G70, G71, R520, R580 die size discrepancy

[Geo]

I have not seen any accurate measurements on G71. It still look bigger than 196mm^2.

You do get that the chip chart isn't from NV PR materials? It's from Wavey measuring the one in his possession with his own ruler. . .and by that standard he was able to confirm what NV said.

The same would be true for G73, Wavey measured it himself: http://www.beyond3d.com/misc/chipcomp/?view=chipdetails&id=111&orderby=release_date&order=Order&cname=

 

[Megadrive1988]

Firingsquad is reporting that G71 has 278M transistors. I don't get it. how would Nvidia be able to keep G71 architecture the same as G70 with only 278M transistors in G71 compared to ~302M in G70?

 

[Dave Baumann]

By reducing the pipelining. G70 probably has more transistors there solely dedicated to incrasing the clockspeed for the 90nm process, which they have removed with 90nm.

 

[rwolf]

By reducing the pipelining. G70 probably has more transistors there solely dedicated to incrasing the clockspeed for the 90nm process, which they have removed with 90nm.

I think you mean 110nm.

 

[Megadrive1988]

ahh I see. interesting.

 

[3dcgi]

Firingsquad is reporting that G71 has 278M transistors. I don't get it. how would Nvidia be able to keep G71 architecture the same as G70 with only 278M transistors in G71 compared to ~302M in G70?

I doubt it was pipelineing. They probably reduced the size of some fifos that were overdesigned in G70.

 

[atomt]

You do get that the chip chart isn't from NV PR materials? It's from Wavey measuring the one in his possession with his own ruler. . .and by that standard he was able to confirm what NV said.

The same would be true for G73, Wavey measured it himself: http://www.beyond3d.com/misc/chipcomp/?view=chipdetails&id=111&orderby=release_date&order=Order&cname=

The link was for G73 and not G71. I already admitted I was dead wrong with G73 since the 20% delta is obvious. I wasn't counting too much on G71 due to high measurement errors
with photos. My last estimate was 200mm-220mm, so it could well be my measurement
error.

If it makes you happy, I start counting my balls the Nvidia way as well. No thanks on Nvidia size efficiency. I rather have big balls than efficient ones.

 

[Geo]

Ah, my bad. There's G71 http://www.beyond3d.com/misc/chipcomp/?view=chipdetails&id=112&orderby=release_date&order=Order&cname=

 

[Fodder]

I doubt it was pipelineing. They probably reduced the size of some fifos that were overdesigned in G70.

Official stance seems to be ...

NVIDIA says it has replumbed the G71's internal pipelines throughout the chip, making them shorter, because those longer pipes and extra transistors aren't needed to help the G71 achieve acceptable clock speeds—the faster-switching transistors of the 90nm process will suffice for that purpose.

http://www.techreport.com/reviews/2006q1/geforce-7600-7900/index.x?pg=1

 

[JoshMST]

I had also heard from a 3rd source that NVIDIA had removed some of the SLI circuitry that wasn't being used, so that is another area of transistor decrease. I've also heard of a lot more custom cell work in the G71 vs. standard cell. That also helps to get rid of some of the pipelining.

Still, damn impressive to pare it down that much yet still have it run as well as it does.

 

[superguy]

Somebody please pull apart their 360 and measure Xenos shader core?

I'm betting with ATI seeming to undercount, and Nvidia getting more density, it might actually be larger than G71 (eyeballing photos is seems possible)

 

[3dcgi]

Official stance seems to be ...http://www.techreport.com/reviews/2006q1/geforce-7600-7900/index.x?pg=1

I'd think that would be a lot of work for minimal gain unless they're talking about math units that were generated with Arithmetica or Module Compiler. I don't have personal experience with these tools, but they can adjust the number of pipes stages automatically in certain types of code.

 

[Mintmaster]

I'd think that would be a lot of work for minimal gain unless they're talking about math units that were generated with Arithmetica or Module Compiler. I don't have personal experience with these tools, but they can adjust the number of pipes stages automatically in certain types of code.

Remember RSX is designed to operate at only 550MHz (and even that is way more than necessary for the bandwidth involved), and with the volume you'll have with PS3, you want to save every transistor that you can. There's quite a bit to gain by doing such heavy retooling.

Still, your theory about FIFO reduction is likely to have plenty of truth to it. If you look at Digit-Life's pixel shader benchmarks, G71 gets a bigger gain for moving to FP16 than G70. This meshes well with DeanoC's comments that NVidia is telling devs to use FP16 whenever possible (I assume that was mostly RSX oriented).

 

[KimB]

Remember RSX is designed to operate at only 550MHz (and even that is way more than necessary for the bandwidth involved), and with the volume you'll have with PS3, you want to save every transistor that you can. There's quite a bit to gain by doing such heavy retooling.

Well, being more memory bandwidth constrained isn't necessarily so bad for a console, as it's more realistic for developers to write their software so that it can operate with less memory bandwidth (i.e. stress ALU operation more, as well as just long pixel shaders in general).

 

[Mintmaster]

I agree, and have said so many times. Just pointing out that there are very good reasons for NVidia to save transistors by reducing the pipelining at the risk of a lower clock speed ceiling.

 

[3dcgi]

Remember RSX is designed to operate at only 550MHz (and even that is way more than necessary for the bandwidth involved), and with the volume you'll have with PS3, you want to save every transistor that you can. There's quite a bit to gain by doing such heavy retooling.

Still, your theory about FIFO reduction is likely to have plenty of truth to it. If you look at Digit-Life's pixel shader benchmarks, G71 gets a bigger gain for moving to FP16 than G70. This meshes well with DeanoC's comments that NVidia is telling devs to use FP16 whenever possible (I assume that was mostly RSX oriented).

I thought we were talking about G71, not RSX, which runs at over 600MHz. I agree that such work makes more sense for RSX.

 

[Mintmaster]

I thought we were talking about G71, not RSX, which runs at over 600MHz. I agree that such work makes more sense for RSX.

It would be quite stupid for any pipeline optimizations in the 90nm G70-based RSX to not be used in the 90nm G70-based G71. I'm sure the projects overlap a huge amount. Looking at the results, they definately made the right decision.

RSX aims for high yeild at 550MHz. G71 aims for just enough yeild at 650MHz to satisfy GTX demand, and the rest goes into the 450MHz GT.

 

[fellix]

OK guys, haven't track the whole thread, but I find it a perfect place for another suggestion, concerning the yelds from ATi and nV top chips being prodused now: G70 & G71 and R580 & R520. All wafers are 300mm.
Sure the numbers are not exact, but I believe the error margin is within 3~6 dies being knocked out during the slicing process.

 

[Razor1]

323 vs. 175. 45% difference of the r580 vs g71 chips per wafer. Sounds about right :smile:

216 for the r520

 

[Jawed]

OK guys, haven't track the whole thread, but I find it a perfect place for another suggestion, concerning the yelds from ATi and nV top chips being prodused now: G70 & G71 and R580 & R520. All wafers are 300mm.
Sure the numbers are not exact, but I believe the error margin is within 3~6 dies being knocked out during the slicing process.

Damn, for a second there I thought you had yield graphs!

Jawed