And still no Desktop APU since ZEN+.(Except Renoir which is only for OEMs). Maybe AMD wants you to buy non APU cpus so that you also buy a GPU from them.
APUs and GPUs are relatively low-margin for AMD on a per-wafer basis. It takes a lot of silicon to produce a product which sells for a not a lot of money.
Consoles exercised options to increase their orders by 50% from the expected amounts during 2020, and reportedly consoles are now consuming up to 80% of AMD's wafer allocation. So they are in a defensive mode to maximize their returns on their remaining wafers.
Monolithic dies need to have IO for memory/etc, which doesn't shrink well on modern nodes, so it consumes a lot of space. The iGPU also consumes a pretty large amount of space. On chiplet based CPUs, all of that is pushed off to a separate, older process which is not in shortage (GloFo 12nm/14nm), so only the CCDs themselves are on 7nm. This is a very efficient way to use silicon compared to the monolithic APU dies.
GPUs are not a great situation for AMD either. NVIDIA is using a cheaper, worse process, but as is customary they have much better silicon engineering and have matched AMD's core efficiency despite the worse process - the only real advantage AMD has is that they use less memory which saves them about 30 watts out of a 300W budget. And at the top end they are exceeding their performance, plus they have tensor cores to accelerate DLSS and provide additional performance, plus more raytracing performance. So AMD is paying more for an expensive process and hasn't managed to beat NVIDIA. NVIDIA is essentially forcing AMD to cut their own throats on margins in the GPU segment, if they want to remain competitive. There is very much a reason that AMD is no longer wildly undercutting NVIDIA like they used to and it has nothing to do with mindshare/etc, AMD has been using more expensive technology just to try and stay performance competitive for years now.
As such, APUs and GPUs have been the products that get the shit end of the stick. To wit: as far as cutdowns go, a 5600X die is 80mm2, they sell that for $300 MSRP, and they keep most of that for themselves (let's say $50 chip cost/manufacturing/packaging, $50 for retailer/distribution, they make $200 a chip). A 6800 die is 520mm2, the MSRP is $580, and they have a significant BOM cost, and they have to incorporate a profit margin (even if slim, it's not great being an AMD partner). So let's say $100 chip cost/manufacturing/packaging, $300 BOM, $50 for the partner, $50 for the retailer, they make $80 per card. But it uses 6.5 times as much silicon per chip, so they make the equivalent of $12 on this GPU versus $200 on the CPU! It's literally an order of magnitude more profitable to make the chiplets than to make a GPU.
Or an APU, Cezanne is 175mm2 for an 8-core die, they can sell that for maybe $50-75 more than the iGPU-less chip, so $350-375. The chip costs them let's say $25 more, so they make $250, but it's 2.2x as big as the chiplet.
Just made up numbers pulled out of my ass, but you can see the math heavily disfavors making APUs and GPUs. Which is why AMD has actually been shunting production away from those throughout 2020, it is not a coincidence that they were announced in January and didn't show up until maybe May in actual products, and then you had reports from partners that they were getting their allocated (basically confirmed and accepted) orders pushed back a quarter or more.
(this eventually ended up with the orders from August getting pushed back into November if you read the series of threads from XMG)
TSMC just doesn't have enough capacity to go around, and right now they are an effective monopoly, people basically pre-write the obituary of products that choose processes that aren't TSMC. And unfortunately the situation is only getting worse this year and next year - there are products that have been waiting for consoles that will immediately soak up all that capacity. Intel is launching GPUs and potentially laptop CPUs (?) on TSMC now. Automakers have been clamoring for more and have had to shut down lines because they can't get enough chips. NVIDIA is reportedly moving a few of their high-end consumer products to TSMC late this year to supplement their Samsung capacity and allow a "refresh" like 3000 Super with higher performance at the top. Apple is moving all of their laptop and desktop products from Intel products (made at Intel foundries) to TSMC, so I expect they will continue to squat the 5nm node for a longer period of time (desktop/laptop will probably stay 5nm and phones/tablets move to 3nm once that's available), and this will echo right down the conga line, products from other brands that would have transitioned to 5nm will be stuck consuming 7nm instead.
Everyone thought it was some snub from TSMC that NVIDIA switched to Samsung, like there is a persistent narrative (which nobody credible will actually get behind) that NVIDIA tried to demand a discount from TSMC and got told to pound sand. It is now very clear that TSMC was not going to be able to deliver the capacity (and certainly not at a price anyone liked) and it's now looking like NVIDIA switching to Samsung was a masterstroke. They've been able to ramp Ampere faster than Pascal (3080 is at the same % marketshare in Steam Hardware Survey as 1080 was at the equivalent point in its lifecycle and the market is >2x as big) and AMD can't get product on shelves. And now NVIDIA can supplement their Samsung capacity with a few higher-performing chips on TSMC at the top of their stack to help broaden their production capacity.
