Discussion Nvidia Blackwell in Q4-2024 ?

[Aapje]

I suppose by the time the refresh comes out, 3 GB chips will be available so perhaps they will use that.

Yeah, I was thinking the same thing. Nvidia may intend to do a big refresh anyway, since RDNA6 is supposed to come out at the end of 2025 already.

Kopite7kimi clearly refer to GB203 onward cause we all know GB202 going to use GDDR7. If people still not understand the words "memory interface configuration" of AD10x. Well, wait for more leaks...

I'm not convinced that he was referring to the type of memory with his comment, rather than the bus size and memory amounts.

 

[Tigerick]

Yeah, I was thinking the same thing. Nvidia may intend to do a big refresh anyway, since RDNA6 is supposed to come out at the end of 2025 already.



I'm not convinced that he was referring to the type of memory with his comment, rather than the bus size and memory amounts.

I am totally believe he meant GDDRX6 memory. And why don't you create a list of specs of your so called GDDR7 lineup with your clamshell guarantee... I am waiting

 

[KompuKare]

Don't be mad, people. Just buy RDNA4. And "Blackwell" sounds ominous.

Now that's what mindshare does: a company is able to do things which even their fans don't like but they buy it anyways!

 

[Aapje]

I am totally believe he meant GDDRX6 memory. And why don't you create a list of specs of your so called GDDR7 lineup with your clamshell guarantee... I am waiting

Surely not if you are going to troll me.

 

[Tigerick]

Surely not if you are going to troll me.

See, you can't even try to list and say I am going to troll you. Clearly you are not confident with what you said or guarantee...

 

[Aapje]

I am just speculating. How could I ever guarantee anything? I'm not Jensen.

 

[jpiniero]

I looked at it again, and I will predict that the faster GB203 is maybe comparable in raster to the 4090 in 4K and 10-15% faster at lower resolutions (?) With the slower being 15-20% less than that.

The faster GB205 being (of course) barely faster (if even that) than the 4070 Ti.

RT being another story perhaps but need more info.

 

[Aapje]

RGT put out an interesting video where he claims/speculates that GB207 and GB206 will be made on Samsung SF4X and the rest on TSMC 3N.

That does make a lot of sense since TSMC prices of the new nodes are so high and the high efficiency is not really needed for the lower end cards.

He claims a 96 bit bus for GB207, which could be good news if that chip will then only go into the 3050 like with Ampere, but probably not so good news if it doesn't. Although it seems that a clamshell is pretty much guaranteed if it goes into the 5060.

For the other chips the claimed bus size is the same as for Ada, so you'd see the same RAM issues/dilemma's.

 

[jpiniero]

RGT put out an interesting video where he claims/speculates that GB207 and GB206 will be made on Samsung SF4X and the rest on TSMC 3N.

That does make a lot of sense since TSMC prices of the new nodes are so high and the high efficiency is not really needed for the lower end cards.

I ain't gonna watch the video cuz RGT... but while I think it's possible they do that, max frequencies would be decently less and mobile efficiency would be worse too. The performance gap between them and the N3 products would be even wider, desktop and laptop.

I'd also say they would also do GB205 there as well.

 

[SmokSmog]

Back to 512bit

 

[Tigerick]

Actually, last time Chiphell already leaked about GB202 is having 512-bit memory bus with GDDR6X and 128MB L2 cache. NV has always been supporting GDDR6X and GDDR7 in the same die. Therefore, when GDDR7 is available, NV has switched from testing 512-bit GDDR6X to 384-bit GDDR7. That's why I reduced the amount of L2 cache from 128MB to 96MB.

Regarding 28Gbps, seems like NV has reduced the clock for GB202. Samsung has stated they are shipping 32Gbps GDDR7. The total memory BW has been reduced to 1.34 TB/s with 84/96 MB L2 cache. Clearly NV feel like 28GBbps is fast enough. Of course, NV might be reserving 32Gbps for Ti model.

I have updated my comparison table in the front page regarding memory bandwidth. Almost all of my higher bus GDDR6X solutions are faster than GDDR7, clearly NV knows that. Yeah, Blackwell series only has flagship models, i.e. RTX-5090 is using GDDR7, the rest of lineup are using GDDR6X/GDDR6 memory.

 

[Aapje]

Regarding 28Gbps, seems like NV has reduced the clock for GB202. Samsung has stated they are shipping 32Gbps GDDR7.

Just because they are shipping some 32Gbps GDDR7 doesn't mean that they are shipping it in sufficient volume for Nvidia. The higher the speeds, the lower the yields.

Anyway, if GB202 is 512 bit, that would suggest we might see a 384 bit GB203, 256 bit GB205 and a 192 bit GB206. Plus a 128 bit GB207. If so, that would fix most or all of their VRAM issues, with 24 GB in the 5080, 16 GB in the 5070. 12 GB in the 5060 Ti and perhaps also 5060. And then 8 GB in the 5050.

Unless they put the GB202 only in the Titan, 5090 Ti and/or pro cards, of course, and put GB203 into the 5090.

 

[Tigerick]

Just because they are shipping some 32Gbps GDDR7 doesn't mean that they are shipping it in sufficient volume for Nvidia. The higher the speeds, the lower the yields.

