AMD: Zen 3 Announced [2020-10-08]

Zen 3 is ahead of Tiger Lake by ~6% in IPC. Rocket Lake is supposedly lower IPC than Tiger Lake so it has to be clocked significantly higher on 14nm to beat Zen3, which is unlikely. Zen 2's weakness in gaming was primarily due to the split L3 cache and that's been fixed in Zen 3. Intel isn't doing anything groundbreaking in Rocket Lake so I don't expect anywhere near 20% in gaming. I see Rocket Lake either matching Zen 3, or slightly ahead at best (at much higher power consumption).

6% ahead in what though? Zen 3 is miles ahead of Comet Lake in most workloads but still only marginally ahead in gaming. And the rumours suggest that Rocket Lake will have at least a 15% IPC uplift over Comet Lake + equal or higher clocks (5.3-5.5Ghz boost) with a specific performance focus on gaming. If true then my take on that would be that RL might trade blows with Zen 3 in the 8-core segment for non-gaming apps but take a fairly significant lead in gaming. The caveat being that it's limited to 8 cores. Would I still prefer a 12 or 16 core CPU for gaming even if in today's games it's say up to 10 or even 15% slower than Intels fastest 8 core Rocket Lake? Most likely yes.
 
I wouldn't make that assumption. AMD test where at1080p and the two really gpu light games where well ahead of skylake.

Yes but the average uplift seems to be about 6%. Granted the 2 outliers saw about 20% more performance but if Rocket Lake does come along with a 15% IPC uplift in games and a small boost in clocks (which looks to be at the low end of the rumoured estimates for gaming performance) then on average it still looks like it might be faster. But if it is, it probably won't be by much and not for another 3-5 months anyway. Zen 3 is definitely still where it's at right now and thanks to it's higher core count, likely until Alderlake or Zen 4 launch.
 
6% ahead in what though? Zen 3 is miles ahead of Comet Lake in most workloads but still only marginally ahead in gaming. And the rumours suggest that Rocket Lake will have at least a 15% IPC uplift over Comet Lake + equal or higher clocks (5.3-5.5Ghz boost) with a specific performance focus on gaming. If true then my take on that would be that RL might trade blows with Zen 3 in the 8-core segment for non-gaming apps but take a fairly significant lead in gaming. The caveat being that it's limited to 8 cores. Would I still prefer a 12 or 16 core CPU for gaming even if in today's games it's say up to 10 or even 15% slower than Intels fastest 8 core Rocket Lake? Most likely yes.

6% in Cinebench single thread. An increase in IPC does not directly translate to gaming performance. If RKL has a 15% IPC improvement, it dosen't mean games will be 15% faster. As I mentioned, one of the big reasons Zen 3's gaming performance improved was because they changed the L3 cache design vs Zen 2, which was not optimal for gaming. Intel doesn't have this design issue in Comet Lake and hence the potential for improvement in gaming isn't as much. And if you're gaming at 4K there is almost no difference in performance anyway.

Latest info directly from Intel says double digit IPC improvement, which to me points towards 10-12%. If it was more I'd imagine they'd have been more vocal about it. And only 8C with a higher PL2 of 250W compared to Comet Lake so that's not looking good.
 
I have a hard time seing where the IPC improvement for Comet Lake, being a Coffee lake derivate, should come from.

It would be a different matter if it was a highly clocked Sunny Cove derivative as it has a much wider microarchitecture.

Guess we will have to see.

Cheers
 
I have a hard time seing where the IPC improvement for Comet Lake, being a Coffee lake derivate, should come from.

It would be a different matter if it was a highly clocked Sunny Cove derivative as it has a much wider microarchitecture.

Guess we will have to see.

Cheers
I think you got the codenames mixed there, Comet Lake is Skylake-derivative (like Coffee Lake) and is out already, but the upcoming Rocket Lake is using new Cypress Cove cores aka Sunny Cove cores ported over to 14nm.
On laptops Cove offered 18% IPC over Skylake (or that's the official figure, actual benefits of course depend on workload)
 
I think you got the codenames mixed there, Comet Lake is Skylake-derivative (like Coffee Lake) and is out already, but the upcoming Rocket Lake is using new Cypress Cove cores aka Sunny Cove cores ported over to 14nm.
On laptops Cove offered 18% IPC over Skylake (or that's the official figure, actual benefits of course depend on workload)
Mea culpa.

Then there is no doubt that there will be an IPC increase. They question is if they have reduced the operating frequency penalty for this microarchitecture (at iso power).

Cheers
 
Mea culpa.

Then there is no doubt that there will be an IPC increase. They question is if they have reduced the operating frequency penalty for this microarchitecture (at iso power).

