Zen 5 vs Zen 4 (desktop), TSMC N4P vs N5

[Xebec]

I’ve been going through all of the recent reviews of the Zen 5 desktop chips — 6, 8, 12, and 16 core models. Zen 5 is a “clean sheet design” (Source: Ian Cuttress, others) indicating there’s going to be a lot more variation in performance characteristics than the usual generation iteration.

However, everything I’ve seen seems to indicate that the overall performance is only around 4-5% better than previous gen (released 2 years ago), and power efficiency is only a few percent (5%?) better at best, despite both being a brand new architecture and on a better node.

Digging a little deeper, the top end frequency seems to be down about 0.5% than previous gen (-25 MHz for top bins). There are also scenarios where Zen 4 efficiency is actually better than Zen 5.

I’d like to understand how much of this may be node vs. architecture. Besides Apple A15/A16, what anre some other examples of products made on both N5 and N4 that I can look at to see how their efficiency differs? (The Apple chips show a clock increase from 3.23 to 3.46 GHz for the ‘big cores’ with efficiency cores all at 2.0 GHz, but I haven’t been able to find any really good deep dives on performance/watt between the two, ).

P.S. Zen 5 (per google) is 28% denser than Zen 4, with both chips having about the same dimensional size.

..

(Some example) Sources:
7950X application performance is 96.5% of 9950X, Gaming is 96.9% in CPU constrained scenarios: https://www.techpowerup.com/review/amd-ryzen-9-9950x/27.html

Example of efficiency regression - Zen 4 and Zen 5 8-cores locked to 50W:

Hardware unboxed showing effectively zero difference in games:

 

[Jozo035]

Its bad. AMD did significant architectural changes (+40% instruction windows IIRC). They claimed ZEN 5 will achieve double digit improvements. There must be huge issue with firmware or architecture itself.

Either way, looks like Intel might catch up sooner than we thought. Maybe even this year (if they won't make mistakes).

 

[nghanayem]

However, everything I’ve seen seems to indicate that the overall performance is only around 4-5% better than previous gen (released 2 years ago), and power efficiency is only a few percent (5%?) better at best, despite both being a brand new architecture and on a better node.

N4P is not that much better than N5P, and this is the HPC versions of the node in the highest voltage regime (so the difference might even be smaller than the low power optimized version or at a lower voltage). So I suppose it shouldn't be too surprising the difference in final chips is smaller, but I don't know enough about chip design to say if it being this meager is unexplainable just from the minor node tick. Only data point that comes to mind of a similar example is intel rocket lake. Same perf in many applications and efficency as last gen with the addition to AVX-512 to client in spite of the core arch being like what 18% better. I'm with you though Xebec I would love to hear what the actual chip designers think instead of a fab engineer like me speculating.

Digging a little deeper, the top end frequency seems to be down about 0.5% than previous gen (-25 MHz for top bins). There are also scenarios where Zen 4 efficiency is actually better than Zen 5.

Makes sense Cdyn should be higher on Zen5 so if the process has a similar capability power must go up or freq must go down.

Its bad. AMD did significant architectural changes (+40% instruction windows IIRC). They claimed ZEN 5 will achieve double digit improvements. There must be huge issue with firmware or architecture itself.

Either way, looks like Intel might catch up sooner than we thought. Maybe even this year (if they won't make mistakes).

I was discussing this with a co-worker last week. We don't know exactly what this means for the server CPUs. On the on hand all that extra AVX-512 performance will likely be a boon for many workloads. On the other if at iso power it clocks lower and with barely any extra performance in many client workloads, then Turin with 33% more cores must now be operating right next to TSMC's V-min (since surely AMD wasn't that far away with Genoa from the bottom of the voltage range of N5P) and the clocks would be going down much more than with the client parts. The freq story gets even more scary for all core workloads (which I hope someone buying a 128c CPU is buying it for the all core performance ) where the regression goes from 25MHz to 690MHz. I was initially expecting Xeon 6 to get close and make up alot of lost ground, but with this information it looks like AMD's server lead might be totally gone. Of course that doesn't mean a whole lot given AMD has sales momentum from years of better server chips, but maybe this can be the start of a DCAI comeback?

 

[bearmoo]

Zen5 has 11% higher interger IPC and 24% higher floating IPC than Zen4. You cherry-picked gaming performance which is heaviely memory latency dependant as a proxy for Zen5 performance?

https://www.anandtech.com/show/21524/the-amd-ryzen-9-9950x-and-ryzen-9-9900x-review/5

 

[bearmoo]

An eight core desktop chip has a normal working power range from 65w to 100+ watt, and you picked 35 watt? What's your game?

