AMD "Zen 3" Microarchitecture Could Post Significant Performance Gains

[btarunr]

At its recent SC19 talk, AMD touched upon its upcoming "Zen 3" CPU microarchitecture. Designed for the 7 nm EUV silicon fabrication process that significantly increases transistor densities, "Zen 3" could post performance gains "right in line with what you would expect from an entirely new architecture," states AMD, referring to the roughly 15 percent IPC gains that were expected of "Zen 2" prior to its launch. "Zen 2" IPC ended up slightly over 15 percent higher than that of the original "Zen" microarchitecture. AMD's SC19 comments need not be a guidance on the IPC itself, but rather performance gains of end-products versus their predecessors.

The 7 nm EUV process, with its 20 percent transistor-density increase could give AMD designers significant headroom to increase clock speeds to meet the company's generational performance improvement targets. Another direction in which "Zen 3" could go is utilizing the additional transistor density to bolster its core components to support demanding instruction-sets such as AVX-512. The company's microarchitecture is also missing something analogous to Intel's DLBoost, an instruction-set that leverages fixed-function hardware to accelerate AI-DNN building and training. Even VIA announced an x86 microarchitecture with AI hardware and AVX-512 support. In either case, the design of "Zen 3" is complete. We'll have to wait until 2020 to find out how fast "Zen 3" is, and the route taken to get there.



View at TechPowerUp Main Site

 

[Prima.Vera]

I hope they will trash Intel and their 14+++++++++ nm CPUs into oblivion.

 

[R-T-B]

I hope they will trash Intel and their 14+++++++++ nm CPUs into oblivion.

I hope we remain in healthy competition, and no one "trashes" anyone.

It breeds better consumer pricing, for starters.

No denying AMD is very well poised on the CPU front now though...

 

[Calmmo]

There were some direct quotes a while ago, something about 4-7%, about half of that being the fab change and the other ~half design improvements.
I'd be surprised if it's anything more than that. Maybe a small 100ish clock boost on top of that 1:1 4%-7%

 

[biffzinker]

If you want to know where Intel is headed just have a look see at mobile Ice Lake's performance. The Sunny Cove Arc definitely pushes the IPC up even with the clock speed deficiency.

Intel Core i7-1065G7 Benchmarked: Ice Lake with Iris Plus Graphics

This has been a long time coming, as we're finally getting around to testing Intel's Ice Lake architecture. This is Intel's first real attempt at a 10nm...

www.techspot.com

 

[ZoneDymo]

I hope we remain in healthy competition, and no one "trashes" anyone.

It breeds better consumer pricing, for starters.

No denying AMD is very well poised on the CPU front now though...

well yeah ultimately we all want that, but Intel is a bit too big for its bridges, a good kick in the nutz is what they need.

 

[R-T-B]

well yeah ultimately we all want that, but Intel is a bit too big for its bridges, a good kick in the nutz is what they need.

They been having that for the past bit, honestly.

 

[Noztra]

If you want to know where Intel is headed just have a look see at mobile Ice Lake's performance. The Sunny Cove Arc definitely pushes the IPC up even with the clock speed deficiency.

Intel Core i7-1065G7 Benchmarked: Ice Lake with Iris Plus Graphics

This has been a long time coming, as we're finally getting around to testing Intel's Ice Lake architecture. This is Intel's first real attempt at a 10nm...

www.techspot.com

Did you even read the conclusion on your own link?

 

[biffzinker]

Did you even read the conclusion on your own link?

Yes, the Razer Blade Stealth doesn't allow for the best showing of the i7-1065G7 as someone in the comments pointed out.

Razer Blade is not the best Ice Lake implementation available.

 

[notb]

Yes, the Razer Blade Stealth doesn't allow for the best showing of the i7-1065G7 as someone in the comments pointed out.

Extrapolating mobile SoC's results on the whole architecture doesn't make sense in general.
Chips in laptops are always limited. There's a huge performance variance between different laptops.
You can find the best one, but you have no guarantee that's all the SoC has to give.

One would have to remove all limits from OEM configuration, disassemble the laptop and run it with a desktop cooler.
I'm not even sure if this is possible.

