NVidia Ada Speculation, Rumours and Discussion

Status
Not open for further replies.
"So I've mapped out the dart board and there's a couple of possible combinations I've not attention-whored yet.

So imma do that now, so I have complete coverage of ALL possibilities and can claim victory when the reality inevitably coincides with one of my fantasies".

This coprolite guy is silly.
 
Kopite7kimi definitely seems to get some real inside info if you look at their past tweets about Ampere from 2020.

Igor Wallossek also seems thinks their info is good:


To me it seems a lot more plausible that kopite7kimi gets actual inside info on Nvidias product development and that the specs get changed during the process than that they're just outright making shit up.
 
To me it seems a lot more plausible that kopite7kimi gets actual inside info on Nvidias product development and that the specs get changed during the process than that they're just outright making shit up.
No shit specs change during development. It's why most reliable leakers don't mention product specs up until they actually go into production with such specs. Even then it is possible for clocks to change. You don't need to have inside info to know that.

Also I would say that SM and such numbers are the least interesting part about any new architecture these days. If someone would actually have a solid inside info they would probably talk more about things which are new, not the things which everyone know will be there, changing their mind with each new configuration some AIB had heard about from IHV sources. An example with Ampere is the doubling of FP32 and G6X RAM - neither of which were reported by kopite7kimi ahead of other leakers AFAIR.
 
No shit specs change during development. It's why most reliable leakers don't mention product specs up until they actually go into production with such specs. Even then it is possible for clocks to change. You don't need to have inside info to know that.
Read again. I didn't claim you need to have inside info to know that specs change during development. I said the accuracy of past info implies they have inside information.

An example with Ampere is the doubling of FP32 and G6X RAM - neither of which were reported by kopite7kimi ahead of other leakers AFAIR.
First or not, interesting or not, probably not just made up is my point:



But maybe I'm just gullible.
 
Read again. I didn't claim you need to have inside info to know that specs change during development. I said the accuracy of past info implies they have inside information.
Sure but the extent of this information can easily be the number of units per each chip in the lineup. Which is the least interesting information and also one which you can't do much with because you won't know final configurations up until products will go into full scale production. Reporting on what such configurations can be is pointless and also misleading - like it was with 3080 20GB which never happened.
 
No shit specs change during development. It's why most reliable leakers don't mention product specs up until they actually go into production with such specs. Even then it is possible for clocks to change. You don't need to have inside info to know that.

Also I would say that SM and such numbers are the least interesting part about any new architecture these days. If someone would actually have a solid inside info they would probably talk more about things which are new, not the things which everyone know will be there, changing their mind with each new configuration some AIB had heard about from IHV sources. An example with Ampere is the doubling of FP32 and G6X RAM - neither of which were reported by kopite7kimi ahead of other leakers AFAIR.
Regarding the doubled FP32, I believe KittyYYuko suggested it earlier on and kopite7kimi "confirmed" it nearer to launch.

 
kopite7kimi hasn't been very reliable lately I know he has been spot on with Ampere but this gen has been wrong about a lot of stuff. for example he said hopper was >1000w and that its die size is 1000mm2. he also says ad102 is 70-80B transistors at 600mm2 which also doesn't make sense.
 
kopite7kimi hasn't been very reliable lately I know he has been spot on with Ampere but this gen has been wrong about a lot of stuff. for example he said hopper was >1000w and that its die size is 1000mm2. he also says ad102 is 70-80B transistors at 600mm2 which also doesn't make sense.

I don’t remember him shotgunning like this for Ampere. He would say a lot less and ended up being right most of the time. With Ada he’s all over the place.
 
To me it seems a lot more plausible that kopite7kimi gets actual inside info on Nvidias product development and that the specs get changed during the process than that they're just outright making shit up.
There's a big difference between 'specs get changed' and 'there are a range of options that are still being decided on'. Saying the specs get changed would imply that decisions are getting made, and then they're changing their mind like every couple weeks. I'm no expert by any means, but that just sounds like it would be very weird to do.

And if nothing is actually being properly decided yet and things are still very fluid, then it makes no sense to report on this information as if things are getting decided, as kopite keeps doing. There's no qualifier that these are just 'options' for Nvidia to pick from, they seem to pretty directly state these are the specs of 'x' product. And this sets up pretty clear expectations. God forbid somebody isn't hanging on to their Twitter account's every word to know when these specs inevitably change again, ya know?
 
Guessing, but maybe “cut” in this case just means that L2 slice capacity on AD102 was designed to be smaller than in other configurations, e.g. to keep die size in check, and not that they are throwing away 25% of the capacity to increase yield.
 
AD102 is 384 bit so 20-24 GB for 4090
AD103 is 256 bit so 16 GB for 4080
AD104 is 192 bit so 12 GB for 4070
Thanks for doing the math lol.

The 16GB might be reason to upgrade from a 3080. Going to have to see price and how hard it is to get one. Also what I can sell my 3080 for
 
Status
Not open for further replies.
Back
Top