Cheers

I don't think the penalty was to do with the architecture, it was more the 10nm process. With the improved 10SF, the architecture can hit close to 5 Ghz even on mobile. So Rocket Lake desktop should be able to hit similar or higher frequencies than Comet Lake, albeit at a higher power consumption (judging from the increased 250W PL2 limit for the 8C part).
 
I'm all at sea with all the Lakes & Coves, got no idea which is which these days :confused:

On top of it the Anandtech article points out there is a contradiction between the released documentation & what they said at the presentation so its not just me :runaway:
 
6% in Cinebench single thread. An increase in IPC does not directly translate to gaming performance. If RKL has a 15% IPC improvement, it dosen't mean games will be 15% faster. As I mentioned, one of the big reasons Zen 3's gaming performance improved was because they changed the L3 cache design vs Zen 2, which was not optimal for gaming. Intel doesn't have this design issue in Comet Lake and hence the potential for improvement in gaming isn't as much. And if you're gaming at 4K there is almost no difference in performance anyway.

Latest info directly from Intel says double digit IPC improvement, which to me points towards 10-12%. If it was more I'd imagine they'd have been more vocal about it. And only 8C with a higher PL2 of 250W compared to Comet Lake so that's not looking good.

The current leaks (obviously to be taken with a big pinch of salt, although the leakers do seem reasonably reliable) suggest that the focus for Intel in their IPC uplift is on gaming. The suggestion also is that they're increasing L1 by 50% and L2 by 100%.

https://wccftech.com/intel-rocket-l...ket Lake CPU rumors,with PCIe Gen 4.0 support.
 
I don't think the penalty was to do with the architecture, it was more the 10nm process.

Wider frontend, much bigger ROB, much bigger scheduler, many more load and stores in flight at the same time. The increase in size of all the supporting structure incur a penalty on power and cycle time. At iso power, Intel will have to clock this lower, but it should still be a net win performance-wise (otherwise, why bother?)

However comparing IPC between two microarchitectures which don't have the same power characteristics at the same operating frequency is pointless/misleading. That is why Zen 3 is such a big feat, not only did AMD increase IPC substantionally, they did so while lowering power consumption at similar clockspeeds. That means the IPC increase translates into a direct performance increase.


Cheers
 
The current leaks (obviously to be taken with a big pinch of salt, although the leakers do seem reasonably reliable) suggest that the focus for Intel in their IPC uplift is on gaming. The suggestion also is that they're increasing L1 by 50% and L2 by 100%.

https://wccftech.com/intel-rocket-lake-cpu-rumors-5-ghz-all-core-oc-huge-uplift-in-gaming-performance/#:~:text=Intel's Rocket Lake CPU rumors,with PCIe Gen 4.0 support.
Dunno what they mean with references to Willow Cove since Intel said it's Sunny Cove. Increased L1 and L2 cache are known part of Sunny Cove, too.

Hey look! PCIe Gen 4.0

What happened to the rumour of Intel skipping Gen 4.0 and going to Gen 5.0?
Was that rumor ever a thing? Regardless, Tiger Lakes already have Gen 4
 
Well 5.0 is supposed to come quite soon for both AMD and Intel. And then 6.0 quite soon after that etc, they decided they don't want to wait years and years between new versions
 
Hey look! PCIe Gen 4.0

What happened to the rumour of Intel skipping Gen 4.0 and going to Gen 5.0?

That was never going to happen, not for consumer anyway. PCIE 5.0 is not going to be required for consumer for a few years at least.
Wider frontend, much bigger ROB, much bigger scheduler, many more load and stores in flight at the same time. The increase in size of all the supporting structure incur a penalty on power and cycle time. At iso power, Intel will have to clock this lower, but it should still be a net win performance-wise (otherwise, why bother?)

However comparing IPC between two microarchitectures which don't have the same power characteristics at the same operating frequency is pointless/misleading. That is why Zen 3 is such a big feat, not only did AMD increase IPC substantionally, they did so while lowering power consumption at similar clockspeeds. That means the IPC increase translates into a direct performance increase.


Cheers

Of course, not denying your points. However, 14+++ (not sure if they're using a further enhanced version for RKL) does have higher transistor performance than 10/10+ as stated by Intel themselves and this is why the clocks on 10nm are not as high (though 10nm does have lower power so more suited for mobile). It's only with 10SF and 10ESF that clocks are able to catch up, as we saw with Tiger Lake. Alder Lake should be the first desktop 10nm part with similar or higher clockspeeds as 14nm.


FPGAs are for data center where more bandwidth is always desirable, and PCIE 5.0 is being fast tracked for this segment as CXL is built upon PCIE 5.0. It isn't really needed for consumer and I don't think we're gonna see it for a few more years (unless AMDs next gen Infinity Fabric is also based on PCIE 5.0 in which case there might be a small chance).
 
Back
Top