Discussion Nvidia Blackwell in Q4-2024 ?

[jpiniero]

Why not? Those cards need PSUs capable of feeding them without any issues.
I believe It's true, but those values could still change depending on final specs.

I doubt even AIBs even know the much about the final details at this point except for the 5090 and maybe 5080.

In my opinion just a more sane naming without any Ti.

NV might feel otherwise, but IMO mediocre perf increases would go over better than huge price increases. Like if the 5060 (8 GB) is a slight cut GB206 you'd have to think it'd be $400 or more.

16 pin power cables are burning as cheap fries, are they insane to push 50+W on that insane crappy disaster!??
Is NV run by only monkeys now?

They sure don't seem to think it's a big enough problem to move away from it.

 

[igor_kavinski]

They sure don't seem to think it's a big enough problem to move away from it.

Yeah, with performance lusty users sending in their failed 4090s for repair to third parties out of their own pocket, Jensen doesn't have to care.

 

[Aapje]

NV might feel otherwise, but IMO mediocre perf increases would go over better than huge price increases. Like if the 5060 (8 GB) is a slight cut GB206 you'd have to think it'd be $400 or more.

I agree. I think that they've pushed the prices as far as they will go, or beyond.

If they want to offer something actually attractive, they should sell a 12 GB 5060 Ti based on GB205.

 

[TESKATLIPOKA]

I doubt even AIBs even know the much about the final details at this point except for the 5090 and maybe 5080.

It's enough If Nvidia told that this is the highest TBP It could be.

 

[jpiniero]

For S&G, how about this?

5090	140 SM	28 gbps	24 GB	4090 + 40-50%	$1999
5080	80 SM	28 gbps	16 GB	4090 - 10%	$1299
5070	48 SM	26 gbps	12 GB	~ 4070 Ti Super	$699
5060	36 SM	26 gbps	8 GB	4070 - 10%	$449
5050	20 SM	20 gbps G6	8 GB	4060 + 5%	$329

With the refresh adding 50% more ram and 10-15% more performance for the G7 models if they also use 32 gbps memory at the top.

 

[SteinFG]

For S&G, how about this?

5090	140 SM	28 gbps	24 GB	4090 + 40-50%	$1999
5080	80 SM	28 gbps	16 GB	4090 - 10%	$1299
5070	48 SM	26 gbps	12 GB	~ 4070 Ti Super	$699
5060	36 SM	26 gbps	8 GB	4070 - 10%	$449
5050	20 SM	20 gbps G6	8 GB	4060 + 5%	$329

With the refresh adding 50% more ram and 10-15% more performance for the G7 models if they also use 32 gbps memory at the top.

This is way off base. 4070 12GB is already at $500. 4060 Ti 8GB is already a "4070 - 20%" for $350. 4070 Ti Super is at $750.
Is your assumption that 4070 will go way above $500 after the launch of 50 series? Or that nvidia wants to sell exactly 0 new cards?

 

[MoogleW]

For S&G, how about this?

5090	140 SM	28 gbps	24 GB	4090 + 40-50%	$1999
5080	80 SM	28 gbps	16 GB	4090 - 10%	$1299
5070	48 SM	26 gbps	12 GB	~ 4070 Ti Super	$699
5060	36 SM	26 gbps	8 GB	4070 - 10%	$449
5050	20 SM	20 gbps G6	8 GB	4060 + 5%	$329

With the refresh adding 50% more ram and 10-15% more performance for the G7 models if they also use 32 gbps memory at the top.

Nvidia is staying in the same node family. No need to:

-Downgrade 5070 from 4070 Super spec while charging 15% more money for only 20% more performance
-Essentially refresh 4080 super for higher than 4080 super msrp at only 10% performance improvement.
-Essentially refresh rtx 4060 at a 10% higher msrp and more or less same performance

All in all, I think this is overly pessimistic and would guarrantee the most scathing reviews in Nvidia history with a nosedive in comaparative launch sales all while not needing to because BOM costs are relatively even.

 

[jpiniero]

Nvidia is staying in the same node family. No need to:

Which I think is part of the problem, especially for a node that's getting more expensive and not cheaper.

 

[jpiniero]

Is your assumption that 4070 will go way above $500 after the launch of 50 series?

MSRP is $549 and most cards are right about there or higher. I think there's one card that's $529 at Amazon but that's about it.

Edit: I do think whatever they end up calling the better GB205 card can get to 4070 NS performance.... it still seems a stretch ATM.

 

[SteinFG]

MSRP is $549 and most cards are right about there or higher. I think there's one card that's $529 at Amazon but that's about it.

Edit: I do think whatever they end up calling the better GB205 card can get to 4070 NS performance.... it still seems a stretch ATM.

499 right now on newegg, with a game

https://www.newegg.com/p/N82E16814932685

 

[Aapje]

@jpiniero

All those prices are silly, except for the 5090.

 

[jpiniero]

499 right now on newegg, with a game

https://www.newegg.com/p/N82E16814932685

View attachment 103217

Oh interesting. Must be a sale related to Prime Day (ha). But that's seemingly rare.

