News Early Zen 5 CPU benchmarks support AMD's IPC claims — Ryzen AI 9 365 shows 15% improvement over the previous gen

D

Deleted member 2731765

Guest
However, they’re slightly lower than the 20% higher CPU performance AMD claims for the Ryzen AI 300-series, but these tests do not include integrated GPU and NPU tests.

Those were NOT claims made by AMD. That was a prediction made by the leaker "Golden Pig Upgrade" regarding MSI's Laptops which were showcased at Computex booth, and on top of that, these were early sample chips.

He also tested the chip using single-core benchmarks, and its Zen 5 and more efficient Zen 5c cores both showed a 15% minimum instructions per cycle uplift compared to the last-generation Ryzern 7 7735U.
Geekbench 5 and 6 delivered a 15% to 17% IPC uplift over the Ryzen 7 7735U.

You sure about that ? Shouldn't that IPC uplift be for the Ryzen 7 7840U, Zen 4 chip instead ?
 
Last edited by a moderator:
D

Deleted member 2731765

Guest
7735U is previous previous generation: Rembrandt (Zen 3). If Zen 5 is only 16% better than Zen 3, it's in trouble.

I suppose that IPC uplift is against the Ryzen 7 7840U, Zen 4 chip. What do you think ?

But anyway, test system was running unofficial system firmware, and this was a very early engineering sample used. So the final silicon should perform much better than this.
 
Last edited by a moderator:
  • Like
Reactions: Roland Of Gilead
D

Deleted member 2731765

Guest
This IPC uplift in Zen 5 comes also with a lower power consumption, so together it is a substantial improvement.

Eh ? Not really. Because the Zen 3 and Zen 4 chips were running a 28W TDP, versus the 54W of the Ryzen AI 9 365 APU. So the power consumption values can't be directly correlated.

The test system was also running unofficial system firmware/software.
 
  • Like
Reactions: rtoaht
D

Deleted member 2731765

Guest
Videocardz did a much better comparison, and it is more clear as well. And they have correctly used the Ryzen 7 7840U, Zen 4 chip as a reference.

https://videocardz.com/newz/amd-ryzen-ai-9-365-zen5-apu-tested-ahead-of-launch-ipc-uplift-measured

DDen75Z.png
 

KnightShadey

Reputable
Sep 16, 2020
147
87
4,670
Was going to talk about not comparing apples to apples, but really as MM says still early firmware & software, there are some big (~30%) boosts Z4 to Z5 and some little ones, averaging tends to losing the weight vs the real world, even more-so with Spec. The notable thing is comparing Z3+ to Z5c which best illustrates where the areas where efficiency may be exploited.

Looking at the original blog ( https://blog.hjc.im/zen5-preliminary-review.html ) linked in the THG article and at the bottom of the VideoCardz artifle, some nice additional stuff including the TDP tests, intruction rate, and also inter-core latency.

The part that surprised me was the inter-core latencies seem very VERY high, even compared to designs that need to communicate between dies like 7950. You can see the very clear split between Z5 and Z5c and even within their groups it seems a bit high vs the Phoenix based 8700G.
Slower than the 2 separate low-power island efficiency cores on Meteor Lake (cores 20&21 on 155H grid), and also 6-8 of them being hamstrung vs just 2.

This is an area I don't expect changing much over time, whereas the performance numbers above may get boost when accounting for improvements in firmware & software.

First grid & second grid from David Huang's blog, following latencies from Anand.

365-latency.png


365-bandwidth.png


Compared to Phoenix desktop APU (couldn't find Ryzen mobile on Anand), Core 155H, and the BigBoy 7950X which has to communicate between dies.


AMD%20Ryzen%207%208700G%20Core%20to%20Core%20Latency%20APU.png


Intel%20Core%20Ultra%207%20155H%20Core%20to%20Core%20Latency.png


AMD%20Ryzen%209%207950X%20Core%20to%20Core%20Latency%20Final.jpg
 

rtoaht

Reputable
Jun 5, 2020
116
124
4,760
That’s pretty bad. 7735U is a a Zen 3+ chip which is pretty much a rebranded 6800U. If that’s the IPC uplift for two generations combined then AMD is in trouble.