another story all together.

after coming off the manufacturing line they're not all guaranteed to work at the same speed. theoretically, they should all be able to run at the highest rated speed.. but that's only on paper. all processors are tested for speed after they're made.. using ur example, if the processor can run at 3.4 ghz within electrical limits they mark it as a 3.4ghz.. if it can only run at let's say.. 3.0 ghz then they mark it as a 3.0 ghz. them marking down a 3.4 ghz to 3.0 ghz is due to supply and demand. i don't know how economics works or nothing.. and neither do you, so i don't think you should be criticizing their decisions without any understanding. that should answer ur previous question too.

If GPUs are so much more advanced(or specialised) then why cant' some of the technology be ported to CPU advances??(don't want to hear things such as patent and business reasons)

cpus are completely different from gpu's.
1. a gpu only does floating point arithmetic whereas a cpu does all. not only that.. but the instruction sets are completely different.
2. cpus don't need the extra floating point precision offered by gpus.. nor would they benefit from it in any possible way.
3. cpus are designed using pencil and paper.. gpus are designed using a programming language called vhdl and are simulated using computers.. how effective do you think a technology designed and tested primarily by cpus could possibly be for a cpu?
4. last but not least.. the technology of both are completely different and are developed in hardly related enviroments. the memory interface for gpus, today, is 256-bit for the top line and speeds around 1 thz. memory interface available for cpus today is 64-bit.. on top of that, speeds don't get anywhere near 1 thz. gpus are designed looking for that bandwidth.. cpus are designed looking to maximize that bandwidth.

 

[sao123]

if there were that many defects that prevented all chips from running at top speed...the quality control group would be fired in a heartbeat.

 

[itachi]

Originally posted by: sao123
if there were that many defects that prevented all chips from running at top speed...the quality control group would be fired in a heartbeat.

in an ideal world there would be no such thing as violence, poverty, suffering, microsoft, etc.. the list goes on and on and on and on.. since the world is covered with them, single out the person(s) to blame for all of the things that go wrong in our world so we can fire them.

my point is.. get your head out of the clouds, the world is not ideal.. nothing will ever be perfect. the world is full of error. defects will result in either a wafer that can't be used, or a die that can only be partially used or can't be used at all.. the first of the latter is one of the reasons we have radeon 9800se's with 4 pipelines and amd athlon 64 clawhammer with 512k cache disabled. aside from that.. variations in the thickness of the doped silicon to thickness of the n/p substrates leads to variations in the capacitance of the transistors.. higher capacitance = higher current consumption = higher power dissipation. consequently if the die made dissipates too much heat at a certain frequency.. the clock gets lowered til the temperature is within bounds. the end result is a processor that doesn't work at the frequency placed on paper.

if you still don't get it then go read some articles on moore's law, manufacturing process of semiconductors, and characteristics of cmos transistors.

 

[Sunner]

Originally posted by: sao123
i am still of the old school thinking that...
right now with binning you have 2 chips are both identical, flawless, and therefore cost the same to make.
one sells @ 3.0 Ghz for price X, and the other sells at 3.4Ghz for x+y.

My belief is that they should both be sold at 3.4Ghz at price x.

It doesnt cost any more to make them...so binning to me feels like price scamming.

now lowering speeds to bypass defects (a legitimate concern) is another story all together.

How the binsplits work out tends to be tightly kept secrets, but lets say from a batch of P4's Intel gets:
5% 3.6 GHz capable CPU's.
15% 3.4 GHz
25% 3.2 GHz
30% 3.0 GHz
15% 2.8 GHz
10% 2.4 GHz
Most people won't buy the expensive 3.6 GHz or 3.4 GHz ones, in fact most people will probably satisfy with 2.4 or 2.8 parts, so Intel will have a "shortage" of these parts.
So, just mark some of those 3.0's and 3.2's as 2.4's and 2.8's, there you are.
Meanwhile, those very rare 3.6's will be sold at very high prices since a few people will pay 300% times the price for 10% extra performance.

In the end, this is all good, Intel makes good money on the high end, and makes good marketshare on the cheap low end, regular customers can get lower end parts for cheap, the rich can get high end parts at high prices, and enthusiasts can get lower end parts that will perform the same or better as the high end parts for cheap.

