Speculation and Rumors: Nvidia Blackwell ...

[Arun]

I split the CDNA/MI300 discussion into a new thread: https://forum.beyond3d.com/threads/...nalysis-speculation-and-rumors-in-2024.63538/ (some tidbits about Blackwell were also moved there as they were part of MI300-centric posts unfortunately, but it wasn't as bad as I thought it might be).

Please try to keep a higher signal-to-noise ratio in both threads, thanks!

 

[Shifty Geezer]

Just a polite reminder that it was only 2 months ago that B3D was on life support and the switch was thrown off, only to be talked into being switched back on again. But behaviour is reverting to old habits and we know where that leads.

You guys have to self moderate if you want B3D to stay alive. Have to. It's not an option. It's not something for everyone else on the board to do because they are the ones that are doing it all wrong.

Do not discuss the board or the posters in your posts.
Discuss the topic or walk away.
And be aware of your own posts and when you can see you are going OT, message the mods for a branch, or start a new thread, or don't post.

When you post, remember the week in January when this place was dead and keep that in the back of your head, "do I want that to happen again?" And if you don't mind if it happens again, please go elsewhere.

 

[xpea]

H100 80GB is at 700w. 1000w for a dual big chip B100 192GB would likely be a significant improvement in perf/watt.

192GB is just the start for Blackwell. 288GB will follow closely with B200. And AMD won't go anywhere as NVDA bought nearly all HBM3e and COWOS capacity up to end of 2025 (NV is the second largest TSMC customer with more than 10% production allocated). The thing is that NV turnover will be over $100B this year with more than $50B of free cash flow generated. So lot of dirty things can be done behind the doors on the supply chain with so much cash in hand...

 

[Kaotik]

192GB is just the start for Blackwell. 288GB will follow closely with B200. And AMD won't go anywhere as NVDA bought nearly all HBM3e and COWOS capacity up to end of 2025 (NV is the second largest TSMC customer with more than 10% production allocated). The thing is that NV turnover will be over $100B this year with more than $50B of free cash flow generated. So lot of dirty things can be done behind the doors on the supply chain with so much cash in hand...

Just because CoWoS and HBM are sold out it doesn't mean NVIDIA bought them all, so do you have some source for he claim? It's not like NVIDIA has some special right to order first and rest get what's left - extra money helps but walking over done deals leads to court quick.
Yes, they grew from 5th to 2nd biggest TSMC customer last year but the likes of AMD and Qualcomm aren't far behind either. NV doesn't account necessarily more than 10% of production, they did 11% of revenue at 7.7B but TSMC is more than just the modern processes NVIDIA etc use. Don't have 23 numbers here but 22 Qualcomm did 8.9% at 6.8B and AMD 7.6% at 5.8B for TSMC (https://exploresemis.substack.com/p/tsmcs-top-10203040-customers-who)

 

[xpea]

Just because CoWoS and HBM are sold out it doesn't mean NVIDIA bought them all, so do you have some source for he claim? It's not like NVIDIA has some special right to order first and rest get what's left - extra money helps but walking over done deals leads to court quick.
Yes, they grew from 5th to 2nd biggest TSMC customer last year but the likes of AMD and Qualcomm aren't far behind either. NV doesn't account necessarily more than 10% of production, they did 11% of revenue at 7.7B but TSMC is more than just the modern processes NVIDIA etc use. Don't have 23 numbers here but 22 Qualcomm did 8.9% at 6.8B and AMD 7.6% at 5.8B for TSMC (https://exploresemis.substack.com/p/tsmcs-top-10203040-customers-who)

It's like that
ALL Micron HBM3e 8-Hi stack produced in the next 4 quarters goes into H200
ALL SK-Hynix HBM3e 12-Hi stack goes into B100/200 in the first 6 quarters of production
SAMSUNG have still open orders for HBM3e but they are at least 2 quarters later for MP
Regarding TSMC, all capacity of new COWOS fab coming online this quarter is already booked up to end of 2025 (paid by NVDA 2 years ago). Of course AMD still have their allocation but they won't grow much (relative to NV). To the point that it creates frictions with TSMC and AMD is in talk with ASE, KYEC, and Winbond for their packaging needs.
For 2024, it's expected that NVDA will represent ~15% of TSMC revenue, double of AMD. AI changed everything...

