Question Zen 6 Speculation Thread

[Cheesecake16]

Since Zen 6 is not a major ground up design from Zen 5 (just a tweak), I see no way SMT will be more effective.

Why?
Considering the amount of changes that happened with Zen 5, there are likely some pretty trivial performance gains that could be had... one big one is that fast/slow pathing hazard with the VEC INT which makes all SIMD instructions a minimum of 2 cycles instead of 1 cycle... Another one is the front end effectively being unable to use both decode clusters without SMT...
These are behaviors that are likely to be fixed/improved in Zen 6...
Then what makes you so sure it's "just a tweak" of Zen 5... Zen 2 wasn't "just a tweak" of Zen 1, Zen 3 wasn't "just a tweak" of Zen2, Zen 4 wasn't "just a tweak" of Zen 3, and Zen 5 certainly wasn't "just a tweak" of Zen 4... the 2 cores that were the most similar was Zen 3 and Zen 4 but...:
1) While it's very clear that Zen 4 is derived off of Zen 3, that in no way makes it a simple "tweak"... Adding an ISA as all encompassing as AVX512 is no easy feat and on top of that AMD improved the branch predictor, made the op cache massive and capable of feeding even more instructions to the back end, made the vector register file 512b natively, doubled the L2 cache size, etc.
2) AMD still got fairly impressive performance gains for what was not a complete redesign of the core which is not always needed every generation...

 

[Cheesecake16]

Turns out it's not true.

I mean... if both our testing can't pick it up and the SOG says that the core functionally can't do said behavior, then I would argue that you aren't gong to be able to get Zen 5 to use both decode clusters at the same time... you are going to need a very convoluted set of instructions which are frankly so rare that it's functionally doesn't matter...

 

[adroc_thurston]

you are going to need a very convoluted set of instructions which are frankly so rare that it's functionally doesn't matter...

Well yeah.
But then again, AMD cores live off their opcache.

 

[Cheesecake16]

Well yeah.
But then again, AMD cores live off their opcache.

Oh sure, I would agree with that for Zen 4 and 5 especially... tho I do wonder if AMD will go back to a 64K L1i in Zen 6 but with a higher associativity then what Zen 1 had...

 

[adroc_thurston]

tho I do wonder if AMD will go back to a 64K L1i in Zen 6 but with a higher associativity then what Zen 1 had...

Yeah, Zen is due for an i$ bumps.
Don't think Z6 has it though. AMD's generally really conservative with SRAM allotments compared to ~everyone else.

 

[basix]

Unfortunately we’re not at a point yet where AI is actually a net benefit for engineers.

Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity

metr.org

“We conduct a randomized controlled trial (RCT) to understand how early-2025 AI tools affect the productivity of experienced open-source developers working on their own repositories. Surprisingly, we find that when developers use AI tools, they take 19% longer than without—AI makes them slower.”

AI-Tool benefits are very use case dependent. I do not use AI for actual production code. But I use it for supportive "helper" code. When I have to analyse data e.g. "this is my input data, write me a short Python script to visualize and analyse this and that" and make a few iterations on the result is often much faster than coding it by yourself.
This works because the scope of the task is rather narrow and the result (aka the visualization or presentation of statistical analysis) does not have to be perfect, just good enough so that I can understand what is going on. That will not work on my production code, especially because embedded software with hardware and system requirements are involved, which are not directly derivable from the code.

But it is always the same:
Use the right tool for the right job. I like using Python but more and more I go back to simple Excel sheet editing, when it is a one-off thing with limited amount of manual effort. This is simply faster. But the AI-generated Python code can get very close to that and exceeds the Excel editing speeds (as weird as Excel and speed in the same sentence sounds ) when you want to add things as an afterthought or you have to re-run tests and therefore have to analyse data twice.

 

[adroc_thurston]

AI-Tool benefits are very use case dependent. I do not use AI for actual production code. But I use it for supportive "helper" code. When I have to analyse data e.g. "this is my input data, write me a short Python script to visualize and analyse this and that" and make a few iterations on the result is often much faster than coding it by yourself.
This works because the scope of the task is rather narrow and the result (aka the visualization or presentation of statistical analysis) does not have to be perfect, just good enough so that I can understand what is going on. That will not work on my production code, especially because embedded software with hardware and system requirements are involved, which are not directly derivable from the code.

