Question Zen 6 Speculation Thread

[Panino Manino]

Sadly SMT is anti-hype now, everyone's avoiding it. We'll never get even our SMT4 fun.

"I know engineers that are against SMT and they all cowards".

BTW, I don't understand why AMD don't have a single desktop chip with Dense cores.

 

[marees]

I don't understand why AMD don't have a single desktop chip with Dense cores.

Is that because desktop dense cores are for losers ?

 

[jpiniero]

"I know engineers that are against SMT and they all cowards".

BTW, I don't understand why AMD don't have a single desktop chip with Dense cores.

Not useful for DIY gaming.

 

[SpudLobby]

Sadly SMT is anti-hype now, everyone's avoiding it. We'll never get even our SMT4 fun.

This owns and I hope SMT dies in client. It most likely will. Intel is very likely done with it in client and Arm/Apple/QC won’t bring that in.

 

[gdansk]

This owns and I hope SMT dies in client. It most likely will. Intel is very likely done with it in client and Arm/Apple/QC won’t bring that in.

I kinda like it personally since my old desktop chips become "server" parts
But I guess more cores would work even better.

 

[marees]

This owns and I hope SMT dies in client. It most likely will. Intel is very likely done with it in client and Arm/Apple/QC won’t bring that in.

Bring back bulldozer ??

 

[adroc_thurston]

Bring back bulldozer ??

They did.
A510 called.

 

[SarahKerrigan]

They did.
A510 called.

That's a stretch. No shared L1, no shared frontend structures.

e6500 is a lot closer to Bulldozer (or rather a strange kind of inverse-Bulldozer) than A510 is.

 

[adroc_thurston]

No shared L1, no shared frontend structures.

Yeah but Bulldozer derivatives axed that, too (for the most part).
A510 looks like a slightly more independent in-order Excavator.

 

[SpudLobby]

I kinda like it personally since my old desktop chips become "server" parts
But I guess more cores would work even better.

Honestly this is like the most reasonable case I’ve ever seen that isn’t just screeching about free nt and hand waving verification and modern heterogeneous scheduling etc.

Sorta respect it.

But the AMD SMT4 stuff was so funny. Just totally missing the forest for the trees about where the industry is going and why.

 

[SpudLobby]

Yeah but Bulldozer derivatives axed that, too (for the most part).
A510 looks like a slightly more independent in-order Excavator.

Hmmm yeah might be, I’m not familiar. The A510 shared thing was super funny.

 

[SarahKerrigan]

Yeah but Bulldozer derivatives axed that, too (for the most part).
A510 looks like a slightly more independent in-order Excavator.

e6500 had both L1's shared, as well as complex-int, vector, and FP - but had no shared frontend. Kind of an odd footnote in clustered multithreading's history.

 

[adroc_thurston]

e6500 had both L1's shared, as well as complex-int, vector, and FP - but had no shared frontend

Yeah, but that's basically a reverse Bulldozer.
A510 is more similar to Excavator instead.

 

[SarahKerrigan]

Yeah, but that's basically a reverse Bulldozer.
A510 is more similar to Excavator instead.

Like I said, an inverse-Bulldozer.

Never performed particularly well either, IME.

 

[adroc_thurston]

Never performed particularly well either, IME.

Never touched funky embedded POWER, unfortunately, my lands were all ARMv5/6 horrors of yore.

 

[Tuna-Fish]

That's the trajectory but crossover timeline seems off.
Maybe a comboPHY for DT/luggable with two platform at once?

Why would you ever do that when you can just provide two different IODs?

My expectation is that LPDDR6 will come first to some Zen6 mobile APU released late to the cycle, with Zen7 (or whatever) following supporting only it released after.

 

[adroc_thurston]

Why would you ever do that when you can just provide two different IODs?

Kinda ruins the idea of maximum commonality across the entire client lineup.
Like they folded desktop into client to do away with that specifically.

 

[dr1337]

Not useful for DIY gaming.

I would still rather have 16/12 mildly slower cores instead of 8/6 mildly faster ones at the same price for my work PC. And with the dense chiplets they should be able to sell single CCD 16c for a price very similar to the single CCD 8C chip, or 12 small cores for the price of 6 big ones right now.

 

[CakeMonster]

Chipset predictions? Everything points to AM5 and DDR5, so the same PM.21 architecture in that case? That would be kind of crazy to have no evolution until 2028 with Z7...

 

[Chicken76]

Wouldn't SMT3 make more sense before SMT4?

 

[Bigos]

Yeah, SMT4 makes no sense and SMT3 makes 2x as much sense.

 

[Tuna-Fish]

Wouldn't SMT3 make more sense before SMT4?

Yes. But the problem is, SMT1 (that is, just not having SMT) makes even more sense.

 

[gdansk]

With 384 threads in a single socket now it seems really unnecessary.
If Zen 6 increases core counts it is a good time to remove SMT.

 

[adroc_thurston]

With 384 threads in a single socket now it seems really unnecessary

Yeah it's kinda optional there.

If Zen 6 increases core counts it is a good time to remove SMT

Oh they're not removing SMT.

 

[SarahKerrigan]

Not sure why the hate for multithreading.

For small cores it gets forward movement when one context is stalled for Free*. Nice for real-time, too.

For large cores it gets you high utilization of functional units for Free*.

* Terms and conditions apply, but the costs are pretty small. Marvell called it a 5% area impact for TX3.