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

[Timmah!]

I am seriously doubting most of the rumors regarding clocks and TDPs.

My FTW3 3090 already has a large, unwieldily cooler. To add a 50% larger heat sink on top of that? To say nothing of lost slots, a GPU support bracket would be absolutely required and not every case supports the use of said brackets.

A 900W TDP? No way.

What I expect we will see is 2-2.2 Ghz tops unless NVIDIA has made huge efficiency gains. The possible mhz increase from the shrink from Samsung 8n to TSMC N5 will be offset by a much larger die.

If NVIDIA has made breakthroughs in power management or cooling, then maybe, but something tells me these GPUs aren’t going to be as fast as people think. The top tier may be 40-60% faster than a 3090. That is my guess.

I dont know about the frequency, but if AMD can allegedly run their RDNA3 at 3GHz, then its possible Nvidia can go past 2,5GHz as well, without melting the chip or needing some exotic cooling.
Alternatively, i think its possible they might stick to those lower 2GHz+ clocks, if they had those 18300+ FP32 cores, but all of these would be pure FP32, and then on top of that, another 9000 INT cores capable of functioning as FP32 as well, same way like in case of Ampere. So potentially it could be 27000+ FP32 and then 27k x2 x2 = 108TFlops.
Or not

 

[eek2121]

I dont know about the frequency, but if AMD can allegedly run their RDNA3 at 3GHz, then its possible Nvidia can go past 2,5GHz as well, without melting the chip or needing some exotic cooling.
Alternatively, i think its possible they might stick to those lower 2GHz+ clocks, if they had those 18300+ FP32 cores, but all of these would be pure FP32, and then on top of that, another 9000 INT cores capable of functioning as FP32 as well, same way like in case of Ampere. So potentially it could be 27000+ FP32 and then 27k x2 x2 = 108TFlops.
Or not

AMD's chips are smaller and were built on a superior process. Next-gen chips from both will be on the same process, however AMD is once again innovating by utilizing a multi-chip approach while sticking with GDDR6. NVIDIA is building a very large monolithic chip while utilizing insanely hot GDDR6X memory. I would be shocked if NVIDIA managed 2.5 GHz with the rumored die sizes. I'm honestly expecting around 2 GHz, give or take 200 mhz.

 

[Timmah!]

https://www.reddit.com/r/nvidia/comments/uivnok

Any idea what this means?

I look at those diagrams and i see that Lovelace:

- has only those FP32/INT cores, no pure FP32 anymore
- each 16 of them has those load-store units, instead of 16+16 before. I suppose that part about "doubling the subcore" refers to that?

i still dont understand what he means by 4*FP32 expansion space

 

[Saylick]

I think Kopite is hanging his hat on this layout instead, given that this is a pinned Tweet:

https://twitter.com/x/status/1522230690880061440

https://twitter.com/x/status/1522253825217908738

 

[Mopetar]

I like how he's covering for his initial mistake. I suspect all along Ada Lovelace was a new architecture. He may have been fed bad info. Take what he says with a grain of salt.

I agree that this is probably just to cover up a mistake on his part. It's very plausible that Nvidia also worked on a refresh, like AMD is releasing with the 6x50 cards, but that they didn't deem it necessary.

There are limits to what NVidia can change in the span of a year. Radical changes to the architecture aren't going to happen I. That span of time and even the downstream changes are limited by earlier decisions.

If changes were made it would have been when NVidia realized that AMD was competitive again, but once the design is taped out there's not a lot of room outside of clocks, memory configuration, and price to change a product.

It's doubtful that NVidia would have started out with a code name that was intended for a refresh or what might be a minor increment on the design of Ampere and then transition that to be a radically different product. Why not just scrap it and have a different code name? The only situation I can think of is someone getting fed multiple rumors and combining them such that what was said to be Lovelace was something else.

 

[Aapje]

It's doubtful that NVidia would have started out with a code name that was intended for a refresh or what might be a minor increment on the design of Ampere and then transition that to be a radically different product. Why not just scrap it and have a different code name? The only situation I can think of is someone getting fed multiple rumors and combining them such that what was said to be Lovelace was something else.

That's based on the assumption that the rumor mill gets their information from neat documents with the code name on each page and such. If I were Nvidia/AMD, I would have a public code name for the stuff that I want to share and various different code names for different internal groups, where only a few people know multiple code names and act as intermediates between groups. So then the fabbing people would be preparing for X Æ A-12, only getting the information they need for their work, while one design team would work on FLCH and another group would work on FSTNG. Then FLCH might actually be for a refresh and FSTNG for the next gen, but that wouldn't be written down anywhere where the cleaning lady can see it or whoever is leaking this stuff (it's unlikely to be the actual designers as they won't want to lose their jobs).

 

[A///]

The entire "simple refresh" rumor came from the people who spout rumors. Some rumors have truth to them but not all. I don't understand why people put so much faith in the rumors of Lovelace being a simple node refresh with minor improvements simply because some random guy from whatever country on the other side of the planet said so on his their twitter?

 

[jpiniero]

https://twitter.com/x/status/1525797225657696257

Ooh. Speculation that the first release could be Early Q3.

 

[Mopetar]

https://twitter.com/x/status/1525797225657696257

Ooh. Speculation that the first release could be Early Q3.

Seems inline with previous rumors and speculation. Maybe a bit sooner than most previously expected since most people seemed to be leaning on an early September launch.

This makes me think either late July or the beginning of August.

 

[moonbogg]

https://twitter.com/x/status/1525797225657696257

Ooh. Speculation that the first release could be Early Q3.

Nice. How long before you can actually buy one at msrp after launch? Mid next year? I predict 4070's on ebay at $1500+

 

[Saylick]

Nice. How long before you can actually buy one at msrp after launch? Mid next year? I predict 4070's on ebay at $1500+

How funny of you to assume that there will even be an MSRP.