But it is always the same:
Use the right tool for the right job. I like using Python but more and more I go back to simple Excel sheet editing, when it is a one-off thing with limited amount of manual effort. This is simply faster. But the AI-generated Python code can get very close to that and exceeds the Excel editing speeds (as weird as Excel and speed in the same sentence sounds ) when you want to add things as a second thought or you have to re-run tests and therefore have to analyse data twice.

Yeah but those are all boring nothingburger toy usecases that in no way justify gorillions in ML big iron capex.

 

[MS_AT]

What they fail to capture however is how much good the developers felt on interacting with AI tools so emotionally, they probably did better and their stress levels probably improved a ton.

Yes, I can call the bot a moron when it says something uterrly stupid without having to wonder if it had a bad day... it won't take offence, yet...

one big one is that fast/slow pathing hazard with the VEC INT which makes all SIMD instructions a minimum of 2 cycles instead of 1 cycle

Have you tried to see how many nops between the int vadds are necessary to see 1 cycle latency. I think SOG mentions that (inserting nops) as an idea but I don't have copy with me atm, so maybe I remeber wrongly.

Yeah, Zen is due for an i$ bumps.
Don't think Z6 has it though.

Do you know if they plan to bump op cache capacity instead?

I think the loop buffer could be probably added back but then I am not sure if its needed seeing opcache should basiacally cover the same niche.

 

[igor_kavinski]

Yes, I can call the bot a moron when it says something uterrly stupid without having to wonder if it had a bad day... it won't take offence, yet...

Being around lesser minds can give one a false sense of superiority. Such little strokes to one's ego are never a bad thing. Well, as long as one stays with the lesser minds and does not venture into the company of geniuses where they would feel like they don't even deserve to be alive

 

[Josh128]

Since Zen 6 is not a major ground up design from Zen 5 (just a tweak), I see no way SMT will be more effective.

AMD's Mike Clark has already said better lithography will allow for a lot of room for improvement in Zen 6.

"But we need that flexibility in our roadmap, and it makes sense. But still that was really hard to try to control having the two technologies and the features, and a feature that looks great in 3nm not looking so great in 4nm because of the power impact of the not-as-efficient transistor and how it affects the floorplan. Normally, we do the architecture in one, and then we port on the next one, and then you have a lot of time to deal in the floor plan with the two technologies. [..] It was just really challenging. But that gives Zen 6 a lot of room to improve. -- AMD's Mike Clark , July 22,2024

 

[511]

AMD's Mike Clark has already said better lithography will allow for a lot of room for improvement in Zen 6.

Wait wut it's the same freaking lithography lol unless he meant the process

 

[blackangus]

Being around lesser minds can give one a false sense of superiority. Such little strokes to one's ego are never a bad thing. Well, as long as one stays with the lesser minds and does not venture into the company of geniuses where they would feel like they don't even deserve to be alive

I would say... thinking you have a "greater" mind is where the false sense of superiority comes from. Great minds are mostly humble as they know the vast scope of what they dont know.
I have rarely met an arrogant genius.

 

[itsmydamnation]

Wait wut it's the same freaking lithography lol unless he meant the process

Of course he means process. Given the very loose meaning of lithograph, a chemical process to create an image. Ya gotta use common sense. He didn't say something like "because of better lithography machines" which you could then take as High NA or something .

 

[Josh128]

Of course he means process. Given the very loose meaning of lithograph, a chemical process to create an image. Ya gotta use common sense. He didn't say something like "because of better lithography machines" which you could then take as High NA or something .

Correct. Lithography comes from two greek words- "lithos" meaning stone, and "graphe" which means to write or draw. Imagine trying to pompously distinguish "a type of drawing" and the "process of producing that drawing" as an actual point worthy of being discussed. This is why I did not respond to it, it was an ignorant post that deserved to be ignored, lol.

