Question Zen 6 Speculation Thread

[adroc_thurston]

Do you know if there is a write-up somewhere.

Don't think so.

I thought the dual fetch on ICache was to help with that, curious if I read it wrong or they hit a funny problem somewhere.

Eh, AMD cores are opcache-driven anyway.

I guess min 2clk latency for otherwise 1 clk instruction from the "scheduler full problem". 2CLK FADD always would be a nice boost, now it's 2CLK only on happy path, 3CLK otherwise but this is not affected by the "scheduler full problem".

yeah the cheaper's the better.

 

[igor_kavinski]

You mean they should slow down some of them? I mean AMD to match Intel?

Referring to the particular instruction here: https://forums.anandtech.com/threads/the-avx-512-thread.2627775/post-41497331

Hope this is made faster in Zen 6.

 

[ToTTenTranz]

Yes but to be fair it is 11.5+
Just so poors can run FSR4 natively.

Why does it matter whether Medusa Point uses RDNA3 or RDNA5? It's just a display head with video codecs and some GPU acceleration for web browsing. It's more of an evolution of Granite Ridge to be paired with dGPUs than it is to Strix Point.

 

[adroc_thurston]

Why does it matter whether Medusa Point uses RDNA3 or RDNA5? It's just a display head with video codecs and some GPU acceleration for web browsing. It's more of an evolution of Granite Ridge to be paired with dGPUs than it is to Strix Point.

Which Medusa? There's a pile of 'em.

 

[ToTTenTranz]

Which Medusa? There's a pile of 'em.

Point.

 

[511]

Which Medusa? There's a pile of 'em.

Medusa doesn't have Stones 🤣

 

[adroc_thurston]

Point.

Point is the lineup.
There's MDS1/2/3/DT/p/h.
So which one?

 

[511]

Point is the lineup.
There's MDS1/2/3/DT/p/h.
So which one?

i just want to confirm one thing please tell me they removed AI from the name

 

[ToTTenTranz]

Point is the lineup.
There's MDS1/2/3/DT/p/h.
So which one?

The ones with a RDNA3 iGPU.

 

[adroc_thurston]

The ones with a RDNA3 iGPU.

That's like 4 parts.
So which one?

 

[ToTTenTranz]

That's like 4 parts.
So which one?

I don't think any of the Zen6+RDNA3 SKUs are meant to play videogames or run AI on the iGPU because there's AT3 and AT4 for that.

 

[adroc_thurston]

I don't think any of the Zen6+RDNA3 SKUs are meant to play videogames or run AI on the iGPU because there's AT3 and AT4 for that.

8CUs is more than enough to play a ton of simpler games.

 

[Josh128]

i just want to confirm one thing please tell me they removed AI from the name

I'd think its more likely that they double down on it. Ryzen AI² Infinity. Mark it. 🤣 🤣 🤣

​

 

[ToTTenTranz]

8CUs is more than enough to play a ton of simpler games.

You can also play a ton of simpler games on Granite Ridge's 2CUs.

 

[511]

Intel HD 4000 is enough for simple casual game

 

[OneEng2]

Therefore I am not sure how AMD could spend those 5% silicon area better than continue to use SMT. I mean I would expect that whatever Apple does requires more than 5% if we even can reason in this way as I guess its deeply ingrained in the core so cannot be singled out.

Since AMD is "DC First" with their designs, there isn't anything (IMO) to better spend that 5% on. I suspect that apple spends much more space on attaining their high ST performance.

This is a take that focuses on the immediate term and ignores the very real long term degradation of one's own skills by offloading these tasks to an AI.

I won't do it. I have enough problems, I dont need to give myself more.

I disagree. When I went through engineering school, drafting was REQUIRED. It was totally and completely replaced by CAD (thank God).

Prior to the internet (yes, there was a time before that), I looked up chips in an encyclopedia of manuals (a different set for each supplier). Thank God those days are over.

AI doesn't always get things right; however, neither does an internet search or someone else's examples you find. What it does do is consolidate a bunch of information for you. It is just a tool (IMO) to achieve a task faster.

Will Zen 6 SMT be even more effective?

