RV350 Die size clue?

[Dave Baumann]

Take a look at the folling images, they show the back of the R350, 9800 board (Red) and the RV350, 9600 board (Green). Take a look at the little curcle of brown chips on the back of the board - from the R350 board shot we can tell that these a placed a little outside of the die area. If this follows for R350 we can begin to get a feel of the the die size of RV350 in relation to R350

- Sorry, thought I'd resized those images to be the same. :?

 

[BoardBonobo]

Would it be possible to cut out those sections, increase the size and put a scale measurement on them? And maybe a little ruler in mm's?

 

[Pete]

If we follow Dave's guideline (assuming the brown chips mark the die boundary, basically just superimpose the die space over the equally-sized RAM chips and compare), it looks like R350 is 1/3 to 1/4th the size of R350 (and eyeball guesstimate: R350 looks to be the size of a RAM chip, RV350 looks to be ~1/4 the size).

 

[mboeller]

I have tried to resize the Cards in Word based on the size of the AGP-Port-Slots.

Then I have drawn an quadrate inside of the small brown chips on both pictures.

In Word the quadrate of the R350 had an size of 1.11 x 1.34 cm² and the RV350 had an size of 0,6 x 0,69 cm².

So it seems the RV350 has only around 28% of the Die-area of the R350 !

This would suggest around 0,28 x 107 x (0,15²/0,13² ) = 40 Mio Transistors.

So IMHO I have made an mistake, cause it seems rather impossible that an DX9 card with 4 PS2.0 Pipelines and 2 Vertex-shaders has only 40 Mio Transistors.


_BUT_ if ATi has really done some magic here and the RV350 is really so small, then Nvidia will loose again with the FX5200/Ultra cause the RV350 can be produced cheaper then the NV34-Chip.

Otherwise what's the purpose of the RV280 / R9200 when the RV350 can be produced so cheap already. Will the RV280 be only for business-PC's and has an price of < $49, or what.


Manfred

 

[Mulciber]

I have tried to resize the Cards in Word based on the size of the AGP-Port-Slots.

Then I have drawn an quadrate inside of the small brown chips on both pictures.

In Word the quadrate of the R350 had an size of 1.11 x 1.34 cm² and the RV350 had an size of 0,6 x 0,69 cm².

So it seems the RV350 has only around 28% of the Die-area of the R350 !

This would suggest around 0,28 x 107 x (0,15²/0,13² ) = 40 Mio Transistors.

So IMHO I have made an mistake, cause it seems rather impossible that an DX9 card with 4 PS2.0 Pipelines and 2 Vertex-shaders has only 40 Mio Transistors.


_BUT_ if ATi has really done some magic here and the RV350 is really so small, then Nvidia will loose again with the FX5200/Ultra cause the RV350 can be produced cheaper then the NV34-Chip.

Otherwise what's the purpose of the RV280 / R9200 when the RV350 can be produced so cheap already. Will the RV280 be only for business-PC's and has an price of < $49, or what.


Manfred

I think you're leaving out of the equation that the R350 is .15u and the RV350 is .13u

 

[mboeller]

I think you're leaving out of the equation that the R350 is .15u and the RV350 is .13u

No, look at the (0.15²/0.13²)-part of the equation. Otherwise it could only have 0.28x107 = 30Mio Transistors.

 

[RussSchultz]

I think Dave's die size estimation idea is a bit dodgy.

All these chips are flip chip BGA packages, which means that the actual die is layed down face down on a substrate (that is much like a small PCB inside the package which routes the pads to the pins/balls). The pins grid will tell you the size of the substrate, but thats all. There's no correlation between the die size and the substrate size (except that its probably smaller than the substrate).

I think the keep out area is more about symmetry and the parts that need to go around the chip being the constraint rather than the chip itself.

Also, the R300 was 250mm^2, so we can assume the R300 is approximately the same size.

 

[PSarge]

I think Dave's die size estimation idea is a bit dodgy.

Agreed, The board layout has everything to do with the substrate design, and nothing to do with the die. Sorry, but I reckon you're barking up the wrong tree.

You'd be better off taking the heatsink off and actually looking at the die. They've probably left it bare to ensure a good thermal connection.

Also, the R300 was 250mm^2, so we can assume the R300 is approximately the same size.

R300 ~= R300. Yes, I'd agree with that also.

 

[V3]

Also, the R300 was 250mm^2, so we can assume the R300 is approximately the same size.

??? From the other thread R300 core is 14.8x14.8mm, so which is which ?

 

[RussSchultz]

Oops, right. At least 14.8x14.8 is 220mm^2, which is in the ball park of 250. (I think I remembered 250 from something else).

 

[Dave Baumann]

You'd be better off taking the heatsink off and actually looking at the die. They've probably left it bare to ensure a good thermal connection.