 

[Tigerick]

As speculated in custom APU thread, here is the SoC die of upcoming Medusa Premium with 4+8 CPU and 24MB L3 cache. Medusa Premium SoC will bundle with AT2 (68CU) and 192-bit GDDR7 to make up as APU for Xbox X2.

AMD will bundle Medusa Premium SoC with AT4 (24CU) and 192-bit LPDDR6 to make up APU for PC market.

There are some features missing in AT2 compared to AT4, I will let you guys think for yourself. People coming to this forum are supposed to be tech savvy and know the difference between the two GPU...And that's why AT2 won't be used as APU for PC market and AT3/4 won't be used as discreet GPU.

 

[igor_kavinski]

People coming to this forum are supposed to be tech savvy and know the difference between the two GPU...

People also come here to learn so please spoonfeed

 

[StefanR5R]

please spoonfeed

AMD codenamed their GPUs in loving remembrance of the AnandTech web site, and you as an AT forum dweller pay it back by not knowing the differences between AT2, AT3, and AT4? You should be BANNED!

(Well, I need to be banned too. Obviously, the iGPU dedicated to Xbox doesn't need some of the features of a contemporary personal computer iGPU. But I don't care enough for video game consoles to know for certain which features those are.)

 

[marees]

the iGPU dedicated to X-Box doesn't need some of the features of a contemporary personal computer iGPU

You mean the iod chip (of the medusa premium) ?

The gpu chiplets are identical between console & PC

 

[ToTTenTranz]

There are some features missing in AT2 compared to AT4, I will let you guys think for yourself. People coming to this forum are supposed to be tech savvy and know the difference between the two GPU...

Missing features? Are you saying AT2 and AT4 have a different featureset? Aren't they both GFX13?

And that's why AT2 won't be used as APU for PC market and AT3/4 won't be used as discreet GPU.

According to the original leaker for these codenames and specs, these are both wrong.
He says the modularity of the ATx chiplets and their ability to connect to either a display head or IOD + CPU chiplets exist specifically so that AMD can release discrete GPUs with AT2-4.

I also think AT2's Magnus is indeed for the PC market, it's just that it'll be PCs with a Xbox logo. It'll run windows and/or be able to sideload win32 apps.

 

[igor_kavinski]

I also think AT2's Magnus is indeed for the PC market, it's just that it'll be PCs with a Xbox logo. It'll run windows and/or be able to sideload win32 apps.

They gonna turn gaming PCs into consoles!

Their plan must be to turn EVERY ONE of them into a console.

Steam better take evasive, defensive and offensive measures ASAP!

 

[marees]

They gonna turn gaming PCs into consoles!

Their plan must be to turn EVERY ONE of them into a console.

Steam better take evasive, defensive and offensive measures ASAP!

Even your car (with smart LG/Samsung TV) is a console as it can play xcloud (streaming off a virtually partitioned AT0)

 

[igor_kavinski]

Even your car (with smart LG/Samsung TV) is a console as it can play xcloud (streaming off a virtually partitioned AT0)

I hate being cooped up in a car.

 

[ToTTenTranz]

They gonna turn gaming PCs into consoles!

Their "This is an XBox" meme shows exactly that intention.

 

[OneEng2]

Why?
Considering the amount of changes that happened with Zen 5

I believe that AMD is filling up the additional transistor budget with more cores (and then some) in Zen 6 over Zen 5. This doesn't leave much transistor budget on the table to make big changes with. ....... at least this is my theory and SWAG.

AMD's Mike Clark has already said better lithography will allow for a lot of room for improvement in Zen 6.

I am guessing that it will mean more cores, and higher clocked cores .... and very little IPC increase.

 

[Markfw]

My 9755 has twice the cores as most of my Zen 4 boxes, and while 500w seems like a lot, and the run a little slower than Zen 5, the compute my GHZ is more. 128 cores at 3 ghz vs 64 cores @ 3.5, I will take the 9755 over 9554. (500 vs 400w the 500 is better used for twice the cores)

Point being that more cores in Zen 6 most likely has the same type of winning power/GHZ.