Question Zen 6 Speculation Thread

[OneEng2]

We're still talking about double stack X3D? I can believe it was tested, but I'll say that the only way it would ever launch is if Intel just blows the top off with Nova Lake and AMD needs it to maintain / tie for the gaming crown. I doubt thats the case, but we'll see.

I agree. Much of what AMD does, or doesn't do will likely depend on how competitive Nova Lake ends up being.

Double 3D cache, and lots of skews built on N2 are (IMO) dependent on if AMD NEEEDS to do it or not. If Intel is in the rearview mirror with AMD's Zen 6 on N3P and single stacked X3D, then it is unlikely AMD will spend the cash on COGS when they can just pocket the profit.

Lets talk media engines, Nova Lake will excel in this. Industry leading.

Will AMD match or come close here with Zen6? ( This is important for CPU only setups)

Well, based on what we have today, Zen5 outperforms ARL in multimedia in general, but I wonder how much of that is due to AVX512.

Given more cores (48 vs 24) and support for at least a 256 bit AVX path in Nova Lake, I think it likely NVL will hold an advantage over Zen 6 in multimedia.

Of course, Thread ripper will completely dominate these workloads which are, after all, professional level applications for the most part.

 

[DavidC1]

Counter: If it is so useless, why is AMD doing it? Do they have a stash of extra die space that they needed to get rid of?

Companies and people do things that's questionable. This isn't even that.

Also, marketing. Muh cash is bigger than your cache!!!

Are you asking me why x3d exists at all? this has literally been the case since 2022 when 5800X3D was launched.

Because current X3D is at a point where you still get decent gains. X3D is a feature. Enabling that feature is what gets you to the knee of the curve. After that is a diminishing return.

Of course, Thread ripper will completely dominate these workloads which are, after all, professional level applications for the most part.

HEDT platforms like Threadripper is slower in games and the whole setup much more expensive. Motherboard cost is ridiculous and you need double the amount of RAM, not to mention the CPU is slower in lower thread counts too with higher memory latency. You have much higher idle power for those who care. And you miss out on the ITX market too.

The higher core counts go on mainstream setups, less desired HEDT platform is.

 

[poke01]

Well, based on what we have today, Zen5 outperforms ARL in multimedia in general, but I wonder how much of that is due to AVX512.

Given more cores (48 vs 24) and support for at least a 256 bit AVX path in Nova Lake, I think it likely NVL will hold an advantage over Zen 6 in multimedia.

Of course, Thread ripper will completely dominate these workloads which are, after all, professional level applications for the most part.

Threadripper is irrelevant when it comes to consumer multimedia workloads.

I’m referring to the media engines, ie Quicksync and AMDs VCN. Will VCN in the the flagship Zen6 CPU be as good as Intels Quicksync in Nova Lake? I don’t think so but hey AMD may surprise us.

 

[yottabit]

Threadripper is irrelevant when it comes to consumer multimedia workloads.

I’m referring to the media engines, ie Quicksync and AMDs VCN. Will VCN in the the flagship Zen6 CPU be as good as Intels Quicksync in Nova Lake? I don’t think so but hey AMD may surprise us.

Are you concerned with encode or decode capabilities?

I never understood the fascination with QuickSync. I remember using it a fair amount back in the Sandy Bridge through Haswell era. Since then I use NVENC for quick encodes and CPU for quality encodes.

I’d love to see the Venn diagram overlap of people that know enough about computers to even be encoding video but decide they want a system without a lot of cores or a discrete GPU

 

[poke01]

Are you concerned with encode or decode capabilities?

both

I never understood the fascination with QuickSync. I remember using it a fair amount back in the Sandy Bridge through Haswell era. Since then I use NVENC for quick encodes and CPU for quality encodes.

I find Quicksync to be cheaper ie than buying a whole GPU. Its also more efficient and good for media PC setups.

I’d love to see the Venn diagram overlap of people that know enough about computers to even be encoding video but decide they want a system without a lot of cores or a discrete GPU

its a niche within a niche but I like media engines so I am biased.

