Question Zen 6 Speculation Thread

[soresu]

UE5 is very very very brand new.

*Lumen and Nanite are brand new, the old pipelines are still there which facilitates UE4 project migration/compatibility.

Nanite still doesn't even cover all the animation solvers of the old geometry pipeline though - morph targets are missing at least, and I think skeletal/bone solvers are still not production ready as of 5.7.

 

[Joe NYC]

Very true, but they CAN average 1 baby/month for 9 months!

But what's the 1% low FPS on that? I mean BPM (babies per month)?

 

[Tuna-Fish]

I think if system architects went at a game engine from the ground up with the explicit purpose of scaling with the number of cores, it is totally possible.

This has been done, multiple times. It's not even that hard. The problem is that it has very real costs.

A system where there is ultimately a single thread driving things and only uses other threads to offload work into is much easier to understand, test and bugfix than using fully distributed scheduling. And the time that QA needs is time that your game is not on the market. Plenty of people have built such engines as tests or hobby projects, but big games still keep getting built around a single core thread. The FPS that you get is not the key thing that the game developers are trying to optimize.

 

[OneEng2]

A system where there is ultimately a single thread driving things and only uses other threads to offload work into is much easier to understand, test and bugfix than using fully distributed scheduling.

I can certainly belief this.

I have also architected systems that did highly threaded work loads (several hundred to 1000 threads) in production critical applications (manufacturing test for major OEMs).

I have never attempted a game engine, so it is possible that this is a more difficult task.

Still, it will have to be the future IMO. IPC and clock scaling will not be increasing nearly as quickly as cores (and hasn't been for some time now). Either games will stagnate with IPC and clock scaling, or they will figure out how to effectively manage lots of threads at a low cost point.

 

[inquiss]

I can certainly belief this.

I have also architected systems that did highly threaded work loads (several hundred to 1000 threads) in production critical applications (manufacturing test for major OEMs).

I have never attempted a game engine, so it is possible that this is a more difficult task.

Still, it will have to be the future IMO. IPC and clock scaling will not be increasing nearly as quickly as cores (and hasn't been for some time now). Either games will stagnate with IPC and clock scaling, or they will figure out how to effectively manage lots of threads at a low cost point.

How many FPS do we need exactly? Like, games have inherently serial processes, but let's say that wasn't true. Do we need 100 core processors running at full tilt, for 1000fps. What are we chasing here?

 

[BorisTheBlade82]

How many FPS do we need exactly? Like, games have inherently serial processes, but let's say that wasn't true. Do we need 100 core processors running at full tilt, for 1000fps. What are we chasing here?

One potential benefit is energy efficiency, if you can employ lots of low clocking cores instead of chasing high frequencies for just a few of them.

 

[blackangus]

So this is a great conversation, but clearly way outside the scope of the Zen 6 thread.

Should this conversation be spit into a thread about how Amdahls law applies to realtime gaming applications?

 

[CakeMonster]

I think its an interesting discussion and as long as it doesn't go completely off the rails, I think it applies well enough to Zen6 in consoles and PC's. We'll be stuck with the next gen consoles until 2034 at the regular 7yr interval and right now it looks like they might be delayed so possibly more than 10 years from today.

 

[randomhero]

I'm quite surprised that AMD went back to distributed scheduler. Have they found some bug in Z5s unified one that lowered performance and had to go back to drawing board? And that would throw off roadmap by several quarters ? Or they found out that they can make Z6 go really brrrrrrr that it was worthwhile to go back distributed ?

One day we will read interesting post mortem on Z6, hopefully.

 

[511]

I'm quite surprised that AMD went back to distributed scheduler. Have they found some bug in Z5s unified one that lowered performance and had to go back to drawing board? And that would throw off roadmap by several quarters ? Or they found out that they can make Z6 go really brrrrrrr that it was worthwhile to go back distributed ?

One day we will read interesting post mortem on Z6, hopefully.

