Question Zen 6 Speculation Thread

[igor_kavinski]

But it's like living in Detroit in the 60/70s watching everything turn to crap.

I wouldn't worry.

CP/M. Dead and no one cares.
Amiga. Dead and almost no one cares.
Pre-OS X Macs. Dead.

You know what's alive? Cobol and Fortran and mainframes.

Why?

Business software.

I'll probably expire before x86 is forgotten and only found out about by those looking at the history of computing.

 

[gdansk]

Apple mobile is just 600MHz lower than Strix now.

Yep. They'll probably use that in the next few years. But that's a 13% clock rate increase combined with already 30% higher SPECint 1T. Do you think Zen 6 stands a chance in 1T?

 

[soresu]

The thing that worries me most about x86 is COST - not in client mind you, that's very reasonable, but in servers it's totally outrageously too expensive - like 15 grand for a single bloody chip? Pricing is way out of whack, power usage is also going up to crazy levels, so again - cost, but ongoing.

Yes, but how many companies are selling actual ARM server chips rather than VM instances in the cloud run on their own semi custom hardware?

I know that Ampere is one, but there aren't a great many others since so many tried and failed to get ARM servers going before the software ecosystem was ready to truly support it.

 

[adroc_thurston]

Do you think Zen 6 stands a chance in 1T?

Yeah they should start with not regressing on freq gen on gen in mobile.

 

[gdansk]

I wouldn't worry.

CP/M. Dead and no one cares.

It's not a worry of mine but I remain befuddled why people act like ARM isn't already better just because Apple doesn't have games.

I'll keep a ghetto x64 machine for old games but future spending is lamentably going to locked down Apple crap because AMD and Intel simply cannot compete (or refuse to do so, I can't tell the difference).

 

[soresu]

I'm curious to see if ARM Ltd, AMD or Intel will pick up on this Blueshift Memory's new Cambridge architecture.

The claims seem outlandish, but even if a fraction of it is true is has serious potential.

 

[gdansk]

Yeah they should start with not regressing on freq gen on gen in mobile.

Okay. Well it doesn't. I'll give you a prediction: The fastest ARM laptop CPU (probably from Apple) will be about 42% higher SPECint 2017 1T score than Zen 6 laptop (whenever it launches).

 

[Win2012R2]

But it's like living in Detroit in the 60/70s watching everything turn to crap.

This deal is getting worse all the time... and where it isn't?

 

[adroc_thurston]

than Zen 6 laptop (whenever it launches).

No such thing. It's just 'client'.

 

[Win2012R2]

The claims seem outlandish

Never heard of it till now... reminded me of Optane initial claims... without good trusted 3rd party verification this all very hard to believe

 

[soresu]

It's not a worry of mine but I remain befuddled why people act like ARM isn't already better just because Apple doesn't have games.

Games are still a pretty significant factor though.

Rockstar made an insane amount of money on GTA5, and GTA6 is just around the corner.

Yet for some reason they seem to have cut off anything older than GTA3/San Andreas from Mac OS or iOS

If Apple's management had one tenth the sense they should have they would get their engineers to punt a Vulkan GPU driver for Mac OS, because it would make it soooo much easier to plug in the benefits of Valve et al's work on Proton/DXVK/Zink/etc and turn Mac OS into a truly viable gaming platform with relatively minimum effort on their part.

 

[adroc_thurston]

Mac OS into a truly viable gaming platform with relatively minimum effort on their part.

Well it had some games. Then Apple killed 32b apps.
Apple goes fast and loose with platform de-featuring so it's never, ever gonna be a viable gaming platform.

 

[Win2012R2]

Yes, but how many companies are selling actual ARM server chips rather than VM instances in the cloud run on their own semi custom hardware?

The only big buyers of ARM server chips are hyperscalers and they are big enough to buy their own - they don't want to pay somebody else's 50-60% gross margin if they can avoid it! TSMC is lucky in this equation because they can't be avoided really.

I've written software that is portable to ARM, however "cloud providers" are such a rip off that it's far cheaper to use your own x86 hardware than their "cheaper" ARM offering. Ampere sells dedicated boxes but it's way too much - might as well get last gen Genoa and get great perf, AVX-512, no aggro compatibility wise.

What's the resell value of an ARM server? It's more or less zero because nobody wants them, memory in it is the only salvageable part.

Yet for some reason they seem to have cut off anything older than GTA3/San Andreas from Mac OS or iOS

They don't want to pay Apple 30%?

And MacOS - not a good market for selling games like that.

 

[soresu]

reminded me of Optane initial claims

Not even remotely related beyond what they claim as perf boosts.

Optane is just a branded variant of PCM (phase change memory), which until very recently has been woefully unsuited to scaling in the same vertical fashion as NAND flash has since the 2010s to get past the area scaling limitations of the tech.

PCM has potential, but it's still fundamentally just a faster memory with theoretically more write cycles and lower scaling.

