Discussion Zen 5 Speculation (EPYC Turin and Strix Point/Granite Ridge - Ryzen 9000)

[DisEnchantment]

Speculate at will

 

[itsmydamnation]

Seems that infos began floating once he published the relevant slide, suddenly a lot of people who were cluless became knowledgeable, there were no info prior to this slide, just wild speculations, you are rewriting history to prop up your own narrative.

Are you doing the same? @adroc_thurston had very specific contradictory performance figures at that time which he hasn't moved from to this day.

 

[adroc_thurston]

@adroc_thurston had very specific contradictory performance figures at that time which he hasn't moved from to this day

Contradictory?

 

[gdansk]

Contradictory?

To consensus?

 

[itsmydamnation]

Contradictory?

to MILD's claim

 

[Mahboi]

Are you doing the same? @adroc_thurston had very specific contradictory performance figures at that time which he hasn't moved from to this day.

I think you meant "contrary".

 

[FlameTail]

Turing Dense packs 192 Zen 5C cores into a 500W package. So, if discount efficiency losses through uncore and data movement, each Z5C core should be using around 1.5 - 2W per core.

So 1.5-2.0W for Turin Dense is per core, and discounts the uncore and everything else?

The 2W X Elite in question does include the uncore power.

 

[adroc_thurston]

To consensus?

Don't think there is one outside of weird ebeggar circles.

The 2W X Elite in question does include the uncore power.

Not realistic, was discussed already.

 

[gdansk]

Don't think there is one outside of weird ebeggar circles.

Eh, most people expect to follow previous AMD habits. Even those without influence from MLID, etc. What you (and now Kepler) have said is above that trend line.

 

[itsmydamnation]

I think you meant "contrary".

Both are acceptable//same thing in this context.

Remember I'm fully onboard Yolo 41% IPC increase ( that was upgraded from 33%)

 

[Mahboi]

Both are acceptable//same thing in this context.

contrary /kŏn′trĕr″ē/

adjective​

1. Opposed, as in character or purpose.
   "contrary opinions; acts that are contrary to our code of ethics."

contradictory /kŏn″trə-dĭk′tə-rē/

adjective​

1. Involving, of the nature of, or being a contradiction: synonym: opposite.
   "contradictory reports about the vaccine's effectiveness."

"synonym: opposite."

English is such a REDACTED unmitigated disaster.

Profanity is not allowed in the tech forums. AT Moderator Shmee

 

[adroc_thurston]

Eh, most people expect to follow previous AMD habits

Yea Zen3-class bump makes sense I guess.

What you (and now Kepler) have said is above that trend line.

Some tocks are indeed bigger than the others.
I've referred to like a bajillion cool Apple cores already.

 

[itsmydamnation]

contrary /kŏn′trĕr″ē/

adjective​

1. Opposed, as in character or purpose.
   "contrary opinions; acts that are contrary to our code of ethics."

contradictory /kŏn″trə-dĭk′tə-rē/

adjective​

1. Involving, of the nature of, or being a contradiction: synonym: opposite.
   "contradictory reports about the vaccine's effectiveness."

"synonym: opposite."

English is such a fucking unmitigated disaster.

adroc contradicted MILD's clam of IPC uplift at the time.....

you happy now

completely unrelated the the phrase happy now always make me think of

im sure MILD can relate......

 

[yuri69]

N32 was held back in the lab so they could do a very thorough postmortem and mitigate the issues as best they could.
Any respin would come out too late to be worth the effort. The good news is RDNA3.5 does not have any of these issues, so mission accomplished. Radeon is back.

Oh...

They couldn't do a postmortem using N31/N33 any of pre-prod/release samples. Right.

It's like they did not know they got a problem until like a month to the N31 release. Right.

N31 got quickly showed out of the door with utterly rushed drivers/firmware - yet there was no time for a respin.

N33 got released 6 months after N31 but still - there was no time for a respin. Right.

N32 got released like 9 months after N31 but still - there was no time for a respin. Right.

 

[Kolifloro]

My cat is VERY creative ... though it's a bit busy ...

 

[branch_suggestion]

Oh...

They couldn't do a postmortem using N31/N33 any of pre-prod/release samples. Right.

It's like they did not know they got a problem until like a month to the N31 release. Right.

N31 got quickly showed out of the door with utterly rushed drivers/firmware - yet there was no time for a respin.

N33 got released 6 months after N31 but still - there was no time for a respin. Right.

N32 got released like 9 months after N31 but still - there was no time for a respin. Right.

TTM, tapeout order and supply are all things that exist. RDNA3 works fine in headless compute so the fatal flaw in gfx power scaling was hidden a bit longer. Nobody here knows when AMD exactly found out it was a dog, but it was late enough that they didn't want to backtrack on the launch event and contradict FAD expectations.
They lied, everyone knows it, mistakes like this cannot ever happen again.
I know you don't trust RTG but guess what? RTG basically no longer exists, the new structure just works, just like the green people who are always trustworthy and can do no wrong!
Just wait for the MS AI PC maymay event where early Strix RDOA3.5+ data might be, if they stop talking about the XDNA2 Si wastage that is.
Forget about RDOA3, AMD already did.

