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

[jpiniero]

AD102 entering mass production at TSMC

I'm skeptical. If they hadn't started production yet of AD102, we'd be looking at best the end of October for a paper launch. And unless they are convinced AMD is delaying N33 they need at least AD104 to launch not soon after AD102.

 

[Aapje]

It makes sense for Nvidia to start producing AD104 earlier anyway, to fill their contract obligations with TSMC, even while producing fewer AD102 and AD103 chips.

If I were Nvidia, I would delay the Ada launch as much as I can without being beaten to the punch by AMD, to clear the 3000 stock as much as possible, while already shipping chips to AID so they can build up stocks). Then release the new cards with a 2-4 week interval, to get that stock moving before pricing fall more (unless we are already deep in a 2nd hand glut, which is possible).

 

[DooKey]

AD102 entering mass production at TSMC

Who to believe? Igor said HVM isn't starting till August/September time frame.

What the leaks about NVIDIA’s new cards are good for, what is true and why kopite7kimi sometimes seems to be “wrong” | igor'sLAB (igorslab.de)

 

[igor_kavinski]

to clear the 3000 stock as much as possible

That would become very difficult if RDNA3 has better RT performance than the RTX series. I hope so. Would love a 16GB RX 7800 for $800.

 

[jpiniero]

Problem is, we don't know where the glut is, if it's everything or if it's just GA102. Inventory of GA102 basically needs to be zero by the time they announce the 4080.

 

[Aapje]

That would become very difficult if RDNA3 has better RT performance than the RTX series. I hope so. Would love a 16GB RX 7800 for $800.

That's one reason why I said that they should still aim to release before RDNA3.

 

[Frenetic Pony]

Who to believe? Igor said HVM isn't starting till August/September time frame.

What the leaks about NVIDIA’s new cards are good for, what is true and why kopite7kimi sometimes seems to be “wrong” | igor'sLAB (igorslab.de)

Just ignore the rumor mill IMO, it's ridiculous right now and so full of self contradictory information that no one can be right.

Also I don't get why Nvidia is trying to worm it's way out of the contract, as it doesn't look like it has any way to do so. TSMC is notorious for being a stickler to its contract rules, AMD looks like it's just barging ahead on its own schedule and could be really competitive with Nvidia, and a slow dump of mining cards looks like it's happening one way or another.

The only thing worse than paper launching on the high end before AMD if RDNA3 is good would be to wait it out. Imagine launching when your product isn't even press worthy anymore. If they really want to wait out any glut of mining cards they could stockpile dies in some clean warehouse and see if that ekes back some of their profit margin, but whatever way they slice it Nvidia looks like it might be in for a bad few quarters profitability wise.

 

[VirtualLarry]

but whatever way they slice it Nvidia looks like it might be in for a bad few quarters profitability wise.

Welcome to the boom / bust that is crypto, Nvidia! Let's hope that you didn't (allegedly) mislead your shareholders about your crypto involvement for a third time...

 

[igor_kavinski]

Welcome to the boom / bust that is crypto, Nvidia! Let's hope that you didn't (allegedly) mislead your shareholders about your crypto involvement for a third time...

Yeah, they made their bed and now they must lie in it. Jensen, you bad boy, you gonna get a spanking now.

 

[jpiniero]

but whatever way they slice it Nvidia looks like it might be in for a bad few quarters profitability wise.

How much of the "Gaming" revenue do you think the DIY sales are? You would think that OEM sales wouldn't be affected by any mining flood, although they could be cutting orders too.

One thing they could do is stick Ampere chips in a warehouse and gamble that mining returns at some point.

 

[Ajay]

One thing they could do is stick Ampere chips in a warehouse and gamble that mining returns at some point.

Till they have to explain the carrying charge (or whatever) in the next quarter's report. Investors will be thrilled - when they win their lawsuit.

 

[KompuKare]

One thing they could do is stick Ampere chips in a warehouse and gamble that mining returns at some point.

Aside from having to point out their inventory and stock on hands in their financial reports as Ajay pointed out, this would only possibility work if mining profitability goes crazy again AND no newer cards are actually much good at mining. With the heavy use of cache from both sides, the latter is possible, but it's still all very risky.
And those miners who calculate with being able to sell their cards on eventually probably wouldn't be that keen on last gen cards anyhow.

 

[coercitiv]

One thing they could do is stick Ampere chips in a warehouse and gamble that mining returns at some point.

If rumors are true, imagine the irony of realizing that catering to miners for years has made them part of your business cycle. Good luck convincing gamers, during a recession, that they should upgrade to the latest and greatest after you left them squirming with bottom of the barrel cards for multiple seasons.