True, but we don't have that luxury at the moment. As I said before though, the board shown above was the one they took to the briefings - that one didn't have a shim round the chip like the R300/r350 does and I looked between the heatsink and the board. At the time I'd estimated the width of the die to be about 5mm, and it also corellated to the chip positioning on the back. Unless they've done something odd with the dimensions of the chip I'd still say it was substancially less than 10mm in width.

We'll get a better idea when we see the chips though.

Oops, right. At least 14.8x14.8 is 220mm^2, which is in the ball park of 250. (I think I remembered 250 from something else).

AFAIK the actual die size is < 200mm2

 

[demalion]

Why is it a good assumption that separation of components, and layout characteristics is simply proportional to process size?

Are the basic processes that identical in materials, electrical properties, etc, such that their doesn't have to be another factor based on a specific ratio independent of just how fine the process is, but specific to the actual details of dielectrics, trace material, and substrate...i.e., varying from 0.13 process to 0.13 process and 0.15 process to 0.15 process?

Assuming you have access to the characteristics for the specific processes in question, are you taking that into account, Russ?

 

[RussSchultz]

I'm not sure what you're getting at, Demalion. Could you rephrase it and maybe give me a hint about what statements I made that you're commenting on?

(Though pre-emptively: the flip chip paragraph is independant of any process; the size equalivalency betweent he R300 and R350 is just an approximation. I'm sure there's a difference in routing, and they may have even improved it slightly, but not down by over 25%, which is what mbeoler took away from measuring the die size by the keep out area on the back)

 

[demalion]

I'm not sure what you're getting at, Demalion. Could you rephrase it and maybe give me a hint about what statements I made that you're commenting on?

(Though pre-emptively: the flip chip paragraph is independant of any process; the size equalivalency betweent he R300 and R350 is just an approximation. I'm sure there's a difference in routing, and they may have even improved it slightly, but not down by over 25%, which is what mbeoler took away from measuring the die size by the keep out area on the back)

Ok, I'm not talking about the die size guessing, sorry for not being clear, I'm talking about the transistor count estimating that has been done in this thread and elsewhere. I remember you doing so elsewhere, Russ, and that is likely the cause of confusion (you didn't in this thread AFAICS).

 

[RussSchultz]

Reply in whatever thread I made the comment and I'll gladly pull some BS answer from some oriface to explain where I got the number from.

 

[demalion]

Reply in whatever thread I made the comment and I'll gladly pull some BS answer from some oriface to explain where I got the number from.

Well, this is a better place for the discussion, I think, so I'll just link to your comment, and encourage you to clarify here.

But...please "wash it off" first, 'kay?

 

[RussSchultz]

I was guessing at the 70 million, based on a vague recolleciton from some numbers somebody (like MuFu?) had posted earlier.

Why is it a good assumption that separation of components, and layout characteristics is simply proportional to process size?

Generally, the limiting factor in the die area is the transistor size. You can grow the chip up (adding more layers to route signals around), but you can't get the transistors to overlap.

As a bit of anecdotal evidence, some of intel's processors have been actually optically shrunk to smaller processes.

But its all a rule of thumb. Copper vs. aluminium might make a difference (less resistance, so you need less "buffers" to get signals from one side to the other, plus it may run faster/less power). Low K might make a difference since you can route things closer, it could mean the chip has less layers (cheaper) and runs faster/less power. A better route/place program or more time to optimize could make a difference. Top level design choices could make a difference (for example, the R300 shown in this article appears to have partitioned their chip to have functional blocks with blank areas between them.)

Its all guestimating.

I think in general though, both companies have access to the same technologies, they're using the same fabs, etc. So I think we can guestimate ballpark numbers without to much fear of being wildly off. (speculating about how intel's fabs vs. TSMCs fabs, for example would add quite a bit of futility)

 

[MuFu]

I was guessing at the 70 million, based on a vague recolleciton from some numbers somebody (like MuFu?) had posted earlier.

Nope. Probably Uttar - he likes do wild, math-based speculation. :?

MuFu.

 

[Arun]

I was guessing at the 70 million, based on a vague recolleciton from some numbers somebody (like MuFu?) had posted earlier.

Nope. Probably Uttar - he likes do wild, math-based speculation. :?

MuFu.

Bad MuFu! Bad!

Anyway, yes, I did do a math-based guess ( power consumption, considering the RV350 also had to be used for the M10 ) a while ago. And yes, I did say 70M.

But that's a loong time ago. Those last months, I've always been saying 75M transistors. I just can't find who said that originally...
I'm sure it was not MuFu, and I'm sure I didn't invent the number ( could have taken it from an unreliable rumor, however - not sure )


Uttar

 

[RussSchultz]

Sweet. Another urban legend is born.