Discussion Nvidia Blackwell in Q4-2024 ?

[jpiniero]

You ignore that Nvidia is already using the more expensive to produce GDDR6X, which uses PAM4, where GDDR7 uses the more economical PAM3. And with GDDR6X, they are stuck with a single supplier, which makes it harder to bargain for better prices

I assume GDDR7 is going to be more expensive despite that.

 

[Timorous]

The 4090 captures the 'money no object'-buyers, as well a professional buyers who use it as a 'cheap' AI card, so they can't demand the same premium from lower tiers. The 4080 non-Super already proved this.

You are just arguing for them to make the same mistake with the 5080 that they made with the 4080.



A useless comparison since the 3080 Ti was priced at double the price of the 3080 because the mining boom exploded the market prices. Those prices had nothing to do what the gamer market would bear.

People were buying at those inflated prices, hence the 4080 being $1,200. Probably would have sold as well if it was 10% faster.

 

[Aapje]

People were buying at those inflated prices, hence the 4080 being $1,200. Probably would have sold as well if it was 10% faster.

They were buying because they could earn back the extra costs by mining Ethereum. That ended.

Nvidia may have thought that they could permanently increase to that price, but the 4080 sales proved them wrong. Hence the 4080 Super at $999.

 

[Saylick]

They were buying because they could earn back the extra costs by mining Ethereum. That ended.

Nvidia may have thought that they could permanently increase to that price, but the 4080 sales proved them wrong. Hence the 4080 Super at $999.

Maybe Nvidia didn’t think they could permanently increase to that price but instead wanted to simply capture the consumer surplus while they still could. If there’s one thing I’ve learned about Nvidia over the last 7 years it’s that they will always price things as high as the market will bear, regardless if they know there’s a fad or hype bubble propping up the price temporarily. They will loudly beat the hype drum to increase short term sales until it no longer becomes useful for them to do so. If beating the drum for one reason no longer proves useful but another does, they have no shame in pivoting overnight if that’s what keeps sales high.

 

[jpiniero]

They were buying because they could earn back the extra costs by mining Ethereum. That ended.

Nvidia may have thought that they could permanently increase to that price, but the 4080 sales proved them wrong. Hence the 4080 Super at $999.

The alternative might be cutting die size (which I suspect has a decent shot of happening with the 5/6/7 dies...). I just don't know how much of an improvement they can get by doing that though.

 

[igor_kavinski]

*puts on tinfoil hat*

I wouldn’t put it past Nvidia to double the throughput of their Optical Flow Accelerator and allow for one more interpolated frame for Frame Generation, so instead of a doubling of frames you could get 3x. Couple that with tensor memory compression and a market ploy where Nvidia pressures game designers to be more liberal with their VRAM usage, you get planned obsolescence of prior RTX generations.

*takes off tinfoil hat*

That should kill the 8GB cards pretty quickly. I sure hope they do that. 8GB needs to die already, in like anything other than a netbook, phone or tablet.

 

[Ajay]

Nvidia may have thought that they could permanently increase to that price, but the 4080 sales proved them wrong. Hence the 4080 Super at $999.

I think it is useful to remember that the 4080 S is a cost reduced model with a very slim performance advantage over the base 4080. If it was a standard 4080 with a 10% perf increase like the 4070 Ti S received, it would have been a more interesting purchase option. (although +20% would have made it much, much more compelling, IMHO). I hope the RTX 5080 will perform close to the 4090 did, though I expect it may be priced even higher than $1200 if so. Also, very likely to stay at 16GB. Not a problem for me, I'm skipping the next gen (or buying the 5000 series after the 6000 series come out to catch a good discount).

 

[jpiniero]

Also, very likely to stay at 16GB.

It's gonna be 16. Maybe by the time the refresh comes 3 GB chips will be available.

 

[Aapje]

The alternative might be cutting die size (which I suspect has a decent shot of happening with the 5/6/7 dies...). I just don't know how much of an improvement they can get by doing that though.

They already cut the die size of the 4080 compared to the 3080 by a huge amount. That's another reason why the 4080 price was (and actually still is) quite obscene.

@Ajay

Yes, the 4080 Super was pretty obviously a 'we screwed up the price, not the performance tier'-card.

