AMD Execution Thread [2024]

[Deleted member 2197]

March 29, 2024

AMD lists mystery MI388X AI GPU in official docs, says it's barred from China sales due to US sanctions

Another mysterious product.

www.tomshardware.com

 

[del42sa]

https://wccftech.com/amd-zen-5-cpu-core-architecture-over-40-percent-faster-than-zen-4/
according Kepler_L2 info Zen 5 core is >40% faster core to core than Zen4 core in SPEC benchmark

\\ could be AVX512 or Epic core count vs freq

 

[nutball]

https://wccftech.com/amd-zen-5-cpu-core-architecture-over-40-percent-faster-than-zen-4/
according Kepler_L2 info Zen 5 core is >40% faster core to core than Zen4 core in SPEC benchmark

Sounds somewhat anomalous. I wonder which particular SPEC sub-tests are skewing those numbers.

 

[Bondrewd]

could be AVX512 or Epic core count vs freq

no it's just SIR2017, iso CC, client part.

Sounds somewhat anomalous.

oh yes. oh yeeees.

 

[Rikimaru]

Sounds somewhat anomalous. I wonder which particular SPEC sub-tests are skewing those numbers.

I've seen rumour that AVX512 is 2x faster, which could explain the difference.

 

[Frenetic Pony]

no it's just SIR2017, iso CC, client part.

oh yes. oh yeeees.

"I am shocked, shocked to find selective benchmarking just to show winnings in the computer hardware industry!"

 

[Bondrewd]

"I am shocked, shocked to find selective benchmarking just to show winnings in the computer hardware industry!"

specint isn't really selective.
FP side arguably more so.

 

[Arun]

no it's just SIR2017, iso CC, client part.

oh yes. oh yeeees.

Do you mean gaming benchmarks are much less than 40%? I could imagine they have quite different cache characteristics (see: relative V-Cache benefit) but from a “core” point of view I’d expect the kind of changes that help SIR2017 to also help most other workloads including games, so I’m curious how it can be so anomalous….

Also the rumours implied a massive peak raw FP throughout increase which tends to be an “easier” (but expensive) win on some workloads so I imagine the perf gain is pretty good there too.

If this holds up and AMD has a competitive V-Cache variant for gamers, this feels like an awful blow to Intel which would likely fail to clearly win the performance crown with Arrow Lake despite their 20A vs N4 process advantage…!

 

[Bondrewd]

Do you mean gaming benchmarks are much less than 40%?

Oh no, gaming and JITs are things Zen5 is particularly good at.

so I’m curious how it can be so anomalous….

Oh it's not.

Also the rumours implied a massive peak raw FP throughout

The FMA/FADD part is bigger yes but that's an efficiency play, bandwidth is too undersized to use it full throttle outside of MIxxxC parts.

If this holds up and AMD has a competitive V-Cache variant for gamers

V$ is a year later at a gigantic premium.
They don't need it, not until 2026 really.

Intel which would likely fail to clearly win the performance crown with Arrow Lake despite their 20A vs N4 process advantage…!

ARL issue is that the big core (LNC) is mediocre and this is me being generous.

 

[Remij]

https://twitter.com/x/status/1776085853804413035

Just a reminder to not believe everything we see/hear.

 

[Seanspeed]

A good reminder especially that these Youtube 'leak channels' are not actually vetting their sources at all, no matter how much they lie and say they are. And that's for information that they're not just straight fabricating themselves to keep up the constant videos.

 

[Bondrewd]

A good reminder especially that these Youtube 'leak channels' are not actually vetting their sources at all, no matter how much they lie and say they are.

worse, he split one e-mail into two 'sources'.
sucker/10.

 

[del42sa]

The issue that most "journalists" just blindly regurgitates super unreliable sources like MLID, without any secondary source. These days is better mostly ignore all of them, unless its either super conservative or just before product launch.

I still remember Charlie Demerjan and his forum, but he was way more cautious about his sources and info.... golden era

 

[Bondrewd]

I still remember Charlie Demerjan and his forum, but he was way more cautious about his sources and info....

Charlie is still semi(pun intended)-active tho.

The issue that most "journalists" just blindly regurgitates super unreliable sources like MLID

Clickbait outlets.
Because clicks matter.

 

[del42sa]

Charlie is still semi(pun intended)-active tho

yes, but forum is long way gone

btw, do you know anything about AM5+ platform ?

https://wccftech.com/amd-am5-plus-platform-granite-ridge-ryzen-zen-5-cpus-microcode-extraction-tool/

 

[Bondrewd]

btw, do you know anything about AM5+ platform ?

I'm a server/laptop guy so not really but afaik doesn't exist outside of like me (and some other people from my circle) memeing on AT forums.

 

[Frenetic Pony]

I could see AM5+ being a clever solution to extending AM5 into larger die SOCs. Apple can sure charge a lot for the Mac Studio, I'd imagine Microsoft and a few other vendors would be interested in that kind of profit margin.

AM5+ could be a pin compatible socket with AM5, that also allows for larger dies with more pins, maybe with more power draw. Single chiplet SOC with 16 CPU cores, 30 RDNA WGP, NPU, etc. Draws 250w? Could sell to someone anyway.

 

[Kaotik]

I could see AM5+ being a clever solution to extending AM5 into larger die SOCs. Apple can sure charge a lot for the Mac Studio, I'd imagine Microsoft and a few other vendors would be interested in that kind of profit margin.

AM5+ could be a pin compatible socket with AM5, that also allows for larger dies with more pins, maybe with more power draw. Single chiplet SOC with 16 CPU cores, 30 RDNA WGP, NPU, etc. Draws 250w? Could sell to someone anyway.

That's a lot of burned motherboards and cpus you just ordered. If the CPU and socket aren't perfect match the CPU will be installed wrong.

 

[Frenetic Pony]

That's a lot of burned motherboards and cpus you just ordered. If the CPU and socket aren't perfect match the CPU will be installed wrong.

Yeah, might not be worth pin compatibility.

But that Strix Halo w/16 CPU cores and 40 RDNA compute units doesn't sound like it will fit into 160mm that the normal AM5 socket has, so AM5+ could launch this year, and just not be compatible with normal AM5.

 

[Bondrewd]

so AM5+ could launch this year, and just not be compatible with normal AM5.

lmao no, -halo parts are never going desktop.
Chill out.

They're ~explicitly~ premium mobile wunderwaffen that AMD's been trying to pitch for a decade and no one was really buying until Apple did it.
And now it works!