All those prices are silly

I expect it to be silly. Maybe the cards will have better RT performance relative to the raster perf, and of course AI. And these really shouldn't be the names, since people get very hung up on them.

 

[Aapje]

I expect it to be silly. Maybe the cards will have better RT performance relative to the raster perf, and of course AI. And these really shouldn't be the names, since people get very hung up on them.

They reached the end of people's budgets. I expect pretty much the same prices for all tiers, but the 5090.

 

[jpiniero]

They reached the end of people's budgets. I expect pretty much the same prices for all tiers, but the 5090.

Would you feel better if they called it the 5070 Ti, 5060 Ti and 5060?

 

[Aapje]

Would you feel better if they called it the 5070 Ti, 5060 Ti and 5060?

I would indeed feel better if they named and priced the 5080 as the 5070 Ti, the 5070 as the 5060 Ti and the 5060 Ti as the 5060.

I know that you didn't mean this, but I am actually serious about using better chips in some tiers. They already did this with AD103 for the 4070 Ti refresh and I expect them to use the GB203 for the 5070 Ti. And I think that they should use GB205 for both the 5070 and the 5060 Ti, both with 12 GB. That would be something actually decent, where I would be positive about people buying the 5060 Ti (assuming that the price stays the same), over going to AMD or waiting for the refresh with hopefully 3 GB VRAM chips.

 

[jpiniero]

I know that you didn't mean this, but I am actually serious about using better chips in some tiers.

With costs only going up, that seems doubtful.

Waiting for 3 GB chips for the 60 tier products wouldn't be a bad idea. It just seems like they won't given that they didn't refresh the Ada models.

 

[Aapje]

With costs only going up, that seems doubtful.

They already prize gouged for Ada. And they should get a better deal with TSMC after buying capacity at the worst possible time (just before the mining boom ended) for Ada, so the extra costs for GDDR7 should be offset by lesser costs for the chips.

 

[jpiniero]

They already prize gouged for Ada. And they should get a better deal with TSMC after buying capacity at the worst possible time (just before the mining boom ended) for Ada, so the extra costs for GDDR7 should be offset by lesser costs for the chips.

There's been recent rumors that TSMC is raising prices yet again. Like 10-15% IIRC.

 

[Aapje]

The companies don't pay the same prices. Again, it was leaked and makes total sense that Nvidia already overpaid massively for the Ada wafer reservations, because they bought massive volume just before the end of the mining boom.

If they overpaid before, and are now sticking to the same process (and prices for the same process go down over time), then it seems unlikely that they will now pay more. Probably less rather than more.

 

[jpiniero]

If they overpaid before, and are now sticking to the same process (and prices for the same process go down over time), then it seems unlikely that they will now pay more. Probably less rather than more.

Everyone would be paying more.

 

[Aapje]

Everyone would be paying more.

You are still not getting it.

Imagine someone hugely overpaying because they get a Strix card, despite it offering no real performance advantage. Then for their next buy, they wise up and get a much cheaper variant. Then they can spend less, even if the card prices went up. After all, their mistake cost them more money than the price increase.

Now again, Nvidia bought lots of capacity for Ada just before the demand for TSMC capacity drastically reduced, due to the end of the mining boom. It's pretty obvious that they paid a lot more for that capacity then if they would have bought it later.

 

[jpiniero]

Now again, Nvidia bought lots of capacity for Ada just before the demand for TSMC capacity drastically reduced,

Prices for N5/N4 are higher now compared to then though. That's the thing. Even if NV really did overpay, you'd have to think they still paid much less than what the going rate is now.

 

[igor_kavinski]

Even if NV really did overpay, you'd have to think they still paid much less than what the going rate is now.

Yeah so the 5000 series is gonna be more expensive coz Jensen isn't suddenly going to become more charitable.

I mean, poor guy is already sacrificing quite a few leather jackets per year to deliver so much to the world.

And AMD Instinct Accelerators are poised to put a further dent in his jacket collection count. God forbid if he had to sell some of those just to pay the bills for his mistresses and associated progeny.

 

[Aapje]

Prices for N5/N4 are higher now compared to then though.

The MSRP of the 4080 is 70% higher than the MSRP of the 3080. Yet there are people that paid $2400 for the 3080 in June 2021. But Nvidia never announced a 240% percentage price increase. So by your reasoning, no one could have paid these huge sums for 3080s during the mining boom, yet they did.

It is pretty obvious that TSMC asked more money during the mining boom, yet they never announced that they were price gouging, for obvious reasons (people don't like price gouging), just like Nvidia never did. You are letting yourself be played by these companies, by believing that they weren't price gouging just because they didn't announce it and believing that their simplistic announcements of price increases actually match reality.

Again, all these contracts are negotiated individually and the actual costs depend on market conditions at the time, the nodes they use (and what competition there is for those nodes), etc.

 

[jpiniero]

You got AI hype (even if packaging limited) + Intel using a lot of TSMC too.