 

[511]

Here is multimedia comparison

画質比較 (2025.08)

 

[igor_kavinski]

Here is multimedia comparison

画質比較 (2025.08)

Great link!

So generally ARC and Geforce are doing better. AMD hasn't improved things much with RDNA4 still.

 

[StefanR5R]

According to the rumors, Zen 6 should reduce memory latency. So +50% LLC, decently higher IPC and core clocks and on top lower memory latency are in the ring. Looks good regarding to gaming.

If that all happens:
To be seen.

As you explicitly mentioned the rumored increase in cache size per core complex, I assume you mean "main memory latency" by "memory latency" (as opposed to general average memory latency which would include the effects of cache hit rate and cache latency). If so, then no: There is not much, if any, left to carve out by improvements of the processor-integrated memory controller and of the processor-internal data fabric. The big stumbling block WRT memory latency which is DRAM technology will affect Zen 6 the same as it does Zen 5.

 

[StefanR5R]

Much of what AMD does, or doesn't do will likely depend on how competitive Nova Lake ends up being.

Much? Pricing, and maybe adding or omitting one or another minimally adjusted SKU.

Double 3D cache, and lots of skews built on N2 are (IMO) dependent on if AMD NEEEDS to do it or not.

2-hi cache stack --> won't play into competitive consideration as long as the popular benchmarks are with first person shooters rather than strategy-type games.
skews = SKUs? if so --> The decision of what to build on N2 is made long before anybody, notably the two CPU makers, know who has the slightly longer bar in benchmark XY.

The tale of vendor 1 leans back to see what vendor 2 does and then outdoes him by just a necessary little amount is almost entirely false, things like pricing excepted. This tale addresses tactics at best, but in no way does it address strategy.

 

[OneEng2]

HEDT platforms like Threadripper is slower in games and the whole setup much more expensive. Motherboard cost is ridiculous and you need double the amount of RAM, not to mention the CPU is slower in lower thread counts too with higher memory latency. You have much higher idle power for those who care. And you miss out on the ITX market too.

The higher core counts go on mainstream setups, less desired HEDT platform is.

The subject was multi-media encoding. IMO, consumers do so little of this that anything that is out there today is already more than enough. Me for example. I make a video a few times a year for family events. The final render generally takes about 5 minutes. Edits are instantaneous as is playing a section of the video in the editor.

Higher core counts without the HEDT bandwidth may work OK with some applications; however, for those that encode for a living, I don't see them skimping on a consumer CPU vs buying a 96 core HEDT. The productivity gain makes the ROI a no brainer to any manager.

Threadripper is irrelevant when it comes to consumer multimedia workloads.

Please explain what a consumer multimedia workload is? I use PowerDirector. My Ryzen 5600g works just fine and is pitiful compared to even entry level Zen 6 or NVL. It wouldn't even cross my mind to shell out for NVL 52 core in order to save a few minutes of rendering time a few times a year. YMMV.

I’m referring to the media engines, ie Quicksync and AMDs VCN. Will VCN in the the flagship Zen6 CPU be as good as Intels Quicksync in Nova Lake? I don’t think so but hey AMD may surprise us.

It is my understanding that these technologies are for low end systems without a graphics card. If that is the case, what difference would this make to flagship Zen 6?

 

[511]

It is my understanding that these technologies are for low end systems without a graphics card. If that is the case, what difference would this make to flagship Zen 6?

it will matter in Mobile CPUs not in Desktop

 

[Tuna-Fish]

There is not much, if any, left to carve out by improvements of the processor-integrated memory controller and of the processor-internal data fabric.

On all existing AMD Zen CPUs, the link from CCD to IOD is not full width and there needs to be serdes on both sides. This adds extra latency to every memory request, and is the main reason why, when using the same memory and with the same settings, AMD platforms have ~10ns worse memory latency than 14th gen intel.

For Zen6 and Zen7 the technology for connecting to IOD is changed to one that allows a much wider connection, people are assuming that this will result in a full-width connection that will then reach parity with Intel.

 

[StefanR5R]

On all existing AMD Zen CPUs, the link from CCD to IOD is not full width and there needs to be serdes on both sides.