 

[fancucker]

Here we go with the relentless hype machine. Process benefits and uarch I'd be surprised if it cracks 10% over Zen 2. Maybe AMD can finally equal the ST performance of Coffee-Lake.

 

[notb]

Here we go with the relentless hype machine. Process benefits and uarch I'd be surprised if it cracks 10% over Zen 2. Maybe AMD can finally equal the ST performance of Coffee-Lake.

And more importantly... what's the point?

OK, maybe they've found a better idea for the architecture and it'll give them extra 5% in benchmarks compared to just optimizing Zen2.
And what about real life? Because after 2.5 years software finally starts to fly on Zen.
Are we looking at another shitstorm of "no one wants to optimize for AMD"?

In other words: I don't believe in the new arch. Maybe it'll be tweaked. Maybe they'll make the I/O die on older 7nm (as Zen dies move to EUV).
A lot of things can be improved in Zen2.

And instead of launching a new architecture every 3 years, they could focus a bit on the software. Clearly, they aren't as poor as some people here suggest.

 

[efikkan]

I do hope AMD will be stepping on the gas rather than slowing down. We need a continued push for higher single core performance (both IPC and SIMD performance).
There is still no word on AVX-512 from AMD, unfortunately.

There were some direct quotes a while ago, something about 4-7%, about half of that being the fab change and the other ~half design improvements.
I'd be surprised if it's anything more than that. Maybe a small 100ish clock boost on top of that 1:1 4%-7%

It's always hard to know what people mean by such figures. Node improvements doesn't yield IPC improvements.

 

[R0H1T]

I hope we remain in healthy competition, and no one "trashes" anyone.

It breeds better consumer pricing, for starters.

No denying AMD is very well poised on the CPU front now though...

I hope we get a return of the "core" wars & see each core retailing for as little as $10

 

[biffzinker]

Maybe AMD can finally equal the ST performance of Coffee-Lake.

AMD has equal or better than Coffeelake IPC.

Intel Core i9 9900KS processor review

Intel has released its flagship desktop processor in the non-HEDT segment. The 9900KS is their most premium Coffee Lake-S eight-core processor that has been discussed so abundantly. This little beast ... Performance - CineBench 15 (and IPC)

www.guru3d.com

 

[oxrufiioxo]

AMD has equal or better than Coffeelake IPC.

View attachment 137573

Intel Core i9 9900KS processor review

Intel has released its flagship desktop processor in the non-HEDT segment. The 9900KS is their most premium Coffee Lake-S eight-core processor that has been discussed so abundantly. This little beast ... Performance - CineBench 15 (and IPC)

www.guru3d.com

Not 100% sure why but people get confused with IPC and clock speed advantage not knowing for some reason that clocks need to be matched to measure IPC from 2 different architectures.

 

[efikkan]

AMD has equal or better than Coffeelake IPC.

Intel Core i9 9900KS processor review

Intel has released its flagship desktop processor in the non-HEDT segment. The 9900KS is their most premium Coffee Lake-S eight-core processor that has been discussed so abundantly. This little beast ... Performance - CineBench 15 (and IPC)

www.guru3d.com

Cinebench is a benchmark of the rendering engine of Cinema 4D, not a measure of IPC. Cinebench is not a representative workload, and should never be extrapolated into "general performance".

 

[ZoneDymo]

Not 100% sure why but people get confused with IPC and clock speed advantage not knowing for some reason that clocks need to be matched to measure IPC from 2 different architectures.

I do, people are stupid.

Cinebench is a benchmark of the rendering engine of Cinema 4D, not a measure of IPC. Cinebench is not a representative workload, and should never be extrapolated to "general performance".

Not saying you are wrong, but could you explain?
They simply test 1 core of each processor, running at the same speed, and see how quickly it gets done.
How does that not indicate IPC ermm levels? vs eachother?

 

[efikkan]

Not saying you are wrong, but could you explain?
They simply test 1 core of each processor, running at the same speed, and see how quickly it gets done.
How does that not indicate IPC ermm levels? vs eachother?

"IPC" is a measure of "workload agnostic" performance per thread per clock. In its literal form it means instructions per clock, but given the code is the same, this should translate directly to performance per clock*.

"IPC" is an approximation of the processor's "general performance", which is why you can't just cherry-pick one benchmark and say it demonstrates IPC. You either need a wide selection of benchmarks, or a benchmark which incorporates this into a single one (hopefully without becoming too synthetic).