It really says it all that AMD hasn't officially launched Milan yet. That's a chiplet-based product (so, good utilization of their wafers) but higher-margin than consumer products. So basically strictly better margins than consumer dies in all respects. But they just can't deliver enough capacity yet to be credible. That is the next thing on AMD's docket, it'll be launching in March. So far hyperscalers have samples and probably small production runs and that's it. Unlike NVIDIA, unlike Intel, AMD has no second-source to run to, it's TSMC or bust (or go back to shitty 12nm/14nm Zen+ chips from GloFo).
(which, btw, long story short, is why "only Zen+ APUs on desktop"! those APUs are manufactured at GloFo on an older node and thus do not compete for precious TSMC capacity.)
This article is just a marketing teaser to try and divert attention from the Tiger Lake-H launch. AMD had a core count advantage over Tiger Lake-U which they could point to (and which gave them a performance and efficiency advantage in highly-threaded tasks), Tiger Lake was faster per core but AMD had more of them. Well, now Intel has 8 cores coming up in March, and they're still faster cores than Zen2-based Renoir. AMD has been sitting on these Zen3-based Cezanne chips as long as possible so they could preserve their margins on chiplet products, with Tiger Lake-H coming they don't have a choice but to at least flash them around to show what they can do.
The APUs launching in March are gonna be a paper launch. Milan is where all their wafers are going in Q2, it's way higher margin and they need to hit that market (especially if they can beat out Ice Lake-SP since they have been making good inroads in server). But they need to show the flag in mobile because Intel is about to surpass Renoir in the mobile-workstation market, not just the ultrabook market. They have to show their successor now and hope to keep people on the hook long enough. It will probably be Q3 if not Q4 before these Cezanne processors show up in any number - there are just too many other companies who booked TSMC's capacity and too many products within AMD that need to go first before APUs.
It's no coincidence these articles are showing up 2 months before the launch event. Hardware Unboxed had one too. It's AMD's last chance to try and get benchmarks against 14nm 8-core chips and 10nm 4-core chips that they can make an argument against. This is "look at the shiny, please pay no attention to that Intel product launch that reviewers will have in-hand within a month or so". And Intel laptops have hard-launched to a much greater extent than AMD.
I wonder what Sony and MS are paying AMD for console chips to get them to push off Milan and their APUs, it's got to be considerable. Milan probably would have at latest launched alongside Vermeer if not before, so they have delayed it 6+ months. It must have been an absolute pile, Milan is a high-margin high-volume product.
tl;dr: fab capacity is fucked, maybe things will be better in 2022
That money is supposed to build a new US-based fab, among other things. I’m very glad to see that, but it does seem hard to scale up manufacturing capacity at the highest performance levels.
Perhaps there will always be a single “best” fab globally, which makes all the top-end products for everyone and commands way higher prices than everything else, while inspiring hand-wringing about risk concentration.
Huge thanks for such a comment. This is why I love HN.
As about consoles AMD likely just losing a lot of money on them simply because both Sony and Microsoft most likely have decade-long contracts that heavily limit AMD margins. There are reasons why Nvidia don't try to compete on this market.
> There are reasons why Nvidia don’t try to compete on this market.
Maybe not at the high-performance end, but the Nintendo Switch is powered by an Nvidia Tegra X1, and they won that contract largely by proving the concept in their own Nvidia Shield “microconsole”.
It really seems AMD's problem right now is fab capacity - as they can't sell their existing inventory fast enough. The new gen GPUs are double-whammied by the resurgence in cryptomining but it's telling that even the 5000-series CPUs still seem to be selling out (even the lowest tier 5600X).
From AMD's earnings report (released yesterday), their margins are healthy and enterprise segment more than 2X in revenue on QoQ basis. Certainly indicates that Milan is their top priority given the high profit margins.
https://www.anandtech.com/show/16455/amd-reports-q4fy-2020-e...
Will fab capacity really be better in 2022? Can't be worse than this year but with Apple buying up all the next-gen 3nm capacity, Intel also sourcing to TSMC, it just seems a persistent issue until more fabs come online (which takes years).
I'm sure it's coming. This market timing is frustrating, though. Especially as it gives mobo vendors (looking at you ASROCK) more slack with getting all their BIOS/UEFI of last-gen models up to date to support them, ever.
The most probable thing is that APU are low margin products and if they released an 8-core APU like in mobile they would have cannibalized the non APU parts