Blueshift's Cambridge architecture on the other hand is more like a radical shift in how the internal data structures of system memories are laid out so that the CPU/GPU utilisation for memory operations and overall data movement are minimised.

The idea in of itself seems to have merit, but even if it does it's going to take years before we see it in any of the big name processor architectures unless they were tracking this from the earliest funding rounds.

 

[soresu]

Well it had some games. Then Apple killed 32b apps.
Apple goes fast and loose with platform de-featuring so it's never, ever gonna be a viable gaming platform.

True, both its greatest strength and weakness in one.

Does that include emulation?

 

[igor_kavinski]

Yes, but how many companies are selling actual ARM server chips rather than VM instances in the cloud run on their own semi custom hardware?

Do companies really want to invest in software development for a particular ARM CPU in the cloud that may not have a successor in 5 or 10 years? Sounds too risky.

 

[Win2012R2]

Not even remotely related beyond what they claim as perf boosts.

Yeah, but that's the point - claiming x1000 like Optane did got one thing in common: BS.

If their stuff really works then file patents and prove it to the world with all the details, otherwise it's snake oil - even if it is not such disclosing behavior should be treated as such as a matter of principle.

Do companies really want to invest in software development for a particular ARM CPU in the cloud that may not have a successor in 5 or 10 years? Sounds too risky.

It's ok if you write (use) cross platform stuff, as anybody really should anyway.

If ARM was 3-5 perf/$ or ideally 10 then people would move over, but it isn't - getting 30-40% with risks + porting costs only works for large volumes and if you got that then you'll have better price on x86 - better even owning your own hardware saves a lot, and the only alternative to that is Ampere which to me seems on their last legs.

 

[soresu]

Do companies really want to invest in software development for a particular ARM CPU in the cloud that may not have a successor in 5 or 10 years? Sounds too risky.

Yup, ThunderX3 comes to mind.

Marvell made big promises about ThunderX4, but it never materialised.

And of course we all know what happened to Qualcomm's Falkor successor.

It is definitely going to be interesting to see if Qualcomm wants to try their luck punting Oryon to the server market with ARM Ltd making it relatively easy for new players to take up Neoverse and make their own semi custom designs.

 

[soresu]

claiming x1000 like Optane did got one thing in common: BS

Oh 100% that was BS, but in Blueshift's case the wording is pretty explicit about "up to x% improvement" terminology, so I was never under any impression that this was anything but the most rare edge cases for the uses they talk most about like big data analysis, and anything that usually requires a lot of heavy data movement under the current system.

 

[gdansk]

Games are still a pretty significant factor though.

If we must accept that Zen CPUs will be offering smaller gains and longer generations then there is less incentive to buy a new CPU for games.

That's fine. I won't buy 'em. Small gains means the old machine is good enough longer. But the upgrade money is gonna go somewhere. Nvidia probably. Apple maybe.

 

[soresu]

even if it is not such disclosing behavior should be treated as such as a matter of principle

True, but then again we all entertain that sort of thinking about Soft Machines VISC, Nuvia's Phoenix etc.

Some will pan out, and others will disappoint, always going to be the case.

 

[soresu]

If we must accept that Zen CPUs will be offering smaller gains and longer generations then there is less incentive to buy a new CPU for games.

If game devs were doing their jobs properly the CPU would have ceased to be a bottleneck years ago with improvements to multithreading to take better advantage of larger core counts.

One of the chief benefits of Vulkan (and DX12 possibly? 🤔) over OpenGL is its significantly better design for multithreading rendering engines for games/content creation, and yet it seems to have been largely unused up to this point sadly.

 

[Win2012R2]

the wording is pretty explicit about "up to x% improvement" terminology

Yeah, that's used not to go to jail for false claims...

If we must accept that Zen CPUs will be offering smaller gains and longer generations then there is less incentive to buy a new CPU for games

But it's all GPUs now - this is the only thing that still gets 50-70% uplift, I am far more excited about 5090 than 9800X3D, even though to drive that beast even at 4k will need at least 7800x3d really, will be VERY interesting to see if even 9800x3d becomes bottleneck for it.

One of the chief benefits of Vulkan (and DX12 possibly? 🤔) over OpenGL is its significantly better design for multithreading rendering engines for games/content creation, and yet it seems to have been largely unused up to this point sadly.

Well, that did improve a lot from DX11 era - main problem with DX12 is that devs have to do a lot more low level stuff, and how many devs have time to do it? It's not a wonder that many game companies now abandoning their game engines to standardise on Unreal Engine 5.

 

[gdansk]

If game devs were doing their jobs properly the CPU would have ceased to be a bottleneck years ago with improvements to multithreading to take better advantage of larger core counts.

PhDs can seldom write working MT code that actually has a world to synchronize. Game devs do not stand a chance. They have to ship. Properly multithreaded games are the future and always will be.

 

[soresu]

Yeah, that's used not to go to jail for false claims...

Or at least to avoid getting sued by investors in any case.