PS: my source is well placed (been working 30 years in the industry). Can't obviously say more. So believe what you want

 

[Bondrewd]

AMD is in talk with ASE, KYEC, and Winbond for their packaging needs.

That's really not for MI300.
Not for 400 either, that's for long-term fanout capacity for parts not GPGPU.

my source is well placed (been working 30 years in the industry). Can't obviously say more. So believe what you want

My source is saying MS thinks otherwise about 350X and the 400 shebangs (which they'll be ordering quite a lot of).
They're buying stuff which according to you is impossible/doesn't exist. Wonder why. Must be miracles and magic.

You really should go back to Reddit.

 

[CarstenS]

Please do not post in a derogatory manner like this („go back to Reddit“).

 

[xpea]

That's really not for MI300.
Not for 400 either, that's for long-term fanout capacity for parts not GPGPU.

My source is saying MS thinks otherwise about 350X and the 400 shebangs (which they'll be ordering quite a lot of).
They're buying stuff which according to you is impossible/doesn't exist. Wonder why. Must be miracles and magic.

You really should go back to Reddit.

I didn't say AMD won't have any QTY. I said that they can't grow as much as they want. Obviously MS being a tiers one account, they will give more allocation to them but it's at the detriment of other customers. And from AMD own financial forecast for this year, they show a relatively slow increase in their DC business relative to NVDA, which proves my point entirely.


PS: if you could stop personal attacks when you have sh*t to say, it will make this place a much better community. And finally, you must be very confused, I don't have a reddit account, nor facebook, nor tiktok, I'm too old for that...

 

[Bondrewd]

I said that they can't grow as much as they want

Yeah they can lol, see analyst estimates.

And from AMD own financial forecast for this year,

It doesn't matter, only the street numbers do.

This is all very low quality bait and you should try harder.
Maybe pay for like UBS reporting.
Or Bernsteins, they both did nice TAM (in units) modeling.

 

[xpea]

Yeah they can lol, see analyst estimates.

It doesn't matter, only the street numbers do.

This is all very low quality bait and you should try harder.
Maybe pay for like UBS reporting.

Bullshit. From AMD CEO on last earnings call:

"For 2024, we expect the demand environment to remain mixed" AMD CEO Lisa Su said on a call with analysts.
.../...
In October, AMD said it expected $2 billion in server GPU sales in 2024. On Tuesday, it said it now expects $3.5 billion in data center GPU sales under its “Instinct” brand this year.

source: https://www.cnbc.com/2024/01/30/amd-earnings-report-q4-2024.html

So during ALL 2024, AMD will barely sell 1/5th of what NVDA is already doing in ONE SINGLE quarter in DC revenue. The obvious reason is that AMD could sale 10 times more MI300 if they had enough supply. So please, stop spreading FUD, it's very annoying

 

[Bondrewd]

Bullshit. From AMD CEO on last earnings call:

What Lisa says doesn't matter much, she has to sandbag in case market changes.

So please, stop spreading FUD, it's very annoying

Come on, post some analyst GPU TAM modeling for 2024 at least.
It's not that hard (even if it's not public).

The obvious reason is that AMD could sale 10 times more MI300 if they had enough supply.

but they will?
Again, see the analyst nodes from the recent wave of $AMD upgrades.
Pretty clear numbers on revenues and their mss% relative to GPU TAM overall.

 

[xpea]

What Lisa says doesn't matter much, she has to sandbag in case market changes.

