Question Zen 6 Speculation Thread

[soresu]

The reason behind Zen 4 slow sales was mainly the horrible premium asked for AM5 + DDR5 compared to AM4 + DDR4.

Yes, this is also part of why Zen1 sold like hot cakes, because it was well behind the initial PC market transition to DDR4.

 

[Win2012R2]

Shared L2

Bad for server CPUs, such as Zen ones.

I'm not saying you can't launch both at the same time, but the theoretical time you can do that is later than the theoretical time you launch without X3D because the process takes time. It's that simple.

Launching new Zen without X3D at least announced to be available very soon is a very bad launch - it's that simple. That matters far more than when technically first chip becomes fully made.

Nvidia must be doing 80 and 90 class in batches, not perfectly parallel - so why won't they announce them separately? Because it would be very dumb commercial move.

 

[blackangus]

Bad for server CPUs, such as Zen ones.


Launching new Zen without X3D at least announced to be available very soon is a very bad launch - it's that simple. That matters far more than when technically first chip becomes fully made.

Nvidia must be doing 80 and 90 class in batches, not perfectly parallel - so why won't they announce them separately? Because it would be very dumb commercial move.

I don't think its horrible, seems to have worked for AMD in the past. The Nvidia comparison isn't relevant as its apples to oranges. The 5090 and 5080 are the same general manufacturing process.
Zen X3D relies on the base Zen chip. So for the X3D to be completed the Zen chiplets have to be available and then have the X3D chips bonded after. They means the plain Zen will always be ready before the X3D version. And a company is crazy to sit on stock of plain zen when they could be selling it. If a product is ready then sell it. You dont see Apple sitting on their compute product line up until every model is available, neither does Intel. So I think the industry precedence is clear. NVidia still hasnt launched the whole 50x0 series....
Not sure why you would hold AMD to different standards than what the other big players do?

 

[Win2012R2]

And a company is crazy to sit on stock of plain zen when they could be selling it.

Was AMD crazy to sit on RDNA4 stock till March?

 

[Gideon]

Bad for server CPUs, such as Zen ones.

Yeah, it looks to be mostly the case for server CPUs. However, AMD is drifting the server and client cores more and more apart with every generation. Client workloads really seem to favor a small number of fatter, area-inefficient cores anyway (4-8), coupled with dense cores for throughput (see Strix Point, or anything Intel, Apple, Qualcomm, or ARM is doing).

These cores don't make too much sense on most server products as well, due to area bloat. On client chips, in comparison, it's a rather small area loss compared to the GPU and other fixed-function blocks.

What I'm trying to say is that AMD already has Zen 5 and Zen 5c (in multiple versions with 256-bit and 512-bit wide vector units). I wouldn't be too surprised if they end up with a "fat" Zen 7X core on top of these 2. And then it wouldn't be that much of a stretch to offer different cache hierarchies for mobile and server.

 

[LightningZ71]

Shared L2 caches in server provessors is so bad that IBM went ahead and designed a whole Power processor generation that shared all the L2 caches among the cores in one dynamic virtual L3 pool. I haven't seen any of their customers complain about it too loudly though...

 

[Win2012R2]

I haven't seen any of their customers complain about it too loudly though...

Have you ever met one? They are making super expensive mainframe level hardware for 0.01% of the market and understand well what kind of software runs on it, the clients also understand it.

PS5 cuts lots from L3, that may work for a console, but not for general purpose CPU used by 99%.

Shared L2 is terrible for multi tenant instances because your neighbor can affect you by running non-cache friendly code, hyperscalers dictate this thing, it's not going to happen.

Client workloads really seem to favor a small number of fatter, area-inefficient cores anyway

"Client" these day is also a server - I've got a few apps open but Windows reports 4k threads, imagine running a game and some of the background process totally trashes L2, this won't affect just one neighbour cores, but whole CCX - one thread will be enough to trash it.

 

[LightningZ71]

In fact, I do actually talk to two different data center techs whose job is the care and feeding of two different late model IBM Power systems on a fairly regular basis. These systems have many tenants each and they have told me that they don't see any performance issues of note. I will deliberately not go into detail on what exactly they do as I am not at liberty to share that, but one of the systems is handling obscene amounts of transactions on a truly massive data set from many different sources.

 

[Win2012R2]

These systems have many tenants each and they have told me that they don't see any performance issues of note.

one of the systems is handling obscene amounts of transactions on a truly massive data set from many different sources

Do they allow a guy paying ten bucks using stolen card to run their custom code on those?

I reckon they don't. It's a unique product for unique enterprise environments, not a mass produced thing that a gazillion of different software will run with very easy access to anybody.