 

[adroc_thurston]

if they stop talking about the XDNA2 Si wastage that is.

no way jose, mr. Nutella gonna haunt the entirety of AMD mgmt if they ever try not talking >AI on June 3rd.
Plus FAD'24 soon, too.

 

[CouncilorIrissa]

Remember I'm fully onboard Yolo 41% IPC increase ( that was upgraded from 33%)

the what now? when did it happen?

 

[Mahboi]

I know you don't trust RTG but guess what? RTG basically no longer exists, the new structure just works, just like the green people who are always trustworthy and can do no wrong!
Just wait for the MS AI PC maymay event where early Strix RDOA3.5+ data might be, if they stop talking about the XDNA2 Si wastage that is.
Forget about RDOA3, AMD already did.

It was taken seriously enough that they restructured all of the GPU division?

 

[yuri69]

TTM, tapeout order and supply are all things that exist. RDNA3 works fine in headless compute so the fatal flaw in gfx power scaling was hidden a bit longer. Nobody here knows when AMD exactly found out it was a dog, but it was late enough that they didn't want to backtrack on the launch event and contradict FAD expectations.
They lied, everyone knows it, mistakes like this cannot ever happen again.
I know you don't trust RTG but guess what? RTG basically no longer exists, the new structure just works, just like the green people who are always trustworthy and can do no wrong!
Just wait for the MS AI PC maymay event where early Strix RDOA3.5+ data might be, if they stop talking about the XDNA2 Si wastage that is.
Forget about RDOA3, AMD already did.

The green ppl simply deliver. They keep showing their laughably misleading slides with 2x/3x improvements but... they deliver.

RTG went through a rework after Raja got finally kicked out. During RDNA2 dev they shifted a lot of ppl from CPUs to RTG. But apparently they delivered RDNA3 after RDNA2.

So far AI PC XDNA2 reminds me AMD TrueAudio/TrueAudio Next.

 

[branch_suggestion]

It was taken seriously enough that they restructured all of the GPU division?

That whole AMD China layoff thingie was a reorg after RDNA3 went oopsie.
Lots of internal changes have been made progressively based on performance and constant reviewing.

RTG went through a rework after Raja got finally kicked out. During RDNA2 dev they shifted a lot of ppl from CPUs to RTG. But apparently they delivered RDNA3 after RDNA2.

CPU ppl were sent to make clocks go big, and they did, and then they tried way harder and nearly pulled it off, just half a gen late. Oh and recent patents point to more CPU-esque scheduling mechanisms and other funny things coming to future GPUs.

So far AI PC XDNA2 reminds me AMD TrueAudio/TrueAudio Next.

Tell M$ that, AMD is only doing what Microsoft is asking of them, like putting Pluton in Rembrandt to get... nothing in return.
MS's PC division is just completely unhinged and out of touch, but getting money from them is usually worth the hassle, and Strix and X Elite are the only premium AI SoC's for some time.

 

[uzzi38]

Tell M$ that, AMD is only doing what Microsoft is asking of them, like putting Pluton in Rembrandt to get... nothing in return.
MS's PC division is just completely unhinged and out of touch, but getting money from them is usually worth the hassle, and Strix and X Elite are the only premium AI SoC's for some time.

Also Lunar Lake but yeah.

Worse bit about the massive AIE is that in order to make room for it, AMD chopped off the original plan of slapping on a SLC. Something that would have been incredible for both CPU and iGPU performance. But yaaaaaaaaaaaay AI bubble.

I hate MS sometimes (actually almost always).

 

[branch_suggestion]

Also Lunar Lake but yeah.

Worse bit about the massive AIE is that in order to make room for it, AMD chopped off the original plan of slapping on a SLC. Something that would have been incredible for both CPU and iGPU performance. But yaaaaaaaaaaaay AI bubble.

I hate MS sometimes (actually almost always).

LNL true, though it addresses a different market so not that comparable outside of the NPU. Also 1Q later at best.
Any missed chance for MALL is very sad indeed, yay STX-Halo.

 

[Kaffeekenan]

Soooooo, according to mlid AMD is disclosing a 17% IPC uplift to partners. He says the source is an AMD employee. How does that fit the optimistic 40% stuff in this thread. @adroc_thurston Can you dismiss the 17% or what is happening. Someone has to be wrong here...and I tend to think an AMD employee should know the correct number.

 

[Tigerick]

I dislike MLID as well but I won't dismiss his leaks totally cause as for any leaks, things could change. Now we have two parties which are providing two predictions...

1. MLID +20% IPC
2. @adroc_thurston and @Kepler_L2 +40% IPC with $999 price tags

Unlike RDNA3, there is no wiggle room here, only one party will be right here...we should know more in coming months and see which one is correct. Hint from Allthewatts:

April fool. The Zen 5 IPC figures aren't too far from reality. The reality is even higher than this. Please wait till June for official figures.

 

[Kepler_L2]

Soooooo, according to mlid AMD is disclosing a 17% IPC uplift to partners. He says the source is an AMD employee. How does that fit the optimistic 40% stuff in this thread. @adroc_thurston Can you dismiss the 17% or what is happening. Someone has to be wrong here...and I tend to think an AMD employee should know the correct number.

Did you miss the part where the leak was fake and MLID fell for it?