*) And if we're pedantic here, IPC isn't actually linear either, due to memory latency being a constant. But it's not a problem as long as you compare all CPUs at the same clock.

 

[notb]

AMD has equal or better than Coffeelake IPC.

He said "ST performance", not IPC.

I do hope AMD will be stepping on the gas rather than slowing down. We need a continued push for higher single core performance (both IPC and SIMD performance).
There is still no word on AVX-512 from AMD, unfortunately.

But what's the point if it's not utilized?
Of course AMD could make an even more extreme design that no one knows how to use. What's the point?

Intel's products are way more stable, so - even putting aside market share - there's way more value in optimizing software for their APIs.
Keep that in mind next time you'll see a topic about unfair Matlab or something like that.

They simply test 1 core of each processor, running at the same speed, and see how quickly it gets done.
How does that not indicate IPC ermm levels? vs eachother?

Cinebench is not a general testing suite (like e.g. SPEC CPU2006).
It's a benchmark for a particular software: Cinema4D rendering engine.

If you run a similar test in different software, you'd get different results.

 

[phill]

Just waiting for the reviews

 

[john_]

AMD's SC19 comments need not be a guidance on the IPC itself, but rather performance gains of end-products versus their predecessors.

Why is that? I think it is specifically said that we are talking about IPC, not for example IPC+frequency+faster memory+whatever else=+15%.

 

[efikkan]

But what's the point if it's not utilized?
Of course AMD could make an even more extreme design that no one knows how to use. What's the point?

Do you mean what's the point of AVX-512?
It has obviously twice the vector width of AVX2, but have also more types of operations and is more flexible.
But new instructions have always the problem of adoption. If you don't ship hardware support, software is not coming along. Right now, support in the client space is pretty much non-existent, but custom softare uses it of course. But once AVX-512 is widespread enough, we'll start to see more applications utilize it, and at that point hardware lacking support will have a major disadvantage. But still, hardware have to come first to lead the way.

Intel's products are way more stable, so - even putting aside market share - there's way more value in optimizing software for their APIs.

What "APIs"?
Desktop software is written and compiled towards the x86 ISA, which is microarchitecture agnostic. The only level of "optimization" in that regard is which optional extensions you choose to use, most of which are supported by both. There is no way to "optimize for Intel" etc., not the way most people think it is.

 

[medi01]

Hyping AMD stock someone is.

 

[Chrispy_]

And more importantly... what's the point?

OK, maybe they've found a better idea for the architecture and it'll give them extra 5% in benchmarks compared to just optimizing Zen2.
And what about real life? Because after 2.5 years software finally starts to fly on Zen.
Are we looking at another shitstorm of "no one wants to optimize for AMD"?

In other words: I don't believe in the new arch. Maybe it'll be tweaked. Maybe they'll make the I/O die on older 7nm (as Zen dies move to EUV).
A lot of things can be improved in Zen2.

And instead of launching a new architecture every 3 years, they could focus a bit on the software. Clearly, they aren't as poor as some people here suggest.

I kinda agree with you there. The 1.0.0.3 AAB AGESA updates were a fix to a non-issue as far as I'm concerned because they addressed a 50MHz defecit in single-threaded performance at the cost of voltage bumps. It mattered only to people who were synthetic benchmarking because in the real world you never have a multi-core CPU using just one core. The OS has thousands of threads running at all times and the only way to actually achieve "single-threaded performance peaks" is by using synthetic benchmarking software to hog all the cores for itself and then run nothing on the idle ones for the synthetic single-threaded test.

Ryzen is now the best processor because in the real world you get more threads and more cache than Intel, which means that outside of niche use-cases like 720p240Hz gaming and synthetic single-threaded benchmarking, stuff just gets done faster on Ryzen. If Intel had more of an advantage than they do, it wouldn't be so clear cut, but Zen2's IPC and clocks mean that in many cases the fact that each core is still slightly weaker doesn't matter, because AMD will give you TWICE as many threads at Intel at many of the popular price points. That's a 100% absolute advantage when Intel's IPC*Clockspeed advantages are 10% at best, even ignoring the massive power draw issues Intel 14nm++++++++++ has.