 

[siliconbruh999]

N4P is not that much better than N5P, and this is the HPC versions of the node in the highest voltage regime (so the difference might even be smaller than the low power optimized version or at a lower voltage). So I suppose it shouldn't be too surprising the difference in final chips is smaller, but I don't know enough about chip design to say if it being this meager is unexplainable just from the minor node tick. Only data point that comes to mind of a similar example is intel rocket lake. Same perf in many applications and efficency as last gen with the addition to AVX-512 to client in spite of the core arch being like what 18% better. I'm with you though Xebec I would love to hear what the actual chip designers think instead of a fab engineer like me speculating.

Makes sense Cdyn should be higher on Zen5 so if the process has a similar capability power must go up or freq must go down.


I was discussing this with a co-worker last week. We don't know exactly what this means for the server CPUs. On the on hand all that extra AVX-512 performance will likely be a boon for many workloads. On the other if at iso power it clocks lower and with barely any extra performance in many client workloads, then Turin with 33% more cores must now be operating right next to TSMC's V-min (since surely AMD wasn't that far away with Genoa from the bottom of the voltage range of N5P) and the clocks would be going down much more than with the client parts. The freq story gets even more scary for all core workloads (which I hope someone buying a 128c CPU is buying it for the all core performance ) where the regression goes from 25MHz to 690MHz. I was initially expecting Xeon 6 to get close and make up alot of lost ground, but with this information it looks like AMD's server lead might be totally gone. Of course that doesn't mean a whole lot given AMD has sales momentum from years of better server chips, but maybe this can be the start of a DCAI comeback?

Xeon 6 have AMX and bunch of Accelerators no one talks about like QAT/IAA AMX for AI™ it is going to be very close 128C/256T Xeon 6 vs 128C/128T zen 5 also Intel really has a very good software Support if you compare to AMD it's what AMDs Achilles heel has been and it's Rivals Excels at i am personally hyped about Clearwater forest on 18A

 

[lefty]

I think the reason why they couldn't clock Zen 5 higher than Zen 4 (despite being on a better node) is because Zen 5 has doubled the AVX-512 data paths and execution units, that means quite a lot of power output concentrated in a very small area. It's not because of any defect in the N4P process.

 

[siliconbruh999]

I think the reason why they couldn't clock Zen 5 higher than Zen 4 (despite being on a better node) is because Zen 5 has doubled the AVX-512 data paths and execution units, that means quite a lot of power output concentrated in a very small area. It's not because of any defect in the N4P process.

Yeahyou can see here the current limit is hit

 

[Xebec]

Zen5 has 11% higher interger IPC and 24% higher floating IPC than Zen4. You cherry-picked gaming performance which is heaviely memory latency dependant as a proxy for Zen5 performance?
View attachment 2178
https://www.anandtech.com/show/21524/the-amd-ryzen-9-9950x-and-ryzen-9-9900x-review/5

Oddly spec is an outlier (and while I know it’s the traditional app for comparison architectures across time, it’s also very easy to compile certain benchmarks to make CPUs look a lot faster than they are in real world applications). I avoided adding a ton of links to my first post because I didn’t want to create a wall of text, knowing that someone would probably indicate “cherry picking”.

Hardware Unboxed, Gamers Nexus, Level1techs, Eurogamer, Techpowerup have all tested applications en masse and 9700X is at best 5% overall better than 7700X, excluding AVX512. For games, there’s maybe 2 games with >10% gains, most average 3-4% (across a wide spread of CPU bound scenarios).

Example for >20 apps: https://www.techpowerup.com/review/amd-ryzen-7-9700x/27.html
Example for >10 games: https://www.techpowerup.com/review/amd-ryzen-9-9950x/17.html. (also includes 9700X, 7700X, Hardware Unboxed which tests 13 games shows the same thing)

For apps, 7700X has 96.4% of the performance of 9700X.

FWIW, For the 12 and 16 core variants, the tech community has observed that the inter-core latency is a lot higher across CCDs in Zen 5 than Zen 4.

I can share more links upon request.

 

[Jozo035]

AMD is fixing things. Kind of.

Ryzen 9000 Series Community Update: Gaming Performance

The new “Zen 5” architecture first introduced in high end laptops earlier this summer is now available in a full stack of desktop processors with the AMD Ryzen 9000 Series. And while “Zen 5” is delivering leadership performance across content creation, productivity and AI applications, we’ve...

community.amd.com

+6% in "Procyon Office productivity benchmark". Which i never heard about but probably it's good. And They suggest participating in Windows insider program and using preview builds... Which is probably not ideal.

 

[samwilde]

I bought AMD Zen 4 but I am sitting this one out. I will probably only get another computer when GAA transistor CPUs come out.

 

[vslee]

when GAA transistor CPUs come out.

Luckily for you, that will be in Oct, assuming Arrow Lake uses 20A