NVidia Ada Speculation, Rumours and Discussion

[dorf]

If the constellation is such, that only with a heavily overclocked Hopper they can claim perf kingship in the desktop and/or gaming space, they are likely to divert a couple of hundred cards, make a lot of marketing fuss around it and then say "sorry, demand is far greater than we estimated. We will be working with our partners to improve supply."

I don't think they are this stupid. There's no way Nvidia and AMD haven't learned with current gen how disappointed people are about poor availability. Whenever Geforce social media accounts post anything, they are spammed with "where can I buy one?", "prices are ridiculous!" etc. This is not a good look for them.

The vast majority of tech press would also scold Nvidia or AMD for this kind of a silly performance-crown-at-any-cost move.

 

[Jawed]

No such thing is possible.

AMD was clearly happy to stick at "about 20%" market share in desktop graphics and their wafer purchase contracts will probably reflect its timidity for a while.

I don't see any reason for NVidia to not go after as much 5nm wafer capacity as it can muster.

https://twitter.com/i/web/status/1257685298118451200

May 2020.

 

[trinibwoy]

If NVidia floods the market with cards that have the same performance as 3070 and 3080 in Q1 2022, by launching Lovelace "104" first, does that count as a major shift in strategy?

Not if we're talking about "winning". Flooding the market with attractive price/performance parts while ceding the crown is of course a change in overall strategy but it's also admitting defeat. I thought the conversation so far was about how Nvidia was going to fight to hold on to that crown.

NVidia would be pretty stupid to have assumed that AMD wasn't going to use chiplets, as for the past five years it's been a hot topic.

Exactly, which is why Hopper as a backup plan makes no sense. They would have already targeted Lovelace at whatever chiplet based perfomance they expect to go up against.

RDNA 3 rumours indicate some cards are chiplet based and some aren't. Hopper and Lovelace might form a similar pairing.

Sure but does that require a whole separate architecture? I would expect the lower end AMD cards to also be considered RDNA3.

 

[Deleted member 13524]

But I give no credence to the "Hopper is NVidia's plan B for desktop (top tier prosumer or consumer)" theory/rumour. NVidia will be selling every last Hopper to data centre people who are quite happy to pay 10x more than consumers.

The beauty about halo products is that you don't really need to sell them to consumers in volume. You just need to get a few dozens in reviewers' hands, and the large bar will always appear in all the charts.

 

[Bondrewd]

I don't see any reason for NVidia to not go after as much 5nm wafer capacity as it can muster.

Already booked to hell, also H100 alone is Q2'22.
Every other nV product is ~later~.

Nvidia was going to fight to hold on to that crown.

lol wat

The beauty about halo products is that you don't really need to sell them to consumers in volume

Idk chief people seethe about 3080/6800XT availability all the same.

 

[Jawed]

I thought the conversation so far was about how Nvidia was going to fight to hold on to that crown.

A conversation captured by a stupid idea. I have no beef with the discussion of stupid ideas as long as I don't feel stupid afterwards. Speculation threads shouldn't reject discussion points out of hand even if they seem really stupid to some.

Exactly, which is why Hopper as a backup plan makes no sense. They would have already targeted Lovelace at whatever chiplet based perfomance they expect to go up against.

Yep.

Sure but does that require a whole separate architecture? I would expect the lower end AMD cards to also be considered RDNA3.

I think you might say that NVidia is being cautious with chiplets, versus AMD.

This presumes that Lovelace has no chiplets!

Who wants to start the "GL102 is chiplet-based" rumour?

 

[trinibwoy]

The beauty about halo products is that you don't really need to sell them to consumers in volume. You just need to get a few dozens in reviewers' hands, and the large bar will always appear in all the charts.

AMD is going all in on CDNA. There’s no way Nvidia’s next datacenter chip has all of the graphics bits needed for gaming. They are feeling the heat from HTPC/AI focused hardware on multiple fronts.

 

[Jawed]

AMD is going all in on CDNA. There’s no way Nvidia’s next datacenter chip has all of the graphics bits needed for gaming. They are feeling the heat from HTPC/AI focused hardware on multiple fronts.

Does anyone have any idea why GA100 has TMUs? Does CUDA depend upon their presence?

Does the die space actually matter? If the power is switched off for 99.99999999999% of the time, do they matter?

 

[Bondrewd]

Does the die space actually matter?

