The mystery of GPU Transister counts?

[QtombeD]

Hi I found a somewhat intereting interview with N. Mohammed on X-bit labs about transister counts....

I know that you can count transister different but usually it's done by a gate count which I believe to be the most accurate one.( I thought ATI uses this and that is prolly one of the reasons they are lower then Nvidia's)

Or as Nvidia counts (I believe they count like this?) count up all transistors on the chip including caches, FIFO's and register files.(not sure if that is done with the human eye+finger precision )

Stating that the all new super Card Chrome20 to only have about 80 million transister really bazzled me. that is about just over the 57 million of the NV20

Nadeem It is not clear to me how some of the big numbers some other folks talk about are generated

So he just haven't got any clue how to count beyond 80 million or S3 is the only company giving the real amount of transistors. I mean most new GPU's are count up to 300+ million are those numbers really just generated by some formula?

Does any1 know how S3 can make a card with just 80million transistors with specs like PS4 dx10 on a .90 chip and unified shaders with the higest clockspeed to date ( I can't even believe this even can support those features)...... and be a real contender So far their cards aren't impressive at all (a real understatement.)

X-bit labs: How many transistors does the Chrome20 have?
Nadeem Mohammad: We currently are not disclosing details, the actual number depends how you count, but for now approximately 60 to 80 million transistors in advanced 90nm, is a reasonable range.

X-bit labs: What do you mean by how to count? Isn’t there only one way count the transistors?

Nadeem Mohammad: Internally graphics companies count their designs in terms of logic gates, which in turn can be translated into transistors by a simple multiplier. In addition modern designs use large amounts of internal memory, for caches and registers etc., their routing density can be very high compared to standard logic. To confuse things further there are analog portions of the design which may have large transistors – many order of magnitude larger than those in logic gates – companies could choose to count them literally or as “logic-transistorâ€￾ equivalent. It is not clear to me how some of the big numbers some other folks talk about are generated – actual transistors or equivalent using logic density or cache density.

 

[Geo]

I'm about this close to having a mini-tantrum over this whole mess on "how do we count/how do they count" that's been going on for about 1.5 years now (okay, some would say longer. . .but very visibly since release of NV40). If I thot it would do any good I'd indulge the instinct.

 

[rwolf]

Transistors are counted by using an electron microscope and counting each one. Sometimes they forget where they left off the day before so the counts are never entirely accurate.

 

[Geo]

That's a better explanation than most!

 

[CMAN]

Transistors are counted by using an electron microscope and counting each one. Sometimes they forget where they left off the day before so the counts are never entirely accurate.

They are actually counted by little elves, and you know how elves forget things.

 

[Ailuros]

So he just haven't got any clue how to count beyond 80 million or S3 is the only company giving the real amount of transistors. I mean most new GPU's are count up to 300+ million are those numbers really just generated by some formula?

Not 100% same sized die pics and I don't think the GammaChrome has more like 60M or so transistors (since it's a single quad GPU), but it might help giving a picture:

http://www.beyond3d.com/previews/s3/s18/index.php?p=01#chip

Die size: 121mm^2

http://www.beyond3d.com/previews/nvidia/g70/index.php?p=03

Die size: 334mm^2

 

[Robbitop]

Just a very simplified calculation:

G70 = 110nm
S18 = 130nm

110nm gets 40% more transistors per area.
G70 got 302 Mio Transistors.

S18 got 121mm^2 , G70 got 334mm^2. So G70 would be 467mm^2 @130nm (+40%).
467/121 = ~3,9.
-> 302/3,9 = ~77Mio Transistors. So 80 Mio seems to be real. And thats for S18. S18 is a one quad variant of columbia. S20 would be 2 quads AFAIK.

 

[Simon F]

I'm about this close to having a mini-tantrum over this whole mess on "how do we count/how do they count" that's been going on for about 1.5 years now (okay, some would say longer. . .but very visibly since release of NV40). If I thot it would do any good I'd indulge the instinct.

Actually, there's not a lot of point in counting gates either because some are much larger than others (e.g. XOR vs AND) and even individual gates of the same type (e.g. two input AND gates) can be different sizes depending on how fast each has to be and what they drive.

AFAIK, the only thing one is interested in is area which is usually measured in mm^2 for an entire chip.

 

[Geo]

So transistor counts are all just flim-flam engaged in as a means of keeping us entertained and annoyed?

 

[Simon F]

So transistor counts are all just flim-flam engaged in as a means of keeping us entertained and annoyed?

Probably.
Well, I guess in the old days (possibly when it was all done by hand!) it was a sensible measure of complexity but not really relevant now.

Of course, the number of transistors is always going up while the area see-saws so the former is certainly a more entertaining "mine's bigger than yours" metric.

 

[SiBoy]

Actually, there's not a lot of point in counting gates either because some are much larger than others (e.g. XOR vs AND) and even individual gates of the same type (e.g. two input AND gates) can be different sizes depending on how fast each has to be and what they drive.

AFAIK, the only thing one is interested in is area which is usually measured in mm^2 for an entire chip.

Transistor count is still a reasonable measure of complexity, given that it is technology independent (i.e. a 130nm chip that is larger than a 65nm chip should not be seen as more complex, so mm^2 is not the be-all-end-all).

Of course the transistor count numbers do tend to get wildly skewed by architectures that have large amounts of on-chip memory.

 

[Simon F]

Transistor count is still a reasonable measure of complexity, given that it is technology independent (i.e. a 130nm chip that is larger than a 65nm chip should not be seen as more complex, so mm^2 is not the be-all-end-all).

I guess it depends on what you mean by complexity. Certainly, if you are only considering cost, then it's got to be area that matters (within the same technology).

If complexity means a measure of the amount of effort that went in to designing it then, given two designs in the same process that perform the same function at the same rate, perhaps the "smaller" chip is the more complex?

 

[olivier]

AFAIK, the only thing one is interested in is area which is usually measured in mm^2 for an entire chip.

so you are saying that ... size does matter !!!!