Anyway, if GB202 is 512 bit, that would suggest we might see a 384 bit GB203, 256 bit GB205 and a 192 bit GB206. Plus a 128 bit GB207. If so, that would fix most or all of their VRAM issues, with 24 GB in the 5080, 16 GB in the 5070. 12 GB in the 5060 Ti and perhaps also 5060. And then 8 GB in the 5050.

Unless they put the GB202 only in the Titan, 5090 Ti and/or pro cards, of course, and put GB203 into the 5090.

So you finally agree that the rest of lineup are using GDDR6X?

 

[jpiniero]

So you finally agree that the rest of lineup are using GDDR6X?

GB203 still very likely uses GDDR7.

 

[SmokSmog]

I think GB202 will have 512bit because of Quadro lineup, AI customers want a lot of vram.
512bit bus + 3GB GDDR7 ICs would double the vram from 48GB on 384bit AD102 Quadro to 96GB on 512bit GB202.

 

[Aapje]

So you finally agree that the rest of lineup are using GDDR6X?

Why do you think that this follows from my comment? With only 2 GB modules for GDDR7 for now, the memory sizes available relative to the bus size are identical.

 

[Tigerick]

Why do you think that this follows from my comment? With only 2 GB modules for GDDR7 for now, the memory sizes available relative to the bus size are identical.

Except GB202 is using 384-bit GDDR7, therefore GB203 will be using 256-bit if using GDDR7, do you understand?

Please read carefully...gods

Clearly you do not understand 512-bit GDDR6X and 384-bit GDDR7 are sharing same memory bandwidth. That's why GB202 can support both memories with one die. if GB202 choose to use GDDR7 solution, then GB203 have to use lesser bus if they use GDDR7...

 

[jpiniero]

The GB202 die could be 512-bit but the gamer GB202 cards could be cut down to 384.

384 bit @ 28 is only 33% more bandwidth than the 4090 though.

 

[Aapje]

Clearly you do not understand 512-bit GDDR6X and 384-bit GDDR7 are sharing same memory bandwidth. That's why GB202 can support both memories with one die. if GB202 choose to use GDDR7 solution, then GB203 have to use lesser bus if they use GDDR7...

I think that you fundamentally misunderstand how GDDR works.

2 GB modules have a 32 bit interface, regardless of whether it is GDDR6(X) or GDDR7. The main thing that is needed for the same chip to support GDDR6X and GDDR7 is to have the right encoders. So the chip then needs to be able to do PAM 3 and PAM 4.

We already know that GDDR7 will support both PAM3 and NRZ for efficiency. It switches over to NRZ is less bandwidth is needed. So it should be pretty easy to then support GDDR6, since NRZ-signalling has to be supported anyway.

The memory bandwidth has nothing to do with any of this.

 

[jpiniero]

I'd have to think it still supports GDDR6X, if only so it can use excess memory chip inventory if needed.

 

[Aapje]

I actually think that the 512 bit confusion is because GB102 will be 512 bit, but that GB202 will still be 384 bit.

The rumor is that Nvidia is splitting up the gaming and pro-chips, a bit similar to how AMD has RDNA for gaming and CDNA for compute. Although AMD limits CDNA to their data center stuff, but GB10x would then be aimed more at everything from AI workstations to cheaper data center stuff.

So then GB10x would be MCM, perhaps with a separate IO-chip like AMD uses. It would also probably be fully GDDR7, going up to 64 GB for the GB102 clamshell cards. HBM would still only go into the highend B100/B200/etc MCM chips.

GB20x would then still be monolithic, but excising the professional stuff from the chips, that is on the chip, but mostly disabled. Of course, ideally Nvidia would make it so that these cards don't work that well for AI, so people would be forced to the more expensive professional cards for that.

 

[jpiniero]

NVIDIA Blackwell "GB203" GPU to feature 256-bit bus & GB205 with 192-bit, claims leaker - VideoCardz.com

GeForce RTX 50: up to 512-bit memory bus, no 384-bit option? NVIDIA might be skipping the 384-bit memory bus in their Blackwell gaming lineup. Kopite7kimi has entered the conversation about next-gen NVIDIA gaming GPUs, conversations that were often sparked out of his rumors. This respected...

videocardz.com

Seems to think that GB202 is 512-bit, GB203 is 256 and GB205 is 192.

Assuming it only has 64 MB L2 (or less), I don't think GB203 will be able to reach 4090 performance until faster GDDR7 comes along.

I actually think that the 512 bit confusion is because GB102 will be 512 bit, but that GB202 will still be 384 bit.

I kind of think GB1xx other than GB100 was cancelled/blown up to make it more AI.

 

[SteinFG]

Assuming it only has 64 MB L2 (or less), I don't think GB203 will be able to reach 4090 performance until faster GDDR7 comes along.

It's not all about cache. Newer architecture will also help. GB203 only needs to be 30% faster than AD103 to beat 4090.

 

[SteinFG]

Interesting how there's now GB205. Is it suppoed to be faster than AD104? Going by the name, it probably has lower number of cuda cores. then 5070 will be kinda meh, and most likely still 12GB.

 

[MrTeal]

512 bit would be a big change of pace for Nvidia, as far as I know they haven't ran a bus that large since GT200/GTX280 16 years ago. 50% more SM should scale well with N3E, but that memory controller really won't. Even by x02 standards, that chip as rumored is going to have a huge and wildly expensive die.