 

[jpiniero]

They already cut the die size of the 4080 compared to the 3080 by a huge amount. That's another reason why the 4080 price was (and actually still is) quite obscene.

That was on a much much cheaper node.

 

[Aapje]

That was on a much much cheaper node.

And yet they increased the price of the 4070 by $100 as it lost about 100 mm2 compared to the 3070, while the 4080 went up by $300 (Super) or $500 (non-Super) as it lost 250 mm2 compared to the 3080.

Please tell me how this makes any sense whatsoever?

 

[jpiniero]

And yet they increased the price of the 4070 by $100 as it lost about 100 mm2 compared to the 3070, while the 4080 went up by $300 (Super) or $500 (non-Super) as it lost 250 mm2 compared to the 3080.

The wafers cost a lot more?

 

[Saylick]

The wafers leather jackets cost a lot more?

FTFY.

 

[Aapje]

The wafers cost a lot more?

Since they reduced the die size for the 4080 much more than for the 4070, they saved more on wafer costs for the 4080 (compared to them keeping die sizes the same). Yet they increased that tier's price much more.

It simply doesn't add up.

So the 4080 price was not decided by looking at the actual costs, but Nvidia thought they could maintain the mining boom prices...and they were wrong.

If you want to argue that this changed, you have to present an argument why market conditions changed, not the same debunked arguments.

 

[Mahboi]

They were buying because they could earn back the extra costs by mining Ethereum. That ended.

Nvidia may have thought that they could permanently increase to that price, but the 4080 sales proved them wrong. Hence the 4080 Super at $999.

I think that's unrelated. MLID said that Nvidia was lowering prices with the Lovelace refreshes to get some gamer goodwill in preparation for a popping of the AI bubble.
I don't think his analysis was wrong there.
Nvidia is surfing on astronomical hype and costs, but that won't last forever. Radeon would also eventually brush off the horrible situation RDNA 3 put them in.
I think the prices were a "F.U price" for customers when they only wanted to sell AI. And now that that money pot seems to get increasingly smelly, they want to sell at better prices.

Which means that I expect AMD to hit quite hard in the 1080p/1440p with Navi 44 and 48, but I really do not expect Nvidia to hold back anything at the higher tiers. Prices will still be high. Not necessarily much higher than Lovelace, but we're going from "no price/perf improvements, much more perf and much higher price" to "some perf/price improvements, more perf and still the same high price".

NV has no reason to lower prices whatsoever, except if it somehow makes their lineup nonsensical (I.E sell a 5060 Ti == N48 for $500, but sell a 20% better 5070 for $800).

 

[Mahboi]

The wafers cost a lot more?

Flimsy excuse. No amount of 5nm costs somehow outstrip an entire extra 100m² of area there.
Nvidia simply didn't want to sell volume. TSMC can only produce so much and wafers spent on 4080s are wafers not spent for H100.
When H100 is $3000 to make and $30000 to buy, while a 4080 is $500 to make and $1200 to buy, it's a no brainer.
Nvidia danced around consumers because they were a literal loss of money.

 

[Aapje]

MLID said that Nvidia was lowering prices with the Lovelace refreshes to get some gamer goodwill in preparation for a popping of the AI bubble.

Nvidia is the company that kept producing massive amounts of Ampere dies, while I was checking the Ethereum news daily to see when exactly they would change Ethereum to PoS. And it is the company that didn't understand that mining boom pricing couldn't last , as was shown by their 4080 pricing.

If they didn't react smartly to that much more predictable thing, or chose to keep GPU prices lower during the mining boom like AMD and EVGA did, for consumer goodwill, then why would they do that for the much more unpredictable end to the AI boom?

Nvidia and AMD seem to actually be remarkably poor at 'reading the room' and setting their prices at a level that satisfies consumers, let alone actually noticing developments that are going to change consumer demand.

Which means that I expect AMD to hit quite hard in the 1080p/1440p with Navi 44 and 48, but I really do not expect Nvidia to hold back anything at the higher tiers.

Keep in mind that the x090 is the low-end cut of their biggest chip. The full chip goes into professional products, at much higher prices.

With the increase in production costs, it makes sense for the gap in pricing and performance to increase between the low-end and high-end products.