If it was 100% 3.2 GHz chips and Intel sold them as such, they'd be cheaper than the 3.6 chips, but more expensive than the 2.4 chips.

Of course, sometimes you'll have far higher demand for some grades, towards the end of Northwoods lifecycle, the process/core was very mature, and Intel probably got very high yields, but still had a large demand for lower end parts, hence the vast majority of 2.4's would make 3+ GHz easily.
On the other hand, during the first big MHz race to 1 GHz, Intel rushed 1 GHz Coppermines out the door before they were really ready to, which resulted in 1 GHz P-iii's that were virtually impossible to find.

 

[BenSkywalker]

1. a gpu only does floating point arithmetic

Wrong.

2. cpus don't need the extra floating point precision offered by gpus.. nor would they benefit from it in any possible way.

Wrong in two ways- processors already support higher levels of FP precission then GPUs and it is used commonly in numerous different applications.

4. last but not least.. the technology of both are completely different and are developed in hardly related enviroments. the memory interface for gpus, today, is 256-bit for the top line and speeds around 1 thz.

1GHZ.

memory interface available for cpus today is 64-bit.. on top of that, speeds don't get anywhere near 1 thz. gpus are designed looking for that bandwidth.. cpus are designed looking to maximize that bandwidth.

Bandwidth on GPUs is designed around what the chip needs, same as processors. GPUs need considerably more bandwidth then CPUs so they are designed as such. GPUs are very much designed to maximize bandwidth- there are considerable lengths taken to ensure that they utilize to make sure that they are making the most effective use of bandwidth possible(HZ, ZCompression etc).

If GPUs are so much more advanced(or specialised) then why cant' some of the technology be ported to CPU advances??(don't want to hear things such as patent and business reasons)

Wouldn't it be nice having processors that run 16 instructions per cycle rather than 2 or 3??

Not realistic for a general processor's useage. I assume you are familiar with a branch mispredict on a processor and its resultant performance hit. Now imagine how bad that would be if your branches were spread out over 16 different units- having to flush multiple pipelines with relatively speaking very low clock speeds would slaughter performance. If the app is coded using multiple threads properly then you could see significant benefit however- watch for multi core CPUs on the horizon to exploit this.

 

[itachi]

damnit.. looks like i have a lot more sht to learn about this.
believe it or not, i actually meant to say 1 ghz.. not 1 thz.

Wrong in two ways- processors already support higher levels of FP precission then GPUs and it is used commonly in numerous different applications.

what's the other way?
with respect to that.. i realize that i was wrong. didn't even think about mmx or sse..

 

[BenSkywalker]

My apologies if my last post came off sounding a bit harsh, I was a bit rushed for time and just summarized.

1. a gpu only does floating point arithmetic

It's actually a fairly small amount of GPU functions, relatively speaking, that are FP. It was only recently that the consumer offerings offered any support for FP color ops at all as an example. Moving forward expect the migration towards FP to continue until almost all functions are FP, but that is a bit away as of now.

2. cpus don't need the extra floating point precision offered by gpus.. nor would they benefit from it in any possible way.

The reason I said wrong in two ways is that one- CPUs do need the extra precission offered by GPUs. The other is that they are already do benefit from it. You could say those two factors are one depending on how you look at it though.

believe it or not, i actually meant to say 1 ghz.. not 1 thz.

Made that mistake myself more then once

 

[sao123]

Itachi Welcome to the community.

my point is.. get your head out of the clouds, the world is not ideal..

Perhaps before you ever attempt to post in the highly technical area again, maybe you should read and understand the rules. And at least try to discuss add to the discussion without insulting people.

in an ideal world there would be no such thing as violence, poverty, suffering, microsoft, etc.. the list goes on and on and on and on.. since the world is covered with them, single out the person(s) to blame for all of the things that go wrong in our world so we can fire them.

my point is.. get your head out of the clouds, the world is not ideal.. nothing will ever be perfect. the world is full of error.

-No value added to the converstaion with any these words.

