Inside your next Apple Watch?I would like to have a Zen 5 Ryzen 9 please, with 2x 6 wide decoders, 8 ALUs, 8 FP pipes, 6 Giggle Hertz and please make it 25 watts.
And Glass.Inside your next Apple Watch?
Heard it here first folks. Zen 5 is an overclocking beast!!So it overclock well, hmm
So far it seems it shouldn't be called Zen anymore - it's not a single balanced product anymore. It's heavily leaning towards SIMD and SMT(?). The core even requires two separate "lighter" versions.So what's the consensus here, is zen 5 good yet?
Seriously now, IMO AMD made the choice of Keeping the TDP low so it is probable that those can pushed (more than Zen4).Heard it here first folks. Zen 5 is an overclocking beast!!
Apple M4 scores nearly 3400 without SME. Here's the result for exactly the same device using Geekbench 6.2 without SME and Geekbench 6.3 with SME:It’s scores ~3700 as well in GB6.2 which doesn’t include SME.
Hmm, it seems then notebookcheck was wrong.Apple M4 scores nearly 3400 without SME. Here's the result for exactly the same device using Geekbench 6.2 without SME and Geekbench 6.3 with SME:
View attachment 101847
A comparison is available here.
It should be higher than 3400 with proper cooling, considering that the Ryzen result also has proper cooling it would be required for a fair comparison.Hmm, it seems then notebookcheck was wrong.
Honestly, Geekbench needs filters. Till then Geekbench 5 is one to use.
To be honest don't get why many peeps are so focused on GB which was always an instrument for Apple to make their products look good.
In this particular case it's not about 1T/nT.Sorry, folks. I created the table for the ES 9950X DDR5-8000 configuration in LibreCalc and I guess I don't know how to make formulas work flawlessly in that AND I was in a hurry to post.
Here's the updated percentage improvement table and I hope I got it right this time:
View attachment 101856
So looking at that, maybe the ST perf isn't that hot but in MT, it's a beast?
If you are perplexed by these numbers, you are not alone
Any mistakes are all mine!
It's on an air cooler.Are you running air cooler or just AIO? I was under impression that properly clocked 7950x can do about 1.4 gflops if thermals can be sustained (one of my friends tested 7900x with custom water loop at release time and he managed to score 1.1 gflops or about that)
Sorry, folks. I created the table for the ES 9950X DDR5-8000 configuration in LibreCalc and I guess I don't know how to make formulas work flawlessly in that AND I was in a hurry to post.
Here's the updated percentage improvement table and I hope I got it right this time:
View attachment 101856
So looking at that, maybe the ST perf isn't that hot but in MT, it's a beast?
If you are perplexed by these numbers, you are not alone
Any mistakes are all mine!
In this particular case it's not about 1T/nT.
It's all AVX-512. "The most overkill AVX-512 implementation ever" or something along these lines according to @adroc_thurston.
AMD specifically mentioned doubled L1 -> FP PRF throughput in their slides even.
I have a different table for DDR5-8000. It doesn't matter. The core has potential. Whether everything can work out in the end is the big question.How did you calculate it?
I have a different table for DDR5-8000. It doesn't matter. The core has potential. Whether everything can work out in the end is the big question.
"It's all AVX-512" - No. This is the fault of all the changes listed below. The ZEN5 architecture is the most revolutionary since ZEN1 and has the most profound changes and is not just an extension of previous architectures. .
View attachment 101868
and this
View attachment 101867
The simulations shown are basically FP throughput tests. They will not translate at all to generalized workloads.
I can’t begin to describe or even understand myself the amusement I will have if Zen 5 is a working Bulldozer design.So far it seems it shouldn't be called Zen anymore - it's not a single balanced product anymore. It's heavily leaning towards SIMD and SMT(?). The core even requires two separate "lighter" versions.
It feels like made for HPC first, not for the general purpose compute as Zen 1-4 were. I'm just afraid of necessity of recompiled binaries, odd performance drops, etc.
I hope the final product won't be that bent towards HPC but...
I can’t begin to describe or even understand myself the amusement I will have if Zen 5 is a working Bulldozer design.
This is definitely not the case. HPC = scalability, which has already been the core of the Zen uarch. Zen 5 is about single thread performance and future upgrade path.It feels like made for HPC first
Does it hinder HPC from achieving higher scalability with SMT?This is definitely not the case. HPC = scalability, which has already been the core of the Zen uarch. Zen 5 is about single thread performance and future upgrade path.