you gave the reason they have the capacity but not money as of now to even ramp things
you gave the reason they have the capacity but not money as of now to even ramp thingsTheir 2.5D start capacity is even more of a joke.
that's because they be broooooooke as hell.
Well they don't.you gave the reason they have the capacity
Yeah that's chicken egg stuff.but not money as of now to even ramp things
Yep. Interesting to observe how quickly a PCIe 6.0 -> 7.0 shift might take place.
they can always pack at Intel who has a bunch of advanced packing capacity lying around lol so much that they stopped the expansion.
I think there's some confusion on that whole topic.
We are talking consoles. "Real" consoles are a piece of mass market hardware that aims to sell 100+ million in a life cycle that usually spans 5+ years. Steam Deck is far far from that, PC handhelds are far far far from that.
With this kind of mass market hardware you don't want to use off the shelf hardware. You want to cost optimize the hardware upfront to have as little cost as possible during the whole life cycle. You want your hardware to bespoke or you are doing it wrong. Microsoft got burned by that with the original gen XBox which they couldn't stop manufacturing quickly enough for that very reason.
For AMD it's free money. They get all the money to do all that necessary upfront work. With it they can hire more staff for that if necessary.
Microsoft is doing it wrong if it is selling too few consoles for the upfront cost to be worth it. Games sell consoles. Microsoft bought plenty of popular gaming IPs they could look at making good/better/any use of before considering launching their next console gen.
What is the max AM6/AM5 APU we are talking of here ?AMD has used monolithic for mobile, to keep the costs down, but is already expected to go with chiplets with high end mobile.
So there is a resistance posed by big die sizes (yields) especially on more costly process, and increased complexity of validating a huge die.
We will see what emerges, and also when.
That pool of talent is not endless. There are opportunity costs. Up to half a trillion projected AI market size.
Do you think AMD would rather have them hand crafting 4 bespoke console dies rather than introduce some efficiency in design, commonality, sharing of IP and as result less staff dedicated to it, and then use the saved human resources in datacenter GPUs, where the big bucks are?
Microsoft, in particular, might welcome lower upfront design costs that could emerge from greater commonality and sharing of resources. Because, who knows if Microsoft will be able to sell enough units to justify the expense.
It would not surprise me if this was not the pitch AMD used, since Microsoft was (apparently) shopping around quite extensively.
What is the max AM6/AM5 APU we are talking of here ?
- 12 core zen 6 CPU CCD ?
- 16+ CU RDNA 5 IOD ??
Ok. I thought you were referring to PC HandheldsThe discussion was about next gen consoles, which are much bigger die sizes that APUs.
Which is why I think it is possible they may go chiplet route...
Not cost. Packaging volume.AMD has used monolithic for mobile, to keep the costs down
TSM is the king of yields. Forget about anything having any yield issues on TSM nodes.So there is a resistance posed by big die sizes (yields) especially on more costly process, and increased complexity of validating a huge die.
who tf cares, GPGPU guys live in their own separate world where crack is infinite, and pipes aplenty.That pool of talent is not endless. There are opportunity costs. Up to half a trillion projected AI market size.
YES.Do you think AMD would rather have them hand crafting 4 bespoke console dies
no they don't, it's the same company that commissioned both Van Gogh and Lakefield for a single meme product.Microsoft, in particular, might welcome lower upfront design costs that could emerge from greater commonality and sharing of resources.
Not cost. Packaging volume.
TSM is the king of yields. Forget about anything having any yield issues on TSM nodes.
who tf cares, GPGPU guys live in their own separate world where crack is infinite, and pipes aplenty.
YES.
because they're gonna ship millions of units.
I get it, 2.5D packaging makes your dick hard, but it ain't happening in the next console gen.
InFO is wafer-level, it's permanently capacity-constrained if you want to ship real real volumes.Companies other than TSMC (such as ASE) can do the this type of packaging (Integrated Fan Out) just fine.
Yield isn't an issue and 2.5D slabs do not make any nodes cheaper.Yield is one issue, cost of advanced nodes (where advanced nodes may not be necessary) is another issue.
That happened in 2020 lol.I am pretty sure AMD client GPU division was raided, to some extend, when bringing up Mi300, Mi300x etc was time critical.
That one died because they thought they couldn't compete.Cancellation of N4c may be related.
Handhelds do not exist. Just boxes.Divide that volume by 4, to get the volume of console / handheld, Sony / Microsoft.
It’s called semicustom for a reason. Sony and MS will use off the shelf AMD IP (read: logic blocks) to build their custom APUs. It’s not like they are developing a SoC where every major block is new for each customer. Think legos.
There is no age of chiplets, the bulk of designs shipped by everyone and anyone are monodies.The question is if (and to what effect) semi-custom evolves in the age of chiplets.
