Question Zen 6 Speculation Thread

[gdansk]

Nvidia earnings and outlook make it look like AMD is stuck in the mud.

A subtle addition to your threadcrapping spree: AMD and every other hardware company. AMD's closer to Micron than Broadcom in revenue growth.

Not that it has much to do with Zen 6 yet.

 

[Joe NYC]

A subtle addition to your threadcrapping spree: AMD and every other hardware company. AMD's closer to Micron than Broadcom in revenue growth.

Not that it has much to do with Zen 6 yet.

How about this, Zen 6 related item.

It may very well be that Venice dense cores are 4 x 9 = 36 cores, and then AMD will never have to sell CCDs with less than 32 cores.

On one of the pictures down in the thread, finds a picture from MLID video that looks like 9 x 4 arrangement.

https://twitter.com/x/status/2026655978339577877

 

[adroc_thurston]

It may very well be that Venice dense cores are 4 x 9 = 36 cores, and then AMD will never have to sell CCDs with less than 32 cores.

It's a 4*9 mesh (4*8 cores and a service row).
32c all in all. Binned ones will go into 192c SKU and friends.

 

[Joe NYC]

IMHO the big datacenter spenders probably tried to use AMD and others as bargaining chips to get sweeter deals from Nvidia

You mean that hyperscallers who spend $100s of billions of dollars use the same strategy as gamers?

 

[Joe NYC]

It's a 4*9 mesh (4*8 cores and a service row).
32c all in all. Binned ones will go into 192c SKU and friends.

I see, so 1 or 2 could be for connecting to IOD. What else?

 

[adroc_thurston]

You mean that hyperscallers who spend $100s of billions of dollars use the same strategy as gamers?

Yeah.
(well, not quite).

I see, so 1 or 2 could be for connecting to IOD. What else?

CPLs, debug and test pads, etc.
Also it's one SDP per mesh column so you get 4 stops for it all.

 

[Joe NYC]

In my (deeply negative) PoV it seems like AMD still desperately needs somebody to buy their AI accelerators at meaningful scale. One would thought that a viable AI product would sell itself at times of an AI craze...

It's a little bit of a chicken and egg problem.

- You can't go from selling from $5 billion per year to $50 billion per year unless you get some $100 billion multi-year orders
- You don't get $100 billion multiyear orders unless you are already shipping multiple $10s of billions of AI gear per year

So, AMD came up with this strategy to break out of this chicken and egg problem.

IMO, AMD may still offer the same deal to Microsoft / Amazon / Google, whichever of these comes first, before this train leaves the station.

 

[adroc_thurston]

- You can't go from selling from $5 billion per year to $50 billion per year unless you get some $100 billion multi-year orders
- You don't get $100 billion multiyear orders unless you are already shipping multiple $10s of billions of AI gear per year

So, AMD came up with this strategy to break out of this chicken and egg problem.

This has nothing to do with that and everything to do with NV's FCF pile.
2018-2021 taught AMD well.

IMO, AMD may still offer the same deal to Microsoft / Amazon / Google, whichever of these comes first, before this train leaves the station.

GOOG has their own stuff.
MS works with AMD normally.

 

[Joe NYC]

This has nothing to do with that and everything to do with NV's FCF pile.
2018-2021 taught AMD well.

GOOG has their own stuff.
MS works with AMD normally.

Yeah, that would be my order of which of the remaining 3 could possibly go for a similar $100 billion, 6 GW deal + warrants:

1. Microsoft
2. Amazon
3. Google

 

[adroc_thurston]

Yeah, that would be my order of which of the remaining 3 could possibly go for a similar $100 billion, 6 GW deal + warrants:

They don't have to.
AMD needs only Meta.
Everyone else for MI500 gen onwards.

 

[inquiss]

They don't have to.
AMD needs only Meta.
Everyone else for MI500 gen onwards.

Only needs them because they have no capacity for anything else? Because they work the most on software? Because they need one hyperscaler to prove it out? What's your thinking.