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.

Will Zen 6 improve the cycle latency for all AVX-512 instructions and finally achieve parity with Intel's instruction cycle times?

I would be careful on this one. I am not certain that Intel's implementation that so heats the core is a great goal to reach. Getting the clock cycles down only to find yourself downclocking the core to keep thermals under control seems like a bad idea.

Will Zen 6 make 10,000 MT/s RAM speed possible and attainable by mere mortals?

Not in the desktop or laptop I think. In server, and threadripper, yes.

Will Zen 6 at least try to go for 8000 1:1 RAM speeds?

Seems logical that it would.

Will Zen 6 finally be the architecture that Intel is not able to beat in any benchmark (been waiting SO LONG for that!) ?

No. In the desktop, NVL will best Zen 6 by a good margin IMO for highly threaded apps with low compute threads.

Will Zen 6 finally usher in the era of dual V-cache CCDs, for the love of all that is holy and sacred?

I don't see why they would do this.

Will Zen 6 have an NPU?

I would think so on some variants.

Will Zen 6 be the first Ryzen that beats Intel CPUs in idle power usage?

Nope. I think that while 18A will have great difficulty with hot spots at higher clocks, it will likely create a very efficient core at lower clocks/idle.

 

[511]

I disagree. When I went through engineering school, drafting was REQUIRED. It was totally and completely replaced by CAD (thank God).

i only had one semester and that sucked like i didn't even like it i still have my drawings though saved up even though its been quite some time.

 

[igor_kavinski]

i only had one semester and that sucked like i didn't even like it i still have my drawings though saved up even though its been quite some time.

My dad claims he saved me from drafting because he just "knew" I could never do it and so always discouraged me from pursuing math or engineering.

Personally, I think I would've loved the art of being precise. Doing something really carefully, getting it checked by a teacher or a peer and getting the all clear. Now I just do stuff on my own and have to check it too because I'm not among engineers!

 

[OneEng2]

My dad claims he saved me from drafting because he just "knew" I could never do it and so always discouraged me from pursuing math or engineering.

Personally, I think I would've loved the art of being precise. Doing something really carefully, getting it checked by a teacher or a peer and getting the all clear. Now I just do stuff on my own and have to check it too because I'm not among engineers!

I suppose it did require a certain level of attention to detail .... or your drawing quickly became a mess.

Still, I felt that requiring any kind of artistic skill from a would-be engineer was pure folly .

My daughter had drafting for 2 years in high school (AutoCAD then SolidWorks). She can get around CAD much better than I can.

Arguably, engineers (most of them) manage CAD drawings, or engineer them rather than actually drawing them (another flaw in the College curriculum IMO).

Still, my point is that AI is a tool. I would never tell people that CAD was silly and that we should go back to drafting on paper. AI is a tool that engineers need to learn to use and use effectively.... no different than CAD.

 

[Hitman928]

I suppose it did require a certain level of attention to detail .... or your drawing quickly became a mess.

Still, I felt that requiring any kind of artistic skill from a would-be engineer was pure folly .

My daughter had drafting for 2 years in high school (AutoCAD then SolidWorks). She can get around CAD much better than I can.

Arguably, engineers (most of them) manage CAD drawings, or engineer them rather than actually drawing them (another flaw in the College curriculum IMO).

Still, my point is that AI is a tool. I would never tell people that CAD was silly and that we should go back to drafting on paper. AI is a tool that engineers need to learn to use and use effectively.... no different than CAD.

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.”

 

[jpiniero]

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

But it's a net benefit for stock prices. Which is all that matters.

 

[blackangus]

Maybe you should use it and help improve it for everyone else rather than impede progress by telling others to not even experience it? It's not like if it says "chickens have green blood", people will start believing that. If someone is that clueless, they are going to be doing and believing stupid stuff without AI anyway.

I mean you realize people follow GPS into rivers, ponds, off road, etc

I went to a store the other day and asked a question about the product they produce and sell.
The person helping me used AI to find the answer which they proudly told me, then after reading the docs on the bottle to make sure they gave me the right answers (pretty important for chemical use), I found they (and the AI) were completely wrong.
So the result was they were happy to give a customer the wrong information about their own product, this kind of AI use mistake is not an uncommon happening.

