Discussion Nvidia Blackwell in Q4-2024 ?

[TESKATLIPOKA]

More than an equal number of 2GB modules, but less than double the amount of 2GB modules needed for a clamshell design. So I expect we will see 3GB memory chips before we see Clamshell models for next gen lower end cards.

That depends how far away the 3GB modules are.

 

[Ajay]

I'd have to find the roadmap but I think 4 GB is some ways off. I am hoping that the Blackwell Refresh cards will come with 3 GB but I don't think that's 100% given.

I mixed up 4GB with 3GB.

 

[jpiniero]

There is a chance I guess that NV does use 32 gbps on some cards rather than 28, which is also sampling.

It'd be interesting what they do on mobile. While I think those are the two options available they could underclock to 26 or 24 perhaps to save power.

 

[Aapje]

I haven't seen 4GB. But I have seen 3GB on all three roadmaps (Micron, SKH, Samsung). They may not be in the first cards, but those will be the high end cards like 5090/5080. They may be in time for lower end 128 bit cards where they are most needed.

I don't think that the 3 GB-chips are anywhere close enough. The memory manufacturers would have to be sampling them now if that was the case and we would have heard about it.

I think that the best case is that we get the 3 GB chips in the 2025/2026 refresh. It seems to me that this would be great for Nvidia, as they can first sell cards to those who are willing to accept too little VRAM and then get a big boost in sales for the refresh, to customers who were waiting for more VRAM.

 

[igor_kavinski]

It seems to me that this would be great for Nvidia, as they can first sell cards to those who are willing to accept too little VRAM

Their diehard loyalists buy their cards twice. Once with little VRAM and then the higher VRAM refresh. It allays their FOMO and the resale value of the older card is good enough for them to keep doing this every generation. Except psolord. He tames the settings to fit his use cases. Super wise guy.

 

[psolord]

Their diehard loyalists buy their cards twice. Once with little VRAM and then the higher VRAM refresh. It allays their FOMO and the resale value of the older card is good enough for them to keep doing this every generation. Except psolord. He tames the settings to fit his use cases. Super wise guy.

Right? xD

Only an idiot buys twice. Smart guys buy once at double the price! xD

 

[DeathReborn]

GDDR6 was specced for 3 GB chips as well, yet that never came.

The 3 GB chips are on the roadmaps for GDDR7 now, but so far there is no sign of any actual chips.

Technically GDDR6(X) was up to 32Gbit (4GB) & GDDR7 is up to 64Gbit (8GB). I don't expect to see 64Gbit GDDR7+ unless there's a demand from AI for it & probably not on consumer devices until the mid 2030's.

JEDEC Publishes GDDR7 Memory Spec: Next-Gen Graphics Memory Adds Faster PAM3 Signaling & On-Die ECC

www.anandtech.com

 

[Heartbreaker]

Technically GDDR6(X) was up to 32Gbit (4GB) & GDDR7 is up to 64Gbit (8GB). I don't expect to see 64Gbit GDDR7+ unless there's a demand from AI for it & probably not on consumer devices until the mid 2030's.

JEDEC Publishes GDDR7 Memory Spec: Next-Gen Graphics Memory Adds Faster PAM3 Signaling & On-Die ECC

www.anandtech.com

There is a difference between JEDEC specs and company roadmaps. I don't remember anyone having GDDR6 3GB chips on their roadmap. All three RAM makers have said they are making 3GB GDDR7 chips, and it's showing up on roadmaps:

So I expect we will get 3GB chips this time.

Also it can be over 6 months between the release of the highest end cards and the low end. So it's very possible that 3GB chips could show up in cards Q3 of 2025.

 

[Ajay]

A bit larger pic than @Heartbreaker . So, 24Gb look to be 2H25-1H2026. No idea yet when GPU makers will use them in their higher end cards.

 

[Heartbreaker]

A bit larger pic than @Heartbreaker . So, 24Gb look to be 2H25-1H2026. No idea yet when GPU makers will use them in their higher end cards.

View attachment 100885

It's 16-24Gb in the first batch starting late 2024.
It's 24Gb+ (note the +), starting early 2026.

 

[Ajay]

It's 16-24Gb in the first batch starting late 2024.
It's 24Gb+ (note the +), starting early 2026.

Okay, thanks. I see it now. Old freaking eyes!

 

[SolidQ]

So 5070 again 12GB

 

[psolord]

Who gives an F about the 5070? What about the 5060/ti? Will these be 8GBs again?