I stopped to read here. By law, the CEO of a public company cannot lie or "sandbag" to investors (forecast numbers are always the most accurate estimations with a small percentage variation for uncertainties) . It's not kindergarten playground. Grow up

 

[Bondrewd]

By law, the CEO of a public company cannot lie or "sandbag" to investors

Yeah they can, Jensen does it routinely.

(forecast numbers are always the most accurate estimations with a small percentage variation for uncertainties)

Christ there's a reason neither vendor gives FY guidance anymore (they literally don't know how or where stuff gonna end).
Please stop talking financial stuff like you understand it.
It's just bad, very low-quality bait.

 

[troyan]

Bullshit. From AMD CEO on last earnings call:


source: https://www.cnbc.com/2024/01/30/amd-earnings-report-q4-2024.html

So during ALL 2024, AMD will barely sell 1/5th of what NVDA is already doing in ONE SINGLE quarter in DC revenue. The obvious reason is that AMD could sale 10 times more MI300 if they had enough supply. So please, stop spreading FUD, it's very annoying

AMD said they have secured enough supply to grow with the demand. Obviously demand is very low for their products. Jensen Huang said in the last conference call, that Blackwell will be supply constraint.

 

[xpea]

(1)Yeah they can, Jensen does it routinely.

(2)Christ there's a reason neither vendor gives FY guidance anymore (they literally don't know how or where stuff gonna end).
(3)Please stop talking financial stuff like you understand it.
It's just bad, very low-quality bait.

Again, you talk nonsense as everybody on this board clearly see.
(1) Give a proof of Jensen lying to investors. NVDA strategy, for many years is to slightly underestimate its forecast in order to create a positive narrative (with a bullish effect on the stock). But its always within a small margin of error. For example, they say they will do $1B +/-2% and they over perform by 3~5%.
(2) CFOs and CEOs always give forecast, especially in semi conductor industry where orders takes 6 months to manufacture due to the complexity and hundreds of processes involved + supply chain constraints. Thus AMD already know exactly how much wafers and packaging allocation they will play with in 2024. They can't double their orders with a snap. That's why Lisa Su gave a forecast for 2024, contradicting your claims.
(3) Looking at your random ramblings that always come without any source nor anything backing your claims, you are definitely not in a comfortable position to talk finance, especially when you don't know my work. Sorry but I don't have any financial lessons to receive from a baseless delusional "always wrong" guy like you

No need to reply, I will ignore it. This conversation is going nowhere and I'm tired of your lies

 

[Granath]

As a result, NVIDIA’s chips account for 40% to 50% of TSMC’s CoWoS packaging capacity.

[News] CoWoS Capacity Shortage Challenges AI Chip Demand, while Taiwanese Manufacturers Expand to Seize Opportunities | TrendForce Insights

With the flourishing development of technologies such as AI, cloud computing, big data analytics, and mobile computing, modern society has an increasi...

www.trendforce.com

 

[Bondrewd]

Give a proof of Jensen lying to investors. NVDA strategy, for many years is to slightly underestimate its forecast in order to create a positive narrative (with a bullish effect on the stock). But its always within a small margin of error. For example, they say they will do $1B +/-2% and they over perform by 3~5%.

It's same old Hock Tan style beat and raise guide.

CFOs and CEOs always give forecast, especially in semi conductor industry where orders takes 6 months to manufacture due to the complexity and hundreds of processes involved + supply chain constraints

THERE IS NO FULL YEAR GUIDANCE FROM EITHER VENDOR

Thus AMD already know exactly how much wafers and packaging allocation they will play with in 2024

They don't.

That's why Lisa Su gave a forecast for 2024, contradicting your claims.

SHE DIDN'T THERE'S NO FY'24 GUIDANCE.
aaaaaaaaaaaa goddamit.

As a result, NVIDIA’s chips account for 40% to 50% of TSMC’s CoWoS packaging capacity.