 

[Mahboi]

Nvidia is the company that kept producing massive amounts of Ampere dies, while I was checking the Ethereum news daily to see when exactly they would change Ethereum to PoS. And it is the company that didn't understand that mining boom pricing couldn't last , as was shown by their 4080 pricing.

If they didn't react smartly to that much more predictable thing, or chose to keep GPU prices lower during the mining boom like AMD and EVGA did, for consumer goodwill, then why would they do that for the much more unpredictable end to the AI boom?

Fair point, maybe I'm wrong.
However the AI boom is directly visible in their sales. I.E they don't have to guess where the market will go like crypto. Crypto had the potential of replacing ETH with another scam coin. AI however is very easy to understand since Nvidia holds 95% of the market and AMD just entered it. Just look at the sales amount per month and the wait times.

Nvidia and AMD seem to actually be remarkably poor at 'reading the room' and setting their prices at a level that satisfies consumers, let alone actually noticing developments that are going to change consumer demand.

Definitely true for AMD (freaking 7600, 7900 xt pricing and 7700 xt...although that one is an upsell).
Not so for Nvidia, they're very good at reading the room.

 

[jpiniero]

Since they reduced the die size for the 4080 much more than for the 4070

I suspect the 3080 was on GA102 because SS yields were bad and because of TTM pressure from having to go back to SS8 instead of SS7 which was originally planned. It was likely intended to use GA103, which the full configuration is only 4 SMs less than the 3080's cut albeit 256-bit memory instead of 320.

 

[Heartbreaker]

I think that's unrelated. MLID said that Nvidia was lowering prices with the Lovelace refreshes to get some gamer goodwill in preparation for a popping of the AI bubble.
I don't think his analysis was wrong there.

I think you need to visit this thread:

Discussion - Should MLID be considered a legitimate source

To keep this topic from polluting other threads, I've created its own. @dullard Here is your list that you wanted to discuss: Nov 2020 June 2021 June 2022 Nov 2022 Dec 2022 May 2023 June 2023 I listed this as wrong because MTL not being Ocean Cove was known well before this...

forums.anandtech.com

MLID made up nonsense, is just nonsense as usual.

4000 Super is a mid cycle refresh that doesn't need motivation beyond that. Mid cycle refreshes are aimed at getting more marketing, to get sales increase on the current generation. Nothing more.

The AI "bubble" has nothing to do with it, and is not going to be "popping" any time soon. AI bubble is going to pop like the e-commerce bubble popped and killed Amazon... Oh, wait...

 

[Mahboi]

I think you need to visit this thread:

I think I can think for myself if an argument I hear makes sense. His did.

MLID made up nonsense, is just nonsense as usual.

Ok.

4000 Super is a mid cycle refresh that doesn't need motivation beyond that. Mid cycle refreshes are aimed at getting more marketing, to get sales increase on the current generation. Nothing more.

Ok.

The AI "bubble" has nothing to do with it, and is not going to be "popping" any time soon. AI bubble is going to pop like the e-commerce bubble popped and killed Amazon... Oh, wait...

Ok.
I got a bridge to sell you.

 

[Aapje]

However the AI boom is directly visible in their sales.

The mining boom was also visible in their sales, but the issue with booms that end is that:
- There is a lot of product in the sales channels
- The end of a boom tends to result in 2nd hand products flooding the market
- The end of a boom tends to result in much less demand

It's very difficult to clear those sales channels from excessive amounts of product, when sales fall of immensely because people buy less and if they do buy, they often buy 2nd hand.

Definitely true for AMD (freaking 7600, 7900 xt pricing and 7700 xt...although that one is an upsell).
Not so for Nvidia, they're very good at reading the room.

Except for the wrong price for the 4080 16 GB and them having to withdraw the 4080 12 GB.

 

[Heartbreaker]

I got a bridge to sell you.

Keep your bridge, you bought it, believing the nonsense story that, 4000 Super cards have something to do with the "AI bubble".

 

[Mahboi]

Except for the wrong price for the 4080 16 GB and them having to withdraw the 4080 12 GB.

I chalk that one up to extreme Chutzpah from Jensen.

 

[Aapje]

I chalk that one up to extreme Chutzpah from Jensen.

I'm not really sure why that doesn't count as misreading the market or why AMD is any different. AMD also often overvalues their products until the market smacks some reality into them.