AMD Intends to Deliver Higher-than-Average IPC Gains with Each New Generation of Zen

Tsing

The FPS Review
Staff member
Joined
May 6, 2019
Messages
11,204
Points
83
The industry's average annual growth rate for IPC stands at a slow 7%. AMD wants to break that trend.

In an interview with AnandTech, CTO Mark Papermaster noted that one of the primary goals with Zen is to bring higher and higher single threaded performance with each subsequent generation.

That would imply a substantial gain in IPC every 12 to 18 months. (According to Anandtech's editors, that works out to a +10.7% improvement.)

The assertion seems plausible enough, being that Zen 2 managed a 15% IPC increase over the previous generation.

Papermaster hinted that Zen 3's uplift would be derived potentially in large part by the improved cache design.

IC: Everyone’s favorite topic when discussing new hardware is about increases in IPC. For Zen 2, we saw a very good 15% increase in IPC over the previous generation. Recently we heard about Forrest Norrod (SVP of Datacenter) talk about Zen 3, in that we should expect to see a very typical increase in IPC again, making another step-function improvement over Zen 2. Do you expect AMD to continue with these sorts of improvements?

MP: So with our roadmap, and as you know I won’t go into specifics about IPC gains for future generations, but we are driving a model of high-performance compute, and there will be performance gains with each new generation. It’s all about picking the right set of IPC improvements, and the right process nodes and design rules for our core designs. There’s always a balance when we design our cores, so match power with efficiency, and at AMD we still expect to exceed the industry standard. We’ve stated before that the industry has been on a slow 7% annual growth rate in single threaded performance, and our goal is to beat that with every generation of our products. We’ve executed better than the industry with our recent products and we exceeded industry expectations.
 
I sure hope Zen 3 comes out of the gate with no issues, like the bios stuff for Zen 2. That caused me to wait and see how things settled and now I am waiting for Zen 3, because it is only a few months away. And I can wait, especially for another 15% performance.
 
The BIOS issues also occurred with the original Zen launch. So far, each new Zen generation has had BIOS issues at launch that have taken some time to get just right. Heck, even Zen+ had its issues and that was a real small step.

But here's hoping for a smoother launch.

Overall I'm not too concerned about it though, AMD gets it right eventually and fixed the issues, like anything, early adopters seem to be beta tests, but wait a few months, and you might have a better experience.
 
Brent_Justice said:
The BIOS issues also occurred with the original Zen launch. So far, each new Zen generation has had BIOS issues at launch that have taken some time to get just right. Heck, even Zen+ had its issues and that was a real small step.

But here's hoping for a smoother launch.

Overall I'm not too concerned about it though, AMD gets it right eventually and fixed the issues, like anything, early adopters seem to be beta tests, but wait a few months, and you might have a better experience.
Click to expand...

I will be waiting at least a few weeks before buying a Ryzen 4xxx, mainly because I don't have a ton of faith in Asus with their BIOS updates for older motherboards (X470 in my case). Worst case I'll get a GPU upgrade first.
 
Hopefully the new architecture changes will help with memory latency.

Getting memory latency <66ms on Zen 2 makes a massive performance difference in games.
 
Wait, zen 3 will still be the old socket?
 
Tsing said:
The assertion seems plausible enough, being that Zen 2 managed a 15% IPC increase over the previous generation.
Click to expand...

I don't think this is 'plausible', outright, because Zen 2 was mostly the culmination of AMDs efforts to get the Zen cores actually fed.

I take this as AMD claiming that they're actually pulling more IPC out of the Zen cores themselves, not just increasing the various on-package cache amounts and bus speeds.

Which if they can do that, and increase all-core load clockspeeds, then cool!

It's just not a 'given' ;)
 
We'll see how long this plan lasts. Continuing to improve IPC while not losing clock speeds will be problematic over the long haul.
 
Dan_D said:
We'll see how long this plan lasts. Continuing to improve IPC while not losing clock speeds will be problematic over the long haul.
Click to expand...

I think Intel's contingency to backport Skylake's successor architecture to 14nm might shed some light. If they're able to keep the +10% to +15% IPC they've shown with Ice Lake while also keeping the ~5GHz clockspeeds, they might have something -- you know, at least for less multithreaded workloads cause it's not like they're going to be able to put as many cores in each socket as they are with Skylake now ;).
 
SilentGhost said:
I think Intel's contingency to backport Skylake's successor architecture to 14nm might shed some light. If they're able to keep the +10% to +15% IPC they've shown with Ice Lake while also keeping the ~5GHz clockspeeds, they might have something -- you know, at least for less multithreaded workloads cause it's not like they're going to be able to put as many cores in each socket as they are with Skylake now ;).
Click to expand...

As we've seen with Cascade Lake-X, its problematic to put lots of cores in these packages. The Core i9 10980XE is a good example of this. Sure, you can clock those 18c/36 thread pretty well through overclocking, but you need serious cooling to do it and your power consumption is utterly insane.
 
Dan_D said:
As we've seen with Cascade Lake-X, its problematic to put lots of cores in these packages. The Core i9 10980XE is a good example of this. Sure, you can clock those 18c/36 thread pretty well through overclocking, but you need serious cooling to do it and your power consumption is utterly insane.
Click to expand...

Yup, that's what I'm getting at: Intel may be able to get eight or ten cores into a socket that are significantly faster than Skylake and Zen 2, but using more than around ten cores is going to be a problem.
 
You must log in or register to reply here.
Become a Patron!
Back
Top