 

[adroc_thurston]

Only needs them because they have no capacity for anything else? Because they work the most on software? Because they need one hyperscaler to prove it out? What's your thinking.

All three!
Stock carrots work against meet the comp discounts, but they sure won't help if NV prepaid for all the HBM4.

 

[Joe NYC]

All three!
Stock carrots work against meet the comp discounts, but they sure won't help if NV prepaid for all the HBM4.

It seems that NVidia managed to shoot itself in the foot by raising HBM4 specs (clock speeds) and there will be a ton of HBM4 left that will not meet NVidia spec, but will meet AMD spec.

At least for Rubin generation + HBM4. We will see what happens with the next gen.

 

[adroc_thurston]

It seems that NVidia managed to shoot itself in the foot by raising HBM4 specs (clock speeds) and there will be a ton of HBM4 left that will not meet NVidia spec, but will meet AMD spec.

NV can sell lower HBM speed SKUs just fine.
It's not rocket science.

 

[Joe NYC]

They don't have to.
AMD needs only Meta.
Everyone else for MI500 gen onwards.

That's a good point. There is only so much that AMD will be able to deliver in the first year of Mi400, and Mi500 may be the starting point for additional demand.

As far as locking up all the demand, I have a feeling that all vendors, from TSMC, to packaging houses to memory makers to assemblers, none of them want to be in position to have only 1 customer.

Because then, that 1 customer will start to make demands they can't refuse, will start to run their business.

 

[adroc_thurston]

There is only so much that AMD will be able to deliver in the first year of Mi400

Not really but NV will try to strangle them.

As far as locking up all the demand, I have a feeling that all vendors, from TSMC, to packaging houses to memory makers to assemblers, none of them want to be in position to have only 1 customer.

they dgaf. demand is demand.

Because then, that 1 customer will start to make demands they can't refuse, will start to run their business.

It's the opposite.
NV is at TSM's (and others) mercy.

 

[1250]

It seems that NVidia managed to shoot itself in the foot by raising HBM4 specs (clock speeds) and there will be a ton of HBM4 left that will not meet NVidia spec, but will meet AMD spec.

Samsung is said to have a lower low-bin (kgsd) count.
Since the base die is different, kgsd is meaningless. Lisa will likely want to buy (TSV kgd) at a lower price.
However, I'm not sure how much cheaper it will be than Nvidia (kgsd $700).

 

[inquiss]

It seems that NVidia managed to shoot itself in the foot by raising HBM4 specs (clock speeds) and there will be a ton of HBM4 left that will not meet NVidia spec, but will meet AMD spec.

At least for Rubin generation + HBM4. We will see what happens with the next gen.

How does AMD grow if there's no HBM supply? Can they?

 

[adroc_thurston]

How does AMD grow if there's no HBM supply?

That's the neat part:

Can they?

Nope.

 

[1250]

Low bin kgsd will go to Broadcom, I know Google is skipping hbm4 and sampling 4e (probably 12Hi)

 

[Joe NYC]

Low bin kgsd will go to Broadcom, I know Google is skipping hbm4 and sampling 4e (probably 12Hi)

Why Broadcom and not AMD? AMD specs call for very low bin of HBM4.

 

[adroc_thurston]

Why Broadcom and not AMD? AMD specs call for very low bin of HBM4.

AMD HBM is kustom.

 

[Joe NYC]

AMD HBM is kustom.

Well, the base is custom, but the actual HBM stack is standard.

BTW, who is making AMD's custom base die?

 

[adroc_thurston]

Well, the base is custom, but the actual HBM stack is standard.

That means the entire stack is custom.

BTW, who is making AMD's custom base die?

Samsung.

 

[1250]

Why Broadcom and not AMD? AMD specs call for very low bin of HBM4.

What I'm talking about is exactly what was made for Nvidia and was dropped. I'm currently asking the current staff how to make HBM.