OMG I can't wait to see that idiot in the 8GB thread, trying to defend them!

xD

 

[igor_kavinski]

OMG I can't wait to see that idiot in the 8GB thread, trying to defend them!

Maybe this time Geforce will be able to upscale 720p textures to 4K, without noticeable degradation. That could solve the issue. IF they can do it.

 

[Aapje]

What about the 5060/ti? Will these be 8GBs again?

The only chance for anything better is if they put the GB207 in the 5050, the GB206 in the 5060 and GB205 in the 5070 and 5060 Ti.

Essentially, the same as with Ampere. I don't see them doing this though, as they seem wedded to sellling the cards as one tier higher than they should be.

 

[jpiniero]

So that's 17% less SMs for GB205 and GB207 but the same for GB206. Kinda strange to cut the other two but leave GB206 alone.

Also probably means no 5070 Ti, at least to start.

So something like this:

5090 - 144 SMs
5080 - 80 SMs
5070 - 48 SMs
5060 Ti - 34 SMs
5060 - 18 or 20 SMs (or doesn't exist)

 

[TESKATLIPOKA]

So GB206 will have 3GPCs * 6 SMs compared to AD106 with 3GPCs * 12SMs?
Or did he actually mean TPCs instead of SM? I don't think so.

Let's say GB*** SM has 2x more Cudas, then GB207 has 2560 Cudas while AD107 had 3072, so -16.7% less.
Either IPC or frequency has to increase to compensate.

Good news is that GB206 will no longer be crippled by BW and config looks the "same" as AD106.
Another interesting thing is the different amount of SMs per GPC, Ada had 12 per GPC for every die.

 

[jpiniero]

Let's say GB*** SM has 2x more Cudas, then GB207 has 2560 Cudas while AD107 had 3072, so -16.7% less.
Either IPC or frequency has to increase to compensate.

Yeah that's what I am thinking. Small IPC increase and 10-15% higher clocks.

 

[Ajay]

3DCenter made this chart:



Pretty lame IMHO. GB202 will still slay. But the GB203 is an awful upgrade (from the 4080). Of course, if Nvidia can pull a rabbit out of it's hat and get better IPC and Clocks, maybe it won't be so bad.

 

[TESKATLIPOKA]

3DCenter made this chart:

View attachment 100964

Pretty lame IMHO. GB202 will still slay. But the GB203 is an awful upgrade (from the 4080). Of course, if Nvidia can pull a rabbit out of it's hat and get better IPC and Clocks, maybe it won't be so bad.

GB206 looks pretty good with that 18->28gbps GDDR7, that's 55% more. There is high possibility for 16GB Vram, but It won't be cheap.

 

[Golgatha]

3DCenter made this chart:

View attachment 100964

Pretty lame IMHO. GB202 will still slay. But the GB203 is an awful upgrade (from the 4080). Of course, if Nvidia can pull a rabbit out of it's hat and get better IPC and Clocks, maybe it won't be so bad.

Being that the 2nd tier is still 256bit and nVidia can best the 4090 with a 448bit 5090, I think they'll go that route. This would give them more 512bit cards for the AI market in terms of yields. Also, there's no competition at the high end, so they can safely get away with releasing something marginally faster at $1500+. Given the state of horrid pricing for graphics cards for gaming being something that seems to be here to stay, I would wager at this point my 4090 will be my next 1080 Ti in terms of spending 5 years or more in at least one of my gaming rigs.

 

[Ajay]

GB206 looks pretty good with that 18->28gbps GDDR7, that's 55% more. There is high possibility for 16GB Vram, but It won't be cheap.

Well, not awesome if only 10-20% faster.

I generally buy xx70 or xx80 cards, so will be looking for something quite a bit better than my 4070Ti before doshing out any more money.

I'll probably be buying an RTX 5000 series card when the 6000 series comes out - if prices drop enough. Similar to what I want to do with Zen5. Or I could skip a generation. I'm guessing my next video card for gaming may be my last (well, for gaming).

 

[TESKATLIPOKA]

Well, not awesome if only 10-20% faster.

You don't need >50% higher BW for that.

@gdansk and what is the speed of the worst GDDR7? Even if it's only 24gbps, that's still 33% more than what AD106 has.

 

[gdansk]

You don't need >50% higher BW for that.

They'll put the worst GDDR7 chips somewhere

 

[jpiniero]

@gdansk and what is the speed of the worst GDDR7? Even if it's only 24gbps, that's still 33% more than what AD106 has.

Right now what is sampling is 28 and 32. They could underclock though.