You forgot the monolithic APUs.

(Also Strix Halo. But its potential gain from having gotten rid of the serdes is masked by LPDDR latency.)

 

[LightningZ71]

The Monolithic APUs face other restrictions that add latency.

 

[511]

They face signal integrity issues going higher than 8000MT/s at least strix halo did

 

[basix]

As you explicitly mentioned the rumored increase in cache size per core complex, I assume you mean "main memory latency" by "memory latency" (as opposed to general average memory latency which would include the effects of cache hit rate and cache latency). If so, then no: There is not much, if any, left to carve out by improvements of the processor-integrated memory controller and of the processor-internal data fabric. The big stumbling block WRT memory latency which is DRAM technology will affect Zen 6 the same as it does Zen 5.

Correct, I had main memory or DRAM latency in mind.

On all existing AMD Zen CPUs, the link from CCD to IOD is not full width and there needs to be serdes on both sides. This adds extra latency to every memory request, and is the main reason why, when using the same memory and with the same settings, AMD platforms have ~10ns worse memory latency than 14th gen intel.

For Zen6 and Zen7 the technology for connecting to IOD is changed to one that allows a much wider connection, people are assuming that this will result in a full-width connection that will then reach parity with Intel.

A few nanoseconds are probably accurate if you look at Alder Lake / Raptor Lake numbers compared to Zen 4 / 5 numbers. 10ns might be a bit optimistic.

But you should gain additional latency due to increased DRAM and data fabric clock rates. Zen 6 is rumored to bring 6400 MT/s compared to 5600 MT/s. Even if sticking to current IFOP SerDes, that would still result in +14% DRAM and fabric clock rates and Zen's memory latency responses very well to increased clock rates.

So overall 10ns DRAM memory latency reduction are not that unreasonable to assume. In best case it could be even more. I wouldn't mind to go back to my Haswell-E system memory latency of <50ns (could even push it <45ns for some bechmark runs)

 

[Kepler_L2]

Zen 6 is rumored to bring 6400 MT/s compared to 5600 MT/s.

DDR5-8000 is the official support

 

[CakeMonster]

Lots of enthusiasts will be looking to buy new memory then at least. We finally have 64GB sticks now too.

 

[basix]

DDR5-8000 is the official support

The new platform delivers DDR5-8000+, correct. But e.g. X870 and Zen 5 do support DDR5-8000 as well. I assume 6400 is the JEDEC verified support with 1:1 clock ratio. Higher speeds might be different.

 

[Doug S]

DDR5-8000 is the official support

In DIMM, or MRDIMM? Because I wasn't aware of JEDEC planning to support DDR5-8000 anytime soon. Or ever.

 

[adroc_thurston]

In DIMM, or MRDIMM

How the hell are you asking this question when Venice .pptware was leaked in full?
Its 8k for normal DIMMs and 12.8k for the buffer gimmick ones.

 

[adroc_thurston]

But e.g. X870 and Zen 5 do support DDR5-8000 as well.

They don't.
Zen5 is 6k with 6.4k in limited SKUs for hyperscaler deployments.

So overall 10ns DRAM memory latency reduction are not that unreasonable to assume.

no the fabric complexity is going up.
I suggest not even thinking about the latencuck wank and sticking to idk, a 9900k.

 

[basix]

Adroc.... as charming as always

 

[Darkmont]

In DIMM, or MRDIMM? Because I wasn't aware of JEDEC planning to support DDR5-8000 anytime soon. Or ever.

The spec has been over 8000 since at least 2022 and was updated to 9200 last summer.

 

[Darkmont]

I don’t know how much they can lower the main memory latency when the IF latency is already 6.5 nanoseconds at 2000 MHz FCLK. If anything, with larger caches and Gear 2 likely being required for client DDR5-8000 support, I’m expecting an increase in system memory latency. Fancy $$$ packaging on client is good for 1. Power 2. Bandwidth 3. Latency in that order, and we already know the new fabric is 32B R/W bandwidth, and it may not be widened, so just expect power to go down.