Question Zen 6 Speculation Thread

[Tuna-Fish]

As I mentioned, 384 bits of LPDDR6-10667 will be the dream for Medusa Halo. ~450 GB/s of bandwidth (66% higher than Strix Halo's 273 GB/s).

It would be great for the user, but unless they move to on-package memory, that doesn't sound fun to route. And with on-package memory it would require a large package.

 

[FlameTail]

It would be great for the user, but unless they move to on-package memory, that doesn't sound fun to route.

Yup, on-package memory would be the way to go.

And with on-package memory it would require a large package.

Apple has been putting 512b packages in their Max SoCs for 3 generations now.

 

[LightningZ71]

It would seem straightforward to either stick with monolithic like Strix Point and move to N3e. On N3e, they can finflex the P cores and the C cores appropriately and also add on the speculated MALL cache. With the MALL cache at 16-32 MB, a single speed bin increase on the LPDDR5X ram could support 20CU of RDNA3.5 quite well. That's just two additional WGPs. The next Kraken could grow to 12CU and keep the same core arrangement, but double the L3 in the ccx.

 

[LightningZ71]

The other possibility is changing packaging and arrangement completely. They can stack a Zen6/6c CCD on a Zen4C I/O die with 20CU of RDNA3.5 and MALL cache. Going wider on the iGPU will allow them to keep clocks lower, keeping thermals on the base die under control.

 

[FlameTail]

If Zen6 is launching in 2027, that's a whopping 2.5 years from Zen5 release. It will be the longest gap in Zen history.

 

[CouncilorIrissa]

If Zen6 is launching in 2027, that's a whopping 2.5 years from Zen5 release. It will be the longest gap in Zen history.

2027 launch would be a disaster.
High-end phones will be running circles around STX successor in 1T at that point.

 

[FlameTail]

2027 launch would be a disaster.
High-end phones will be running circles around STX successor in 1T at that point.

Sounds damning when you put it that way.

Apple A series chips are already faster than laptop chips from Intel/AMD, and so it will continue to be that way.

What's more interesting is what Qualcomm/Mediatek are bringing to the table. The former will be relying on the formidable Oryon cores, while the latter will be using the stock Cortex cores which are getting rapidly stronger.

Going by leaks on the Dimensity 9400/8G4, it seems 2025 flagship Android will be trading blows with the latest laptop CPUs from AMD/Intel. This is a historic moment for mobile computing.

 

[Ghostsonplanets]

2027 launch would be a disaster.
High-end phones will be running circles around STX successor in 1T at that point.

It's incredible that Apple and Arm are able to do IPC/Performance gains almost every year while Intel, AMD and QCOM are roughly on a 18 months cadence. And sometimes the latter 3 might not even bring a good performance uplift.

 

[poke01]

I’m still on the 2026 train.

 

[gdansk]

AMD on a 18 months cadence

Nope. They're really far behind now. It's more like 22 months.
Zen 1 -> Zen 2 = 28 months
Zen 2 -> Zen 3 = 16 months (only time they approached 18 months)
Zen 3 -> Zen 4 = 22 months
Zen 4 -> Zen 5 = 22 months
And 22 months from now is? June 2026. If Zen 6 launches in 2027 it'd be a new Zen record for lateness.

This slow rate of iteration will be their downfall in the consumer space if they're delivering 10-15% cores like Zen 5/6. In the server world even ARM has slower cadences so they may remain competitive there.

 

[poke01]

If Zen 6 is 2027 then it should be N2.

 

[FlameTail]

It's incredible that Apple and Arm are able to do IPC/Performance gains almost every year while Intel, AMD and QCOM are roughly on a 18 months cadence. And sometimes the latter 3 might not even bring a good performance uplift.

Why did you lob Qualcomm into that group?

The only info we have is the Dell leaked slide, which I find to be somewhat sketchy (2.5 year gap between Oryon V2 and V3??).

Otherwise, they haven't released an official roadmap, and we don't have past history to extrapolate how they might do in the future.

 

[FlameTail]

If Zen 6 is 2027 then it should be N2.

Even better if it is A16.

 

[FlameTail]

I’m still on the 2026 train.

Careful my friend. Trains can be dangerous!

