Discussion Zen 5 Speculation (EPYC Turin and Strix Point/Granite Ridge - Ryzen 9000)

[DisEnchantment]

Speculate at will

 

[Ajay]

It's only AI everything if you can compete with nVidia which no one but AMD really can.

From the con call, Lisa even talked about Ryzen AI. Curious if there is some new branding coming along. Anyway, the MI300 series is the first real competitor (IMHO) to Hopper. It is going to take allot of time and effort for AMD to grab significant market share from the incumbent. I hope they execute well.

 

[Joe NYC]

From the con call, Lisa even talked about Ryzen AI. Curious if there is some new branding coming along. Anyway, the MI300 series is the first real competitor (IMHO) to Hopper. It is going to take allot of time and effort for AMD to grab significant market share from the incumbent. I hope they execute well.

I wonder if AMD is going to try to play the AI angle to knee-cap Intel Raptor Lake... Because AMD can play the AI angle with Phoenix and Hawk Point mobile client APU, but Intel can't.

As far as Data Center, I think we need to wait for some performance numbers from AMD next month before we can make projections. If Mi300 outperforms H100, than on some categories of workloads, it may gain traction faster than in Datacenter CPU.

There was a lot of legacy in that market. There are companies still buying Ice Lake CPUs, which is not even in a ballpark of being competitive, but there are a lot of legacy deployments, and people are now just throwing bad money after - bad money.

 

[adroc_thurston]

Anyway, the MI300 series is the first real competitor (IMHO) to Hopper.

It's a big stick, yea.

It is going to take allot of time and effort for AMD to grab significant market share from the incumbent

They'll very much continue dropping the hammer with MI400.

it may gain traction faster than in Datacenter CPU.

Lisa pretty openly said it's their fastest-to-$1B rev product ever.

 

[Ajay]

There was a lot of legacy in that market. There are companies still buying Ice Lake CPUs, which is not even in a ballpark of being competitive, but there are a lot of legacy deployments, and people are now just throwing bad money after - bad money.

A professor of Warren Buffet had said that (stock) markets are able to sustain a level of irrationality for unexpectedly long periods of time. Apparently the same is true CPU purchases.

 

[Goop_reformed]

It's a big stick, yea.

They'll very much continue dropping the hammer with MI400.

Lisa pretty openly said it's their fastest-to-$1B rev product ever.

It might be but MI300 is lagging wayy behind nvidia now they've just moved to yearly cadence unfortunately

 

[adroc_thurston]

It might be but MI300 is lagging wayy behind nvidia

No.
You'll see why pretty soon.

now they've just moved to yearly cadence unfortunately

They moved to that cadence because comp between them and AMD in big schlong GPGPU sticks gonna be insane.

 

[branch_suggestion]

It might be but MI300 is lagging wayy behind nvidia now they've just moved to yearly cadence unfortunately

It is still a 2 year cadence for a fully new product, they are just doing really aggressive posturing to try to seem indomitable.
Aka late 90's Cisco, they declared a plan to dominate the world, several companies worked hard to surpass them and Cisco failed to hit their completely unrealistic roadmap, leading them to go from the #1 stock on the planet to another IBM.
NV is Cisco, as soon as they no longer have technical superiority (software moat is nearly dead thanks to mainly 3rd party stuff ala PyTorch/Triton et al, hardware moat fully dies by MI400) their margins will crater to oblivion, but I expect overall AI demand to crater at the same time, just like the 2000 crash.
Plus H100 took ages to ramp, it was really 6 months or so late to HVM.
NV is a boom-bust company with many historical periods of unsustainable growth, followed by inventory corrections and a return to the mean CAGR. As soon as the LLM juice has dried up, they will have to cope with the collateral damage. I'm sure they have their next unsustainable paradigm in mind already, but sadly for them, they are now the tall poppy to the whole industry, taking over from Intel.

 

[BorisTheBlade82]

The trouble I am seeing with MI300 is, that they will be competing with NV for the same precious CoWoS capacity. Early on they wanted to use InFO-RDL like in N31, but that did not seem to work. So supply might be rather sparse as NV seems to have allocated the lion share. Costs are higher as well, but that should not be a big deal in that segment.

 

[adroc_thurston]

The trouble I am seeing with MI300 is, that they will be competing with NV for the same precious CoWoS capacity

We're so lucky to have OSATs.

So supply might be rather sparse as NV seems to have allocated the lion share

AMD's communicated pretty clear revenue targets.

 

[H433x0n]

The trouble I am seeing with MI300 is, that they will be competing with NV for the same precious CoWoS capacity. Early on they wanted to use InFO-RDL like in N31, but that did not seem to work. So supply might be rather sparse as NV seems to have allocated the lion share. Costs are higher as well, but that should not be a big deal in that segment.

They should unironically contract IFS for dual sourcing MI300 packaging. I don’t think there’s much of an IP risk through packaging alone so it should be a serious consideration.

Nvidia is already dual sourcing packaging to keep up with demand and I’m fairly certain it’s through IFS.