Because it's a console business.Really, it has been pretty much NADA from AMD on semicustom front.
Why do that when a custom SKU does just fine.Many people expected semicustom designs for, perhaps, hyperscalers in datacenter from AMD and AMD delivered exactly zero on this front so far.
Why would it ever?Merger with Xilinx renewed expectations for semicustom
InFO is wafer-level, it's permanently capacity-constrained if you want to ship real real volumes.
Yield isn't an issue and 2.5D slabs do not make any nodes cheaper.
They're win more things.
That happened in 2020 lol.
no they didn't, N33 was explicitly changed into N6 monodie because they considered InFO capacity insufficient for serving the mainstream.AMD already tried with RDNA 3
No they never inteded to ship InFO in mainsteam. Which is why N33 was an N6 monodie.The actual volume shipped was depressed because of bad cycle in GPU sales and RDNA 3 underwhelming performance.
No you can't.The can substitute cheap node for expensive node
Well good news, N3p is N-1 by the time consoles ship.Also, reduce capacity demand on the most advanced node
not how any of that works.Some of the staff get assigned to new active projects, and other projects, in hibernation lose staff.
AMD literally did a small round of layoffs last year. wtf are you on.I don't think AMD is the kind of company (any more) that hires and fires people based on a quarterly report. Intel has become that type of company under Gelsinger.
lmao they swallowed like 6 ML startups in a year.I think AMD will resist hiring too many people, will try to do more with fewer people.
That's more for workloads where you aren't sure what your prefetchers have to grab next. Consoles run a limited suite of software where latency can be hidden through clever optimizations. The things you need/want in a desktop or server may not translate well to a dedicated gaming box that only ever runs a handful of different game engines.Still, at least a small amount of V-cache (even 32MB) would be far better than no V-cache due to the much higher latency of GDDR6/7 RAM used in consoles. The elimination or minimization of CPU idle wait times between accessing data from RAM will help to smooth things over.
That seems very plausible.BTW, here is Olrak's post that he calls schizopost. Year and a half later, it does not look insane...
There is no age of chiplets, the bulk of designs shipped by everyone and anyone are monodies.
Why do that when a custom SKU does just fine.
Why would it ever?
no they didn't, N33 was explicitly changed into N6 monodie because they considered InFO capacity insufficient for serving the mainstream.
No they never inteded to ship InFO in mainsteam. Which is why N33 was an N6 monodie.
Well good news, N3p is N-1 by the time consoles ship.
lmao they swallowed like 6 ML startups in a year.
Yeah.N32 sold for $400 - $500
That's not mainstream.That's as mainstream as you get, while N33 is low end.
I'm not talking imaginary numbers or whatever, just how the product stack was sliced.Just because N32 sold far less than it's potential (if it was a very good generation) does not prove this point.
There's panel-level fanouts but they're 2028+ and not for this console gen.for which you expect zero progress on InFO capacity.
Kind of.AMD has been pioneering in this area
Yeah they did, they postponed 2.5D-fication of their lineup 3 times.and, until now, did not have a very effective and affordable 2.5D solution.
yeah because you have no clue.Is it illogical to expect that AMD will expand it further?
That's about as encroaching as soviets in Africa doing the cold war. Gibs stop, and so will the DC MSS creep.To stop Arm encroachment into datacenter
Super8 already has extensive influence on AMD's big socket roadmap. They do not need to do anything more.to change value proposition for Hyperscalers so that they buy more AMD product and spend less on designing their own.
It's Google, not Broadcom.Even Broadcom is beating AMD in semicustom, which is sad.
No, no matter how many yield kings TSMC's manufacturing process if the die is big Of course, the number of chips that can be taken from the wafer will be reduced.Not cost. Packaging volume.
TSM is the king of yields. Forget about anything having any yield issues on TSM nodes.
who tf cares, GPGPU guys live in their own separate world where crack is infinite, and pipes aplenty.
YES.
because they're gonna ship millions of units.
no they don't, it's the same company that commissioned both Van Gogh and Lakefield for a single meme product.
Come on, keep it clean here.
esquared
Anandtech Forum Director
Is it a semi-custom to AMD hyperscaler…The question is if (and to what effect) semi-custom evolves in the age of chiplets.
Really, it has been pretty much NADA from AMD on semicustom front. Many people expected semicustom designs for, perhaps, hyperscalers in datacenter from AMD and AMD delivered exactly zero on this front so far.
Merger with Xilinx renewed expectations for semicustom, but still zero...
You know that besides raw there's also parametric yield?No, no matter how many yield kings TSMC's manufacturing process if the die is big Of course, the number of chips that can be taken from the wafer will be reduced.
It's no custom at all.Certainly, there was also a MI300C, a customized MI300 product for Azure.