 

[GodisanAtheist]

Let's introduce the CSRP: The Consumer Suggested Retail Price. I'm saying $650 for the 80ti/90, $550 for the 80 non to, $350 for the 70, $250 for the 60, and $150 for the 50.

Can have $50 mid point pricepoints for mid-gen refresh/Ti parts.

Guess NV/AMD won't leave us much choice if they just bail on MSRP entirely.

 

[amenx]

Nice. How long before you can actually buy one at msrp after launch? Mid next year? I predict 4070's on ebay at $1500+

Dude, the dynamics for GPU prices is completely different than what it was last year. For one, ethereum, the main mining coin that was responsible for all the massive mining farms is moving to proof of stake by 3rd or 4th quarter this year. No one is stupid enough to invest $ millions in new gen GPUs at this stage, esp with the recent crypto crash. The market will be flooded with used Ampere cards when POS happens which will further depress prices even for new gen cards.

Furthermore, Nvidia paid TSMC $ billions up front for a much larger stake in the 5nm pie simply to avoid the sort of shortages with had Ampere. I very much expect GPU prices to return to sane levels within next 6 months. They may be a little more than Ampere as cost of manufacture has gone up, but 4070s at $1500 is not going to happen. It just aint.

 

[jpiniero]

https://twitter.com/x/status/1526135149246976001

Now claims that the 4090 will be a decently cut AD102 (126/144 SM) and the official TDP is "only" 450 W.

Also says Mid July for release.

 

[igor_kavinski]

2x3090 in "just" 450W??? WOW. That's got even me salivating, even though I would never (unless I become a millionaire) get a 90 series card. Hoping for this to be true! Bodes really well for the lower end 4000 series cards.

 

[Saylick]

Damn, looks like Nvidia is bringing the heat this generation. Being able to deliver the full product in mid-July is fast considering it taped out less than a year ago. Mighty impressive stuff.

 

[Aapje]

I'm not convinced that Q3 early is going to be mid-July, especially since kopite phrased it a bit weird. Is 'Q3 early' the same thing as early Q3 or does it mean Q3, but earlier than expected? In that case, it can also be August.

 

[Timmah!]

https://twitter.com/x/status/1526135149246976001

Now claims that the 4090 will be a decently cut AD102 (126/144 SM) and the official TDP is "only" 450 W.

Also says Mid July for release.

126 out of 144 is rather big cut for my liking
was hoping for 130+
Hope at least price will be acceptable, by which i mean no more than 2000 with vat included

 

[Aapje]

126 out of 144 is rather big cut for my liking

There needs to be room for a ridiculously priced 4090 Ti. Or a 4100 Ti. Because at this point it wouldn't surprise me if they add another tier.

 

[Saylick]

I'm not convinced that Q3 early is going to be mid-July, especially since kopite phrased it a bit weird. Is 'Q3 early' the same thing as early Q3 or does it mean Q3, but earlier than expected? In that case, it can also be August.

Probably mid-July announcement, but availability like a few weeks later in early August?

 

[Saylick]

There needs to be room for a ridiculously priced 4090 Ti. Or a 4100 Ti. Because at this point it wouldn't surprise me if they add another tier.

Yeah, there's always going to be room for a Uber tier, money-no-object flagship. If the 4090 is a 20% cut-down AD102 and it ties N31 (not sure if they assume full N31 or a cut-down SKU), then that leaves Nvidia with the opportunity to sell a Titan-class GPU that's the real 600W monster at like $2500 or whatever.

I'm more interested in what Nvidia will ask for the 4090 tbh. If it offers 2x the performance of the 3090, which was $1500 MSRP and gave Nvidia great margins due to SS 8nm, if they ask for anything close to $1500 it will likely mean less margins for AD102 because TSMC N4 is probably around twice the cost of SS 8nm. Are they going to take the hit to gross margins and sell it at $1500? Knowing their love of money, I think it will be $2000.

 

[eek2121]

2x3090 in "just" 450W??? WOW. That's got even me salivating, even though I would never (unless I become a millionaire) get a 90 series card. Hoping for this to be true! Bodes really well for the lower end 4000 series cards.

Damn, looks like Nvidia is bringing the heat this generation. Being able to deliver the full product in mid-July is fast considering it taped out less than a year ago. Mighty impressive stuff.

Aren’t the AMD claims something like 3x navi21? If that is true ( which I personally doubt), AMD might have a significant performance lead this gen. I guess we will see.

Also, I knew those absurdly high TDP numbers weren’t going to hold. They are still too high, however. Hopefully both NVIDIA and AMD focus on lowering both TDP and power consumption moving forward. Energy is expensive, and renewable energy is limited. Climate change is a thing. yada yada.

Anyways, not trying to be political. My current GPU doubles as a space heater when gaming. If they keep inching up, next gen cards will be unbearable, not to mention PSU requirements. Not to mention many homes in the US only have 15 amp circuits. 🙃

 

[Aapje]

You can buy a lower tier, undervolt, etc. No one is forcing you to get a 4090 Ti Super Extreme OC Ultra and run it at stock.

 

[exquisitechar]

I'm not convinced that Q3 early is going to be mid-July, especially since kopite phrased it a bit weird. Is 'Q3 early' the same thing as early Q3 or does it mean Q3, but earlier than expected? In that case, it can also be August.

He said mid July in another tweet.

 

[jpiniero]

Aren’t the AMD claims something like 3x navi21? If that is true ( which I personally doubt), AMD might have a significant performance lead this gen. I guess we will see.

2.5x but IIRC only Navi 33 is supposed to come out soonish. The 4090 could be on the market alone for several months .