Question Zen 6 Speculation Thread

[adroc_thurston]

Im pretty sure that is how XBox and PS hardware have been handled since AMD has been involved. An initial design fee, then a flat IP /licensing fee per chip sold, manufacturing handled between TSMC and Sony/MS.

yea.

 

[soresu]

And they certainly won't be getting as favorable a deal from AMD this time around.

Why?

Surface bad blood?

 

[LightningZ71]

AMD is not in the same position that they were in last time around. Previously, they were desperate for ANY degree above COGS. Now, anything that they do in semi custom is essentially an opportunity investment. They could put more resources towards their own branded devices, or they could put them towards MS's chip. Each has an expected return on investment. MS has to offer them enough to not only cover their costs, but to also beat the return on doing anything else.

This is why I suggest that a single CCD Medusa Halo would be a good fit. It's a completed design, so there's no new R&D to pay for. Unused IP in the chip can be fused off, reducing licensing costs. So, pay the smaller integration cost up front and buy the rest on a volume discount.

 

[adroc_thurston]

AMD is not in the same position that they were in last time around

Yeah they are.

Previously, they were desperate for ANY degree above COGS

in 2018? lmao.

This is why I suggest that a single CCD Medusa Halo would be a good fit.

could you please stop?
Console parts are ALWAYS bespoke designs.
The sooner you give up, the merrier it'll be!

 

[Joe NYC]

As for the discussion on how price sensitive a gaming console is, the current XBOX retails for about $600. IME, that means the entire BOM needs to be around $200. I would be shocked if the APU is over $100.00 of that.

I don't think so. There is no OEM profit margin. Microsoft does not sell to OEMs for further mark up. Microsoft sells directly, likely at cost or below at first. Later in the life cycle, the costs come down somewhat.

If new XBOX sells for $600, your cost estimates are off by factor of 2x-3x.

 

[Thunder 57]

Why?

Surface bad blood?

Didn't even think of that. No, I'd say because AMD is in a much better position these days and will demand higher premiums for hardware. It doesn't make sense to use fab space on cheap console APU's when it can be used for Epyc or Instinct. Even AMD dGPU's are an afterthought and they make more of of them than they would a chip for the next Xbox (at least I would think so).

 

[adroc_thurston]

I'd say because AMD is in a much better position these days and will demand higher premiums for hardware

That's not how any of that works.
Consoles are a strategic investment that keeps the ISV moat going.

It doesn't make sense to use fab space on cheap console APU's when it can be used for Epyc or Instinct

They're not limited by wafer counts for either of those.
Semicon ain't a zero-sum game.

 

[Joe NYC]

AMD is not in the same position that they were in last time around. Previously, they were desperate for ANY degree above COGS. Now, anything that they do in semi custom is essentially an opportunity investment. They could put more resources towards their own branded devices, or they could put them towards MS's chip. Each has an expected return on investment. MS has to offer them enough to not only cover their costs, but to also beat the return on doing anything else.

This is why I suggest that a single CCD Medusa Halo would be a good fit. It's a completed design, so there's no new R&D to pay for. Unused IP in the chip can be fused off, reducing licensing costs. So, pay the smaller integration cost up front and buy the rest on a volume discount.

I think it is better than 50% probability that new XBOX uses LPDDR5 or LPDDR6, rather than GDDR7. You can have 2x or more memory for the same cost, and they can just go up memory bus width.

Everyone is enamored with AI and AI likes to have a lot of memory.

If the choice is LPDDR, then the design will likely heavily borrow from Strix / Medusa or even share silicon...

 

[Thunder 57]

I don't think so. There is no OEM profit margin. Microsoft does not sell to OEMs for further mark up. Microsoft sells directly, likely at cost or below at first. Later in the life cycle, the costs come down somewhat.

If new XBOX sells for $600, your cost estimates are off by factor of 2x-3x.

They will happily sell at cost or even a reasonable loss and make it up on Game Pass, game sales, accessories, etc.

 

[adroc_thurston]

I think it is better than 50% probability that new XBOX uses LPDDR5 or LPDDR6, rather than GDDR7

lmao.

You can have 2x or more memory for the same cost, and they can just go up memory bus width.