 

[Gideon]

"Client" these day is also a server - I've got a few apps open but Windows reports 4k threads, imagine running a game and some of the background process totally trashes L2, this won't affect just one neighbour cores, but whole CCX - one thread will be enough to trash it.

Umm, I don't believe it's that simple.

Snapdragon X also has the same 4k threads in windows and Macos has similar amount, and these ARM CPUs have shared L2 only.

Find me a single real world usecase where it trashes client performance on those CPUs but not on x86 ones.

While this is anecdotal, I have a M1 Pro as well as a 6850U (Zen 3) on top of my windows Desktop. I've run lighter games on both having docker, webserver stuff, bloaty IDEs and multiple browsers and electron apps running. Never had an issue.

I find it hard to believe all Apple and Qualcomm chip designers are blithering idiots unable to grasp (nor simulate) basic MT usecases like that. That they just add 10+ cores to their mobile CPUs with shared L2, unable to think of this.

 

[OneEng2]

Clearly a forum full of tech savvy cpu enthusiasts is NOT the best place for Intel and AMD to get product management advice from .

 

[Makaveli]

Clearly a forum full of tech savvy cpu enthusiasts is NOT the best place for Intel and AMD to get product management advice from .

They have people making alot of money on the payroll for this

 

[Win2012R2]

Snapdragon X also has the same 4k threads in windows and Macos has similar amount, and these ARM CPUs have shared L2 only.

So in the case of problem with software (which is validated in app stores before allowing it to run) that will cause problems to 1 (one) person at the time who had chosen to use that software.

In server situation bad code (possibly maliciously) will affect lots of other people whose server apps also likely to server many more other people.

See the difference?

P.S. Obviously in mobile segment reducing area is of big importance for power reasons, and cost of silicon, so of course they want to cut SRAM that does not scale.

 

[Thibsie]

They have people making alot of money on the payroll for this

Which clearly a lot more efficient 😁

 

[511]

Clearly a forum full of tech savvy cpu enthusiasts is NOT the best place for Intel and AMD to get product management advice from .

Just hope no actual product manager for Intel/AMD is on AT we never know 🤣

 

[marees]

Was AMD crazy to sit on RDNA4 stock till March?

Yes & no

You can argue this both ways

 

[Joe NYC]

The reason behind Zen 4 slow sales was mainly the horrible premium asked for AM5 + DDR5 compared to AM4 + DDR4.

Which also says that performance of Zen 4 without V-Cache was not compelling enough to pay the premium.

It also forced AMD to offer deep discounts.

 

[511]

Raptor Lake was better at that time as well and people on LGA1700 had a good upgrade path

 

[Win2012R2]

Yes & no

You can argue this both ways

Was AMD crazy to delay RDNA3 in order to sell remaining stocks of RDNA2?

Delays for non-technical reasons happen all the time - the ultimate goal is to make maximum money, not ship 2 months earlier.

 

[blackangus]

Was AMD crazy to sit on RDNA4 stock till March?

Depends on their reasoning which I am not privy to. If there were issues or they believed they would gain a very strategic advantage the no.
If they did it just to build more stock.... then yes I would say that is pretty silly.

 

[DrMrLordX]

They have people making alot of money on the payroll for this

Intel may have 20% fewer of them now.

 

[Timorous]

Was AMD crazy to sit on RDNA4 stock till March?

In a vacuum, yes they were.

It only remotely became a good idea due to how badly NV had fumbled the ball and in it could have really bitten them in the bottom if NV had gotten their supply sorted just a month earlier than they seem to have so far.

It absolutely was an edge case because in virtually every other situation this kind of thing backfires massively.

 

[jpiniero]

It only remotely became a good idea due to how badly NV had fumbled the ball and in it could have really bitten them in the bottom if NV had gotten their supply sorted just a month earlier than they seem to have so far.

That's why they did the delay though, they knew that was the case.

 

[burninatortech4]

Depends on their reasoning which I am not privy to. If there were issues or they believed they would gain a very strategic advantage the no.
If they did it just to build more stock.... then yes I would say that is pretty silly.

Rumor I've heard is that Navi 32 needed multiple respins before it was ready. Not sure how much that influenced the delay.

 

[inquiss]

Bad for server CPUs, such as Zen ones.


Launching new Zen without X3D at least announced to be available very soon is a very bad launch - it's that simple. That matters far more than when technically first chip becomes fully made.

Nvidia must be doing 80 and 90 class in batches, not perfectly parallel - so why won't they announce them separately? Because it would be very dumb commercial move.

Is this a good faith argument? Genuine question.

1)Beausee you can parallelise their production. You don't need to make 80s or 90s for the other one to be available. X3D production is inherently serial between the bit that makes the chip and the bit that puts the cache on the chip.

2) there's a lot of capacity to make them. X3D requires packaging that doesn't have capacity for mainstream.