Discussion Ada/'Lovelace'? Next gen Nvidia gaming architecture speculation

[igor_kavinski]

All the strong competition from AMD will force their hand?

Good joke. You so funny!

 

[SmokSmog]

https://twitter.com/x/status/1684497812921147392

 

[psolord]

https://twitter.com/x/status/1684497812921147392

This could probably mean that everything will return to their normal tiers.

 

[SmokSmog]

This could probably mean that everything will return to their normal tiers.

Or that Jensen is greedy and 24Gb/s GDDR6X on 512bit bus is cheaper than 32Gb/s GDDR7 on 384bit bus.

 

[Heartbreaker]

Or that Jensen is greedy and 24Gb/s GDDR6X on 512bit bus is cheaper than 32Gb/s GDDR7 on 384bit bus.

Only expected in 2025, so I expect it will be GDDR7.

 

[SmokSmog]

Only expected in 2025, so I expect it will be GDDR7.

So 2TB/s then for 512 bit.

 

[jpiniero]

Sounds expensive.

 

[Heartbreaker]

They could run GDDR7 under max spec, IIRC both GDDR6 and GDDR6X can deliver 24 Gbp/s, but no one acutally runs that.

Either way it will be VERY expensive. 512 bit bus means 5090 will have 32 GB of VRAM.

 

[jpiniero]

They could run GDDR7 under max spec, IIRC both GDDR6 and GDDR6X can deliver 24 Gbp/s, but no one acutally runs that.

Either way it will be VERY expensive. 512 bit bus means 5090 will have 32 GB of VRAM.

I have a feeling that this would be more like a Titan.

 

[Aapje]

This could probably mean that everything will return to their normal tiers.

Unless they introduce a tier above the 5090, which wouldn't surprise me.

But I don't really see it as an option for Nvidia to keep selling 8 GB x060 cards. I think that they need to go back to 192 bit 12 GB for the 5060, and go to 256 bit 16 GB for the 5070.

 

[Heartbreaker]

I have a feeling that this would be more like a Titan.

Aren't the x090 cards just Titan renamed?

 

[Timorous]

Aren't the x090 cards just Titan renamed?

Kinda but without the drivers so they don't cannibalise the pro cards.

Also Titans were 255W.

 

[jpiniero]

Aren't the x090 cards just Titan renamed?

Yes and No but I think the price is not high enough for the Titan brand. I'm guessing this will be well over 2 grand.

 

[Saylick]

Wtf, why would they use a 512-bit bus... Just seems like complete overkill, unless they found a way to make an MCM design that is actually two 256-bit chips acting as one.

 

[igor_kavinski]

Wtf, why would they use a 512-bit bus

Maybe they are scared of RDNA4.

 

[Saylick]

Maybe they are scared of RDNA4.

Nahhh, no way.

 

[igor_kavinski]

Nahhh, no way.

AMD has probably ruined a few thousand sales of the 4080 with the 7900 XTX and this is with flawed execution. Imagine if they execute perfectly...

 

[Aapje]

Wtf, why would they use a 512-bit bus... Just seems like complete overkill, unless they found a way to make an MCM design that is actually two 256-bit chips acting as one.

So they can sell 64 GB clamshell cards for AI. The ever bigger AI models need ever more VRAM.

Keep in mind that the 102-chips are primarily designed for professionals. They won't even sell the full chip to gamers.

 

[Mopetar]

Wtf, why would they use a 512-bit bus... Just seems like complete overkill

Keep in mind that that NVidia isn't just using the die for gaming. Just because they give us a full-bus flagship doesn't mean that they won't have a more cut down version or two.

 

[jpiniero]

So they can sell 64 GB clamshell cards for AI. The ever bigger AI models need ever more VRAM.

Keep in mind that the 102-chips are primarily designed for professionals. They won't even sell the full chip to gamers.

I could see this being a *101 die.

 

[Saylick]

So they can sell 64 GB clamshell cards for AI. The ever bigger AI models need ever more VRAM.

Keep in mind that the 102-chips are primarily designed for professionals. They won't even sell the full chip to gamers.

That makes sense, but the Twitter leaker claims that the gaming flagship of Ada-Next uses a 512-bit bus... It's gotta be some kind of new Titan card that has 32 GB of VRAM and retails for >$2000 USD. No way in hell this thing will be built on TSMC N4P. It's gotta be N3E.

 

[Ajay]

Only expected in 2025, so I expect it will be GDDR7.

Hmm, completely bypassing 2024? Risky, unless AMD is behind on RDNA4 as well. Curious though, that we are still calling NV's next GPU Ada-next. The fact that there isn't a codename yet, could be telling - or it could just be that NV is being super tight lipped.

 

[Heartbreaker]

Hmm, completely bypassing 2024? Risky, unless AMD is behind on RDNA4 as well. Curious though, that we are still calling NV's next GPU Ada-next. The fact that there isn't a codename yet, could be telling - or it could just be that NV is being super tight lipped.

AMD still hasn't finished releasing RDNA3. How soon do you really expect to see RDNA4?

 

[CakeMonster]

I'm not so sure about the 2025 rumors either, its and awfully long time to go without offering new products, especially as the 4xxx series aren't selling that much. People do want to buy, they just are not too keen on current products and prices, but in 2024 they might be more willing to pull the trigger as they've been missing out on newer games.

 

[Heartbreaker]

I'm not so sure about the 2025 rumors either, its and awfully long time to go without offering new products, especially as the 4xxx series aren't selling that much. People do want to buy, they just are not too keen on current products and prices, but in 2024 they might be more willing to pull the trigger as they've been missing out on newer games.

Not all the 4000 series is selling poorly. 4090 and 4070 appear to be selling relatively well.

When Techepiphany was still report Mindfactory sales, 4070 was the best selling card from the week it launched until he stopped reporting results about 2 months later, and it made the steam survey in only 6 weeks.

For the slow sellers, there isn't much that price cuts wouldn't fix, so if they wanted to extend the life an extra 6 months, price cuts would get them there, unless the AMD RDNA4 launch is truly epic.

Though that might give AMD the first launching advantage that NVidia usually has.