You can't do a wider bus, analog shoreline on N3 is $$$.

If the choice is LPDDR, then the design will likely heavily borrow from Strix / Medusa or even share silicon...

consoles are ALWAYS bespoke silicon with zero commonality with other parts.
Give up.
just give it up already.

 

[Joe NYC]

Console parts are ALWAYS bespoke designs.

XBOX console, XBOX handheld, PS6 console, PS handheld. That's a lot of bespoke designs there.

I think there is a good possibility there will be some sharing of designs to silicon or even chiplets.

If AMD has designed standardized chiplet to chiplet connectivity, why not use it, rather than designing multiple massive monolithic dies?

 

[adroc_thurston]

XBOX console, XBOX handheld, PS6 console, PS handheld

Well handhelds do not exist.

That's a lot of bespoke designs there.

No that's normal for products with 6-7 year lifecycles.

I think there is a good possibility there will be some sharing of designs to silicon or even chiplets.

THEY CAN'T.
Console vendors can't use actual AMD parts or chips or tiles or whatever.

If AMD has designed standardized chiplet to chiplet connectivity

For themselves.

why not use it, rather than designing multiple massive monolithic dies?

Because 2.5d volumes are tiny.

 

[Thunder 57]

That's not how any of that works.
Consoles are a strategic investment that keeps the ISV moat going.

They're not limited by wafer counts for either of those.
Semicon ain't a zero-sum game.

I never said it was zero sum but generally speaking any company is going to prioritize parts with higher profitability. What consoles get them is reliable long term income while making less. AMD can try to make a more lucrative deal this time around as they are in a stronger position. At the same time you don't want to piss off what has been a good partner so they won't go too far IMHO.

 

[adroc_thurston]

but generally speaking any company is going to prioritize parts with higher profitability

That's not how any of that ever works.
AMD's revenues in server are not gated by wafers available, but their competitive positioning.

What consoles get them is reliable long term income while making less

They get them a gigantic share of gaming ISV brains.

AMD can try to make a more lucrative deal this time around as they are in a stronger position.

No because then console vendors would just defect to Nvidia.

 

[Joe NYC]

You can't do a wider bus, analog shoreline on N3 is $$$.

32 GB or more of GDDR7 is also $$$.

BTW, AMD used chiplets in RDNA3, with MALL cache on them, to improve the effective bandwidth. Even though GDDR already has good bandwidth.

Why not use that approach with LPDDR5, with mall, on trailing node. this would be the best of both worlds: big memory sizes, high effective bandwidth, and die size savings on more expensive N3 node.

consoles are ALWAYS bespoke silicon with zero commonality with other parts.

Do you think it is a good approach to spam bespoke parts with zero commonality? Devote 100s of people doing duplicative work?

It doesn't seem like the best idea, given all the opportunities out there and only so much human resources that AMD has or can hire.

 

[adroc_thurston]

32 GB or more of GDDR7 is also $$$.

24G is enough.

BTW, AMD used chiplets in RDNA3, with MALL cache on them, to improve the effective bandwidth. Even though GDDR already has good bandwidth.

AMD's been using MALL since RDNA2 in case you forgot that.

Why not use that approach with LPDDR5, with mall, on trailing node.

Really expensive, a ton of packaging overhead for the most mainstream products alive.
Just don't.

Do you think it is a good approach to spam bespoke parts with zero commonality?

YES.

Devote 100s of people doing duplicative work?

YES.

It doesn't seem like the best idea, given all the opportunities out there and only so much human resources that AMD has or can hire.

they can always hire more.
also, they should pay more.

 

[Joe NYC]

I never said it was zero sum but generally speaking any company is going to prioritize parts with higher profitability. What consoles get them is reliable long term income while making less. AMD can try to make a more lucrative deal this time around as they are in a stronger position. At the same time you don't want to piss off what has been a good partner so they won't go too far IMHO.

I think there is a misconception that the fabless companies are constantly in a trade-off situation where they misjudged demand for multiple products and have to constantly shuffle capacity between them.

I think it is much simple than that. Company, like AMD, places and order for a product based on their projection, and then just takes the delivery, with no changes made. If there is a need more wafers, AMD can first go to TSMC to see if TSMC can accommodate.