 

[adroc_thurston]

They should unironically contract IFS for dual sourcing MI300 packaging

EMIB pitch insufficient and I don't think Intel can reticle stitch passive slabs.

 

[H433x0n]

EMIB pitch insufficient and I don't think Intel can reticle stitch passive slabs.

If they’re able to handle the packaging on Ponte Vecchio, Gaudi & H100 I don’t see how MI300 would be an issue.

I think they’d be foolish to not try to inquire about it at least.

 

[adroc_thurston]

they’re able to handle the packaging on Ponte Vecchio

It's all very not tight pitches (like, 45um EMIB is not sufficient to run USRs).

Gaudi

It's TSMC lol.

H100

Is not packaged at Intel.

I think they’d be foolish to not try to inquire about it at least.

AMD works with actual, real OSATs and not the joke that is Intel.

 

[H433x0n]

It's TSMC lol.

The dies are fabricated at TSMC but packaging is done at Intel afaik.

Is not packaged at Intel.

H100 packaging is being dual sourced now.

AMD works with actual, real OSATs and not the joke that is Intel.

This attitude could be the difference between 2.0B in revenue vs 2.5B in revenue from MI300.

 

[adroc_thurston]

The dies are fabricated at TSMC but packaging is done at Intel afaik

No, lol, it's CoWoS-S.
The only Intel ML chip ever designed for Intel packaging was River Crest.

H100 packaging is being dual sourced now.

Yea, not at Intel.
There are actual, relevant OSATs, you know?
Like ASE.

This attitude could be the difference between 2.0B in revenue vs 2.5B in revenue from MI300.

They'll get a lot more than $2.5B from it lmao.

 

[H433x0n]

They'll get a lot more than $2.5B from it lmao.

That’s not possible with CoWoS constraints. I was quoting AMD’s MI300 projections (2B in 2024) from their earnings call but okie dokie.

 

[adroc_thurston]

That’s not possible with CoWoS constraints

Jfc there are OSATs that also provide chungus 2.5D slabs.
Come on.
ASE exists and AMD loves them to death.

 

[H433x0n]

Jfc there are OSATs that also provide chungus 2.5D slabs.
Come on.
ASE exists and AMD loves them to death.

If they want to be packaging constrained for all of 2024 and not explore potential alternatives for dual sourcing such as Intel or Samsung then that’s their right I suppose.

Also I’m fairly certain Nvidia’s dual source isn’t ASE, the overflow from TSMC has been going to ASE for a while now. It may not be Intel doing the packaging but it’s definitely outside of the usual ecosystem they’ve been using. Based on Intel’s earnings call that mentioned new packaging customers I had assumed it was Nvidia but it could be an incorrect assumption. I also saw it mentioned on SemiWiki so I assumed it was true.

 

[adroc_thurston]

If they want to be packaging constrained for all of 2024 and not explore potential alternatives for dual sourcing such as Intel or Samsung then that’s their right I suppose

Intel has no usable packaging options for MI300 and Samsung big chongus carriers barely got qualified.

I had assumed it was Nvidia but it could be an incorrect assumption.

Intel would be screaming from the roofs if they had it.

 

[H433x0n]

Intel would be screaming from the roofs if they had it.

issuing public press releases would be breaching contract I would guess. I saw it mentioned on SemiWiki so I had assumed it was the case but I guess we’ll find out in a few quarters.

It’s the second post in this thread.

 

[TESKATLIPOKA]

225 IIRC

That looks like a pretty big increase in size compared to 178mm2 of Phoenix.
Difference is 225-178=47mm2 or 26%
Specs wise you gain only 4 extra CU(+33%) and 4 extra cores(+50%).
I think It can be attributed to Zen5 cores, which are bigger than Zen4.

 

[Hitman928]

That’s not possible with CoWoS constraints. I was quoting AMD’s MI300 projections (2B in 2024) from their earnings call but okie dokie.

Lisa Su said they had secured supply for sales beyond their $2B for 2024 projection if the market demand is there.

Edit: Exact quote,

So certainly, for the current forecast of greater than $2 billion, we have adequate supply. But we have also planned for a supply chain forecast that could be significantly higher than that, and we would continue to work with customers to build that out.

 

[SpudLobby]

Intel has no usable packaging options for MI300 and Samsung big chongus carriers barely got qualified.

Intel would be screaming from the roofs if they had it.

This. Lmao there is no way Intel grabbed a packaging contract of that scale and just went silent about it.

Jfc there are OSATs that also provide chungus 2.5D slabs.
Come on.
ASE exists and AMD loves them to death.

Also this. Packaging is more diverse than fabs.

 

[rtxtwt]

It is possible that former "8950X" "8950HS" leaks could be fake

this is a new advertising material for old users of Alienware.

https://twitter.com/x/status/1722384775665930262

 

[adroc_thurston]

It is possible that former "8950X" "8950HS" leaks could be fake

AMD pre-prod parts have no names, just an OPN.