5% 3.6 GHz
15% 3.4 GHz
25% 3.2 GHz
30% 3.0 GHz
15% 2.8 GHz
10% 2.4 GHz

That might be a binning curve for a new manufacturing process. Something more mature like the northwood 478 chips is more like:
10% 3.6 GHz
30% 3.4 GHz
30% 3.2 GHz
20% 3.0 GHz
7% 2.8 GHz
3% 2.4 GHz
Next, I'd like to point out, that intel usually only has 5 bins that are sold to system builders at any givent time. So those 3% are actually below tolerance and discarded.

Now, what my point was... I understand binning and defects and the such. They are gonna happen.
But, what I think is dumb is ...
Taking a chip that was binned at 3.6 working perfectly. Then reducing its multiplier. Disabling some of its cache. or crippling a perfectly working chip in any way just to sell it in a lower price point.
If your sales are lacking...dont cripple the chip. Just drop the prices.
You wouldnt be losing any money, because it costs the same amount to make all the chips. No matter what bin they actually perform at.

 

[itachi]

Originally posted by: BenSkywalker
My apologies if my last post came off sounding a bit harsh, I was a bit rushed for time and just summarized.

ehh.. don't worry about it, doesn't bother me.

It's actually a fairly small amount of GPU functions, relatively speaking, that are FP. It was only recently that the consumer offerings offered any support for FP color ops at all as an example. Moving forward expect the migration towards FP to continue until almost all functions are FP, but that is a bit away as of now.

i guess i got a lot more to learn about this stuff..

Perhaps before you ever attempt to post in the highly technical area again, maybe you should read and understand the rules. And at least try to discuss add to the discussion without insulting people.

aww did i hurt your feelings ? i'm sorry..

jokes aside, i'll try to play nicer.

 

[kedvale]

Originally posted by: sao123
Now, what my point was... I understand binning and defects and the such. They are gonna happen.
But, what I think is dumb is ...
Taking a chip that was binned at 3.6 working perfectly. Then reducing its multiplier. Disabling some of its cache. or crippling a perfectly working chip in any way just to sell it in a lower price point.
If your sales are lacking...dont cripple the chip. Just drop the prices.
You wouldnt be losing any money, because it costs the same amount to make all the chips. No matter what bin they actually perform at.

I respectfully disagree with your logic here. When you say you will not be losing any money because it costs the same amount to make all the chips, you are only looking at the supply side of the story. But, to get the whole picture, you always need to take both supply and demand into account. As a simple example, let's say there are only 2 speed bins, and a total of 10 chips made by the chip company in question. Let's say 6 of the chips rate 3000, and 4 of the chips rate 3400. Additionally, let's assume there are 10 chip buyers out there. Two of the buyers are willing to pay $800 to get the best of the best chip, while the other 8 are only willing to pay $500 for a chip. Through market research, the chip company knows these buying habits of its customers. Let's look at three possible scenarios:

1. The company will sell all 4 of the faster binned chips at 3400 for $800, and all 6 of the slower binned parts at 3000 for $500. The 2 high end customers will happily buy their chips for $800, and 6 of the lower end customers will buy their chips for $500 (2 of the lower end customers will be out of luck...they will probably either go buy from a competing company, or wait until all prices come down...either scenario leads to excess inventory of 2 chips, which is not good), resulting in total revenue of $800*2 + $500*6 = $4600.

2. The company takes your advice, and lowers the price of the high end chips to $500 (the highest price they can charge to get the "low end" customers to buy the high end chips, which is what they need to sell all their chips), and the slower chips to $400. Now it will sell all of its chips, for a total revenue of $500*4 + $400*6 = $4400.

3. The company takes 2 of the faster chips, and artificially bins them to the slower bin, therefore offering 2 chips at 3400 for $800 and 8 chips (2 of them artificially "crippled" - but the end user has no way of knowing which 2 they are) at 3000 for $500. With this pricing, they will sell alll of their chips, for a total revenue of $800*2 + $500*8 = $5600.

Since all companies are in business to maximize profits, and since, as you mentioned " it costs the same amount to make all the chips", profits will be maximized when revenues are maximized. Hence, the artificial "crippling" to make the desired bins actually makes quite a bit of sense, and is very important to the chip making industry.