There are alot of very real world examples where people shouldn't be using AI because they are not capable of understanding what is right or wrong.

This is not to say that AI doesnt help at all, as there are clear cases where it is a win, but those are still mostly more process/function specific and not something the average person would do.

 

[Saylick]

But it's a net benefit for stock prices. Which is all that matters.

Sadly true. My firm is trying to adopt AI where it makes sense, but it always requires someone knowledgeable and experienced to double check the output. I fear for the future as I suspect the next generation of working professionals who inevitably will leverage LLM and AI as part of their education will fail to be as good as those who came before.

 

[igor_kavinski]

Spoiler: Sorry I know it's getting very OT BUT

Sadly true. My firm is trying to adopt AI where it makes sense, but it always requires someone knowledgeable and experienced to double check the output. I fear for the future as I suspect the next generation of working professionals who inevitably will leverage LLM and AI as part of their education will fail to be as good as those who came before.

That's been true even without AI. I read recently that Gen Z are the masters of justified effort. Previous generations used to love working hard because you got recognition for it. Now all everyone wants is to get work done ASAP so they have more time for other activities like Netflix or gaming or sports or dating or clubbing etc. and thus we are getting more and more people who want to do the most amount of work done in the least possible time, not because they want to maximize their productivity but because they want to get the minimum needed done so they can pursue their favorite activities outside of work.

The crazy amount of activities available to people now means the focus on hard work will grow weaker and weaker, leading to an overall watering down of the professional workforce's capabilities. It was always going to happen with or without LLMs. Heck, I'm from the generation where parents knew that too much computer time would affect academic performance. What do you think happens when we have a generation whose parents were brought up on computers and loved wasting time on them so they don't bat an eye when their own kids do the same?

There are alot of very real world examples where people shouldn't be using AI because they are not capable of understanding what is right or wrong.

This is not to say that AI doesnt help at all, as there are clear cases where it is a win, but those are still mostly more process/function specific and not something the average person would do.

For certain tasks such as vibe coding, LLMs are awesome. The two little tools in my sig were created with 95% help from Deepseek offline LLM. I could never have done them on my own. I just would never have gotten motivated enough to research and write them. All it took was about a day or two of messing around with the LLM and going back and forth between compiler errors and the LLM and finally I had working code. The possibilities are amazing for people who have ideas but not the will to slog through the research phase.

“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.”

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.

 

[OneEng2]

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

I spent the last day doing a little project using OpenCV and Embarcadero C++ builder. Had a hell of a time getting rid of endless linker errors with the newly built library. Tried repeatedly to use Chat and googles AI to help me figure out what was going on.

After a full day of crap, it ended up being something I tried on my own that fixed it (changed the linker I was using). After letting the AI know what happened, it was like "That makes sense because ....".

Now, I do have to give it to the AI, I learned a great deal about cmake that I never knew before. Stuff all that away with a bunch of other things I am likely never going to need again .

Another fail was a thermal problem I was working out (heat build up within a closed polycarbonate case and a PCB under different conditions). Chat got the thermal conductivity (C/W) all wrong. In fairness, Google got it right, but the only reason I switched was because Chat's answer wasn't making empirical sense with my data.

Chat did quickly give me the constants to the curve fit of F(t)=Text+P*Rth(1-e^-t/tc) which saved me a little time.

Could any of these AI's have actually DONE the job I was trying to do? (create a field analysis tool and a 2 min dynamic EOL test that can detect high resistance in an EV charger)? Not even close.

Were the AI tools useful? Sort of.

Another use was a graph component I was using that wasn't particularly well documented in its help file. I'd give AI about 30% correct answers and 70% wrong. I ended up looking up functions in the header and source files to figure out how to use the silly thing.

One of the younger coders on my team told me that if the AI is trained on libraries and projects used frequently in your organization that it is actually much better vs general usage questions.

Maybe that is true.

For now, I would say it is a useful tool in the right hands. I can easily see how it could send a junior engineer WAY off into left field. You have to know when what it is giving you is crap.