[News] CoWoS Capacity Shortage Challenges AI Chip Demand, while Taiwanese Manufacturers Expand to Seize Opportunities | TrendForce Insights

With the flourishing development of technologies such as AI, cloud computing, big data analytics, and mobile computing, modern society has an increasi...

www.trendforce.com

Yep.
AMD (well, with Xilinx) is >20%-ish.

 

[Arun]

SHE DIDN'T THERE'S NO FY'24 GUIDANCE.
aaaaaaaaaaaa goddamit.

The problem across multiple threads is that while you might be right for a lot of things, you effectively claim 100% confidence for practically everything. There is absolutely no way you (or anyone, ever) is right 100% of the time. The only way anyone can claim to be right 100% of the time is to distort words in such a way that they can claim to have been correct, despite those words not univerally having that meaning (even if they do have that meaning in some contexts, you absolutely cannot assume that is the context everyone will understand it in, unless you explicitly specify exactly what you mean).

This is a clear example in my opinion: https://www.fool.com/earnings/call-...ed-micro-devices-amd-q4-2023-earnings-call-t/

Looking ahead, our prior guidance was for Data Center GPU revenue to be flattish from Q4 to Q1 and exceed $2 billion for 2024. Based on the strong customer pool and expanded engagements, we now expect Data Center GPU revenue to grow sequentially in the first quarter and exceed $3.5 billion in 2024. We have also made significant progress with our supply chain partners and have secured additional capacity to support upside demand.

There's no way you can claim there's no FY'24 guidance unless you distort what "FY'24 guidance" means in a way that isn't obvious (e.g.: there's no FY'24 guidance for the entire business, only datacenter GPUs, but the discussion was about datacenter GPUs, so... hopefully you can agree with that). And if so, doesn't this level of anger and confidence seem unjustified?

It's same old Hock Tan style beat and raise guide.

Also, you could definitely make an argument that the way NVIDIA has beat guidance every time is by being slightly pessimistic in their guidance, and that this constitutes lying. That's a perfectly reasonable thing to claim, for a certain definition of lying and with, say, 90% confidence or less... The way you state everything vaguely and as if it was an obvious fact when it is only a very likely conjecture is extremely counterproductive

 

[Bondrewd]

The problem across multiple threads is that while you might be right for a lot of things, you effectively claim 100% confidence for practically everything

This isn't what I'm talking about here.
It's not about right or wrong, it's vendors presenting hard yearly rev numbers, which is something they did pre-bubble and no longer do since no has any concrete ideas as to how the market gonna play out.
Best you've got is analyst models/estimates.
I've quoted the specific UBS note. If the guy's as well connected as he >implies he is, he can find it willy nilly.

(e.g.: there's no FY'24 guidance for the entire business, only datacenter GPUs, but the discussion was about datacenter GPUs, so... hopefully you can agree with that)

That's exactly what it means.
On-ER call remark about orders on the books to an analyst isn't binding the way a proper FY'24 guidance is.
No ML player, neither Intel, nor AMD nor NV game a FY'24 guidance because no one has any idea how the final revenue gonna look like.
Only some vague "H2 demand mix" statements and they don't really amount to much.

Also, you could definitely make an argument that the way NVIDIA has beat guidance every time is by being slightly pessimistic in their guidance, and that this constitutes lying

No it's a pretty standard "guide very conservatively, beat and raise", there's no lying about that, especially in very competitive markets like these.
As I've said, Hock Tan loves doing that in particular.

 

[Shifty Geezer]

You really should go back to Reddit.

This is all very low quality bait and you should try harder.

Bullshit. From AMD CEO on last earnings call:

Again, you talk nonsense as everybody on this board clearly see.

There's way too much impoliteness and rudery in this thread. Discuss the posts and points, and not the posters. If you disagree, express your point clearly and patiently without assumptions. If that's too much like hard work, just don't post.

If you aren't enjoying the discussion, you should find something more fun to do. I'll be happy to help those struggling to self-moderate to find time to do that other more enjoyable thing.