For Higher Fmax

 

[adroc_thurston]

Or they found out that they can make Z6 go really brrrrrrr that it was worthwhile to go back distributed ?

fmax is nice.
FP schedulers were distrubuted even in Zen5 anyway.

 

[Jan Olšan]

9 pregnant women can't make a baby in 1 month. Some tasks have limitations to how parallel they can go

With pipelining they can absolutely achieve 1 baby per month rate.

 

[OneEng2]

How many FPS do we need exactly? Like, games have inherently serial processes, but let's say that wasn't true. Do we need 100 core processors running at full tilt, for 1000fps. What are we chasing here?

I think it is all about doing more with the hardware (and power envelope) available). I just think that clock and IPC scaling is over for all intents and purposes.... or at a minimum, it will never again be like it once was.

One potential benefit is energy efficiency, if you can employ lots of low clocking cores instead of chasing high frequencies for just a few of them.

Exactly.

This is also where server DC applications live. Run more cores at lower clocks and you can do more in the same power window.

 

[Doug S]

Okay, I take it back! Zen 5 didn't show the improvement "some" were hoping for.

So the "Zen 5 40%+" people can maybe achieve their dream, if they are willing to wait until the final iteration of Zen 6 in 2028.

 

[adroc_thurston]

if they are willing to wait until the final iteration of Zen 6 in 2028.

There's just one iteration of Zen6 in 2026.
2028 is Zen7.

 

[inquiss]

There's just one iteration of Zen6 in 2026.
2028 is Zen7.

Won't things like halo be out later? It takes a while to get the whole lineup out....

 

[adroc_thurston]

Won't things like halo be out later?

it haez the CCD.

 

[inquiss]

it haez the CCD.

Sure, but doesn't mean the doc will be out in 2026. Also the monolithic parts might come out at different times. If I understand you, you're saying that zen 6 comes out then, but I guess I'm saying (and I think the previous poster) that parts that use a zen 6 core may well be staggered, and significantly so if this gen is anything to go by?

 

[adroc_thurston]

but doesn't mean the doc will be out in 2026

mean the CPU IP is the exact same.

that parts that use a zen 6 core may well be staggered, and significantly so if this gen is anything to go by?

Zen5 dumped out the entire product stack across 3-ish quarters?

 

[inquiss]

mean the CPU IP is the exact same.

Zen5 dumped out the entire product stack across 3-ish quarters?

The last point is a good one. The stack released then. So zen 6 is a 2026 release with products on the same ISH cadence through summer 26 and early 27?

 

[marees]

The last point is a good one. The stack released then. So zen 6 is a 2026 release with products on the same ISH cadence through summer 26 and early 27?

Medusa halo & medusa premium could be 2028

Medusa point 1 is early 2027

Medusa point 3 is late 2027

Medusa point 2 depends on TSMC making a more cheaper / efficient node available

 

[SteinFG]

2+2, 2+4, and 4+4 core combination seems too much. MDS3 will be cut, I think. No one wants a Quad-core with copilot+ cert for the price of an 8-core R7 8840U. And even if someone wants that, just get the bad bins of 2+4 die

 

[yuri69]

Zen5 dumped out the entire product stack across 3-ish quarters?

Wut? Strix June 2024 - Shimada July 2025.

Zen 6 product portfolio even broader. Keep dreaming about 3Qs. We will be lucky if the entire stack is out at Q3 2027.

 

[adroc_thurston]

Strix June 2024 - Shimada July 2025

The last relevant part was KRK1 in Q1'25.
Which is 3Q for all.

So zen 6 is a 2026 release with products on the same ISH cadence through summer 26 and early 27?

across 27 for other stuff.

 

[OneEng2]

Wut? Strix June 2024 - Shimada July 2025.

Zen 6 product portfolio even broader. Keep dreaming about 3Qs. We will be lucky if the entire stack is out at Q3 2027.

Seems like every processor release, people stack improvements to grossly over estimate performance, and ignore past cycles to predict remarkable release schedules.

After release, a great sigh of disappointment erupts.

Rinse and repeat.