Spoiler

 

[branch_suggestion]

Medusa Z6-256 (laptop, enhanced INT)
Medusa Z6-512 8c CCX V-C compatible (desktop, enhanced INT)
Venice Z6 16c CCX V-C compatible (classic server, might have enhanced INT)
Venice Z6c 32c CCX (dense throughput)
Z6LP, low power island core for low demand tasks to avoid lighting up the big cores, used where makes sense.
That is what I'm hoping for but who knows.

 

[Jan Olšan]

It's incredible that Apple and Arm are able to do IPC/Performance gains almost every year while Intel, AMD and QCOM are roughly on a 18 months cadence. And sometimes the latter 3 might not even bring a good performance uplift.

ARM yeah, but ARM is still catching up. Apple is mostly clocks now. Since 2020 (M1/A14), they had two IPC-improving new cores - M3/A17 Pro (2023) and M4 (2024). The M3 probably would have launched sooner if it wasn't for 3nm issues, but it's unknown if M4 would have been sooner. Apple's other "yearly" cores only really improved performance through higher clocks (M2, A15, A16).

AMD also had two IPC-improving new cores since 2020 (Zen 3) - Zen 4 (2022) and Zen 5 (2024). So on average it's the same, just more evenly spaced because Zen 4 was not waiting on a process node and came sooner. The IPC gains may even be bigger for AMD (with smaller jump in clocks).
It would be nice if they were able to get new nodes sooner like Apple and throw more money at yearly-designed SoCs even if they have the same architecture. But the notebook market would have to actually buy from them instead of sticking to Intel no matter what. The operating profits of their Client segment doesn't seem to allow this, they would get slightly better chips but lose money.

 

[FlameTail]

ARM yeah, but ARM is still catching up.

I would say ARM has caught up in terms of IPC. Their brand new core: Cortex X925, beats the IPC of the P-core in A17/M3, which also means it's better than Oryon.

 

[Keller_TT]

I would say ARM has caught up in terms of IPC. Their brand new core: Cortex X925, beats the IPC of the P-core in A17/M3, which also means it's better than Oryon.

I was looking at David Huang's post on the comparison of figures between various cores. The Exynos 2400 on Samsung 4LPP+ with the X4 is already pretty neat, and if X925 holds up, then it'll definitely be nudging into M3 territory.
It's impressive how much Apple have pulled ahead. More and more apps are offering native ARM for Windows, and Linux already has excellent ARM64 support. In a year, the final frontier for WoA will be gaming.

 

[FlameTail]

In a year, the final frontier for WoA will be gaming.

Bet Nvidia ARM SoC will crush in that sector.

A lot of the game issues with Snapdragon X Elite is not due to the ARM CPU, but due to the Adreno GPU- which is lacking in both the software (drivers) and hardware (architecture and feature set). Nvidia's SoC will have their excellent GPU, with their impeccable drivers.

 

[Joe NYC]

If Zen6 is launching in 2027, that's a whopping 2.5 years from Zen5 release. It will be the longest gap in Zen history.

What is the origin of "Zen6 is launching in 2027"?
Or "Just asking questions?"

 

[FlameTail]

What is the origin of "Zen6 is launching in 2027"?

Kepler L2 has said he's heard AIBs have been informed that Zen6 client is 2027.

 

[Joe NYC]

Kepler L2 has said he's heard AIBs have been informed that Zen6 client is 2027.

That's a relief.

If Kepler said Zen 6 is definitely shipping 2025-2026, I would be worried.

 

[Mahboi]

Medusa Z6-256 (laptop, enhanced INT)
Medusa Z6-512 8c CCX V-C compatible (desktop, enhanced INT)

Would love that, but seriously, how do you do "enhanced INT" on the same multi use core?
If they put all the Zen6c cores as server cores and beeeg Zen6 cores for DT, different story, but...doesn't seem to be the goal yet.

 

[CouncilorIrissa]

I would say ARM has caught up in terms of IPC. Their brand new core: Cortex X925, beats the IPC of the P-core in A17/M3, which also means it's better than Oryon.

Let's wait until someone has actually run SPECint on it. ARM's slides are about as trustworthy as recent AMD ones.

 

[MS_AT]

Out of curiosity what do people mean by enhanced INT performance? Enhanced memory sub-system performance? Better branch prediction? Different decode scheme?