add arrow-down arrow-left arrow-right arrow-up authorcheckmark clipboard combo comment delete discord dots drag-handle dropdown-arrow errorfacebook history inbox instagram issuelink lock markup-bbcode markup-html markup-pcpp markup-cyclingbuilder markup-plain-text markup-reddit menu pin radio-button save search settings share star-empty star-full star-half switch successtag twitch twitter user warningwattage weight youtube

Zen3 IPC reality check

yawumpus
  • 8 months ago

Could Zen 3 really improve IPC by 15%? Zen[1] to Zen+ got ~3%. Care to guess how much Skylake has improved since Sandy Bridge (10 years ago)? 30%. 30% after 10 different microarchitectures and chances for improvment. If each improved IPC by 2.5% (less than the improvement from Zen to Zen+), you would wind up with the 30% improvement, and AMD is trying to get to 15% in a single leap.

That said, this is the easiest time to do it, as they will certainly see everything that Zen left on the table and can work on the lowest lying fruit. But 15% is a lot to improve an already great microarchitecture, lets see if they can do it. I suspect that most of it will be from more efficient threading and they will claim those additional instructions as "IPC". I guess it technically counts, but really isn't why we usually want higher IPCs.

Source: https://www.realworldtech.com/forum/?threadid=185396&curpostid=185439 (I took the root sum of squares of the IPC improvement for the lot).

Comments

  • 8 months ago
  • 5 points

Care to guess how much Skylake has improved since Sandy Bridge (10 years ago)? 30%. 30% after 10 different microarchitectures and chances for improvment.

There's only been 2 architecture changes and 2 fabrication shrinks in those years. Not 10 architectures.

From this corrected perspective, Intel actually did have approximately 2 X 15% IPC leaps from one major architecture+fabrication revision to the next. They just happened to split their architecture and fabrication revisions into separate advancements.

AMD is launching both a major architecture update with a major fabrication update at the same time.

15% is almost directly inline with what Intel has given us with those same 2 major activities in previous launches (a new architecture and a new fab). For example, going from Sandy Bridge to Haswell, is the "2-step" required to get both an architecture and fab change, which bought about 15%...

I think 15% is very feasible, and expect to see even more IPC uplift in multi-core workload measurements due to the increasing width of the instruction pipeline that doesn't scale as well to a single thread. (more uplift from SMT).

From PileDriver to Ryzen was something like 40% IPC uplift single thread, ~25% multi-thread. (module vs core)

  • 8 months ago
  • 1 point

AMD claimed 52% from bulldozer to zen. I guess piledriver fixed a few more things than I thought.

And while Intel might have only changed the microarchitecture on 2 of those changes, zen to zen+ is said to net 3% IPC improvement (presumably due to caching and memory access, no microarchitectural changes) which was higher than what Intel needed to get where it was going (it needed 2.5% per tick/tock). I'd be surprised if they hit near 15% both times they did a major change.

And just because Intel hasn't had a fabrication shrink in ages, I doubt they failed to meddle in all those 14nm to 14+nm to 14++nm changes.

Skylake to Sunny Cove is said to have a higher IPC improvement, but they are sacrificing clockspeed to do it which really is a completely different subject. And probably closer to the type of changes AMD did with zen2 than the "major changes" done since Sandy Bridge (outside of AVX and related instructions).

AMD might have pulled it off. But don't expect anything like another 15% for zen3.

EDIT: Looks like Anandtech has a lot of information I was looking for. Won't be able to read it all and post back today.

  • 8 months ago
  • 1 point

From what has been released at this time regarding architecture and cache subsystem improvements, I have no reason to doubt the 15% claims.

  • 8 months ago
  • 4 points

Zen to Zen+ was just a refinement. Zen 2 is a new microarchitecture. There's a big difference.

As Allan_M_Systems pointed out already, you're drastically misrepresenting Intel's releases since Sandy Bridge as all being new microarchitectures; they weren't.

All that being said I agree with you cautioning people in a general sense. I think it is prudent to withhold excitement until we get third party benchmarks. 15% is indeed a pretty big gain.

  • 8 months ago
  • 3 points

I suspect that most of it will be from more efficient threading and they will claim those additional instructions as "IPC". I guess it technically counts, but really isn't why we usually want higher IPCs.

They reintegrated the Micro-OP cache that was faulty in Zen and disabled.

They doubled up on branch prediction and prefetch.

They doubled cache sizes.

They doubled internal bandwidth.

They doubled floating point performance.

Calling it IPC gains may be a little misleading from a certain point of view, but they do potentially have the makings of pulling off more usable clock cycles that would have gone to waste idling on Zen.

Zen was already capable of higher potential IPC then Intel's Lake architecture but between package design limitations and software not making best use of the architecture that performance hasn't been realized.

The performance increase is likely there on 3000 series looking at the above changes.

The question is going to be more if that performance is usable with how much more reliant on data fabric capabilities the design is, and if software can make use of the design changes or not.

  • 8 months ago
  • 1 point

All very interesting....and add to that the fact that Intel MUST look at ROI for each 'advancement.' That is, there must be a financial payoff given the major investment in capital for each change.

It is interesting that the figure 15% seems common...there must be a financial model where 15% is the answer from the finance department to go ahead with the development.

  • 8 months ago
  • 1 point

IPC is not exact and because of that there is no number you can assign to it to show how good it is like you can do with frequency. IPC is basically how much work a CPU core can do per clock cycle and the size of that work that fits within that IPC varies depending on workload.

As AMD stated the IPC gain is 15% from 2000 to 3000 series we don't know if that is overall across many types of workloads or under very specific workloads that may or may not be common. Well there is quite likely going to be an overall IPC increase but I would wait and see actual tests when the CPUs are released to get a better feel of the actual speed increase. We all know all companies love to shine their own products in the best possible light and AMD is no exception to the rule.

Of course IPC is not the only thing to affect CPU speed. The added cache could have an effect along with the redesign on the infinity fabric in the CPU. I do have to say I am excited by what we have heard so far about Ryzen 3rd gen and can't wait to see how good they turn out. I don't think I will bother upgrading my CPU as my 1800x is still doing what I need it to do very well but perhaps in a year or 2 when prices come down depending on the actual improvement it is an idea.

Sort

add arrow-down arrow-left arrow-right arrow