TSMC Gigafab concept is designed to have a lot of flexibility to meet such demands.

Only if a product turns out to be a complete dud that does not sell, would there be a desire to cut wafer orders. Overall, I don't think these changes are made on week to seek sales and profit margins of individual products.

 

[Thunder 57]

That's not how any of that ever works.
AMD's revenues in server are not gated by wafers available, but their competitive positioning.

They get them a gigantic share of gaming ISV brains.

No because then console vendors would just defect to Nvidia.

Do you really think Nvidia would even be interested? They half assed the 5000 series. They are riding the AI wave. And who knows what their ARM + GPU product will look like just yet. I'm sure they will want to sell it at a premium though.

 

[adroc_thurston]

Do you really think Nvidia would even be interested

Very.

I'm sure they will want to sell it at a premium though.

Switch 2 silicon surely sells at a huge premium right.

 

[Thunder 57]

Very.

Switch 2 silicon surely sells at a huge premium right.

Switch 2 uses far older/cheaper technology for graphics. Or to put it in words you can more easily understand, Switch 2 uses poverty hardware.

 

[adroc_thurston]

Switch 2 uses far older/cheaper technology for graphics

Yeah but that's Nintendo gimmick.
NV ain't minting any more margin there.

 

[LightningZ71]

Yeah they are.

Their balance sheet is notably better in 2025 than it was in 2018.

in 2018? lmao.

My mistake, "degree" was supposed to be "revenue", and, yes, their balance sheet wasn't that great back in 2018. It is arguable that the volume and revenue for the two game console SoCs was a key factor in them staying cash flow positive, even with low margins.

could you please stop?
Console parts are ALWAYS bespoke designs.
The sooner you give up, the merrier it'll be!

I won't give up the fantasy, first because it annoys you, and second because, unlike back in 2018, when AMD didn't have anything even close to what a console needed, they currently have something much more applicable to the target product in the pipeline.

 

[adroc_thurston]

Their balance sheet is notably better in 2025 than it was in 2018.

Idk how that impacts their semicustom biz.

It is arguable that the volume and revenue for the two game console SoCs was a key factor in them staying cash flow positive, even with low margins.

No they weren't lmao.

I won't give up the fantasy, first because it annoys you, and second because, unlike back in 2018

It doesn't annoy me; just turns this place into more of a petting zoo than it already was.

they currently have something much more applicable to the target product in the pipeline.

They don't?
Consoles need heavily cost-optimized, graphics-focused SOCs with cheapest DRAM BOM available (in $/offchip byte moved).

 

[LightningZ71]

Idk how that impacts their semicustom biz.

It impacts it by allowing AMD to demand increased margins for semi custom projects. Taking them on isn't make-or-break.

No they weren't lmao.

In 2019, semi custom represented 37% of AMD's total revenue (2.5 billion out of 6.7 total). That's enough to be a key factor.

It doesn't annoy me; just turns this place into more of a petting zoo than it already was.

They don't?
Consoles need heavily cost-optimized, graphics-focused SOCs with cheapest DRAM BOM available (in $/offchip byte moved).

Cost includes R&D sunk costs. Spending massive amounts (series x SoC was estimated to costs hundreds of millions in 2018 money) to develop a bespoke design has to be folded back into the cost of each chip. They may move 7 million units a year on average, something that takes hundreds of millions to develop adds a nontrivial cost to each chip.

If the package meets their needs, they can afford higher per unit costs on the back end by not fronting development costs up front.

 

[adroc_thurston]

It impacts it by allowing AMD to demand increased margins for semi custom projects.

No it doesn't, that's not how semicustom BU operates.

In 2019, semi custom represented 37% of AMD's total revenue (2.5 billion out of 6.7 total)

At basically zero opmargin.
It might as well not existed.

(series x SoC was estimated to costs hundreds of millions in 2018 money

That's spread across foundry, EDA, AMD and MS.

to develop a bespoke design has to be folded back into the cost of each chip

Yea and?
Worth every penny for a design tailored to move 40m+ units lifetime at least.

they can afford higher per unit costs on the back end

no they can't.