It's such a pity Intel screwed the pooch with their entry in the dGPU market, it would have been glorious to have them arrive early this summer and inject even more affordable cards before the end of the year.

 

[jpiniero]

Till they have to explain the carrying charge (or whatever) in the next quarter's report. Investors will be thrilled - when they win their lawsuit.

OK scratch that. I will say that I almost wonder if the issue isn't current inventory, but they also made purchase commitments to Samsung as well.

Aside from having to point out their inventory and stock on hands in their financial reports as Ajay pointed out, this would only possibility work if mining profitability goes crazy again AND no newer cards are actually much good at mining. With the heavy use of cache from both sides, the latter is possible, but it's still all very risky.

Doesn't look like infinity cache helped RDNA 2's mining performance at all. It's looking like the 4080 and 4070 are going to have less memory bandwidth than the 3080 and 3070. So mining performance shouldn't be any better and could even be worse.

 

[CakeMonster]

It's looking like the 4080 and 4070 are going to have less memory bandwidth than the 3080 and 3070. So mining performance shouldn't be any better and could even be worse.

Can't believe I'm cheering for this. I'm sure there's some gaming and productivity scenario that will suffer and bite ourselves and NVidia in the ass later on though.

 

[jpiniero]

Can't believe I'm cheering for this. I'm sure there's some gaming and productivity scenario that will suffer and bite ourselves and NVidia in the ass later on though.

It still shouldn't be bad at mining. 4080 has more memory bandwidth than the 6900 XT for instance.

 

[Leeea]

It still shouldn't be bad at mining. 4080 has more memory bandwidth than the 6900 XT for instance.

A 6900xt only gets 60ish MH/s. That does not mean to much.

 

[Mopetar]

Infinity Cache isn't useful for any mining algorithm (e.g., ETH) that's designed to be memory bound. The memory access pattern is essentially random so the cache is just getting thrashed. A wider bus and faster memory will always perform better.

Personally, I'm fine with AMD cards being less optimal for mining. It puts a price cap on how much the actual retail price can inflate due to miners wanting the card.

 

[moonbogg]

If reduced memory bandwidth means poor mining performance, then as far as I'm concerned they can just remove the damn memory from the graphics card completely. I won't be buying another price hiked card anyway, so not like I care.

 

[igor_kavinski]

they can just remove the damn memory from the graphics card completely.

With native PCIe 5.0 cards, they might do just that for the lower end SKUs.

 

[moonbogg]

With native PCIe 5.0 cards, they might do just that for the lower end SKUs.

Good. Screw video cards completely then. Bring on the integrated stuff with fast DDR5. I'd get a good laugh watching miners try to fill their garage with motherboards and a single SOC onboard. "I make 50 cents per day everyday off my $10,000 investment. Ask me how."

 

[igor_kavinski]

I'd get a good laugh watching miners try to fill their garage with motherboards and a single SOC onboard.

We would get a new crisis: scalping and rarity of mini-ITX mobos and possibly NUCs.

 

[Tup3x]

With native PCIe 5.0 cards, they might do just that for the lower end SKUs.

TurboCache 2.0.

 

[biostud]

And a performance "leak"

NVIDIA GeForce RTX 4090 is reportedly 66% faster than RTX 3090 Ti in 3DMark Time Spy Extreme test - VideoCardz.com

NVIDIA GeForce RTX 4090 reportedly 66% faster than RTX 3090 Ti, 82% faster than RTX 3090 Kopite7kimi claims to have the first benchmark result from the upcoming NVIDIA RTX 4090 GPU. Over at Twitter, NVIDIA hardware leaker Kopite7kimi shared first benchmark result of flagship Ada GeForce RTX 4090...

videocardz.com

 

[jpiniero]

And a performance "leak"

NVIDIA GeForce RTX 4090 is reportedly 66% faster than RTX 3090 Ti in 3DMark Time Spy Extreme test - VideoCardz.com

NVIDIA GeForce RTX 4090 reportedly 66% faster than RTX 3090 Ti, 82% faster than RTX 3090 Kopite7kimi claims to have the first benchmark result from the upcoming NVIDIA RTX 4090 GPU. Over at Twitter, NVIDIA hardware leaker Kopite7kimi shared first benchmark result of flagship Ada GeForce RTX 4090...

videocardz.com

If the 4090 is only 66% faster than the 3090, yeah, that might be a reason to cut orders. Because RDNA 3 would be much faster.