Actually yea.
Literally can only stick so much onto a 3x reticle CoWoS piece.
Like a pair of ~700mm^2 + 8 HBM2e tops.

 

[Jawed]

Blimey, it's also got ROPs:

NVIDIA GA100 GPU Specs | TechPowerUp GPU Database

but they do global memory atomics, don't they?

 

[Deleted member 13524]

Not if we're talking about "winning". Flooding the market with attractive price/performance parts while ceding the crown is of course a change in overall strategy but it's also admitting defeat. I thought the conversation so far was about how Nvidia was going to fight to hold on to that crown.

Moreover, I suspect that during 2022 the price/performance crown will be coming from all the Ampere GPUs that are being withheld in China for cryptomining and will flood the 2nd hand market once ETH goes proof-of-stake.
If they thought 2018 was bad, 2022 will probably be 3x worse, with the post-covid recession the mining craze lasting for much longer.

AMD is going all in on CDNA. There’s no way Nvidia’s next datacenter chip has all of the graphics bits needed for gaming. They are feeling the heat from HTPC/AI focused hardware on multiple fronts.

Yes, I agree. I was just clarifying what the rumors meant.

 

[techuse]

Agreed. If both IHV's are on the same node Nvidia definitely has the performance/efficiency advantage.

How do you know this?

 

[xpea]

It doesn’t make sense that Hopper and Lovelace will be ready at the same time with the same graphics feature set. If that’s the case there’s no reason for Lovelace to exist.

Well the thing is that Lovelace is late and Hopper in on time. Nvidia is working all the time on 3 generations + a large team doing fundamental research for future techs.
This time, the situation is a bit special with the transition to chiplets and overlapping schedule / market...

 

[trinibwoy]

Blimey, it's also got ROPs:

NVIDIA GA100 GPU Specs | TechPowerUp GPU Database

but they do global memory atomics, don't they?

A100 also has GPCs and TPCs which presumably means rasterizers, tessellation and all the other geometry setup hardware. The only thing that was clearly dropped is RT acceleration.

 

[PSman1700]

AMD has to redesign their GPUs to provide better support for Raytracing and DL. Yet people here believe that this huge shift in graphics doesnt exist for AMD. nVidia has changed the game like Apple with the iPhone.
Ampere delivers 50% more compute, twice the raytracing and 4x performance for DL.

nVidia is far ahead and they will continue to push towards raytracing.

Once so far behind its quite hard to compete i think, atleast in the high performance gpu market. Its the PS5 that missed tbe boat.....

Sure, when you said it. I mean you are this guy who claimed that RDNA2 would be way more efficient than Ampere. What happened to this?

Same guy who has gotten banned before.

Anyway, theres more AMD talk than NV here

 

[CarstenS]

I don't think they are this stupid. There's no way Nvidia and AMD haven't learned with current gen how disappointed people are about poor availability. Whenever Geforce social media accounts post anything, they are spammed with "where can I buy one?", "prices are ridiculous!" etc. This is not a good look for them.

The vast majority of tech press would also scold Nvidia or AMD for this kind of a silly performance-crown-at-any-cost move.

Maybe, but maybe if they can seed a couple of YTers and have a co-op Video accompanying the launch, there'd be enough mindshare. Current situation teaches: Even with normal planning, you're not safe from bad availability.

H100 has no display outs or RT gear.
Impossible.

Possibly, but then, there might be contingencies.

Idk chief people seethe about 3080/6800XT availability all the same.

Which are clearly not halo products.

 

[CarstenS]

Blimey, it's also got ROPs:

NVIDIA GA100 GPU Specs | TechPowerUp GPU Database

but they do global memory atomics, don't they?

It's also got display controllers and rasterizers. It's fully graphics capable, just lacking RT-Cores.

 

[Bondrewd]

It's also got display controllers

Has none of that iirc.
Even the A100 Station uses a discrete 5th dGP to output displays.

 

[DegustatoR]

Possibly, but then, there might be contingencies.

The reason why this idea smells fishy isn't because GH100 is a compute chip - the rumor specifically mention GH202 for a reason.
The reason is why even have Lovelace if you can have Hopper in its place if it's so good that it would be a better fit competitive wise?

 

[CarstenS]

Has none of that iirc.
Even the A100 Station uses a discrete 5th dGP to output displays.

Talking about the GPU, not the card. The card has no display outputs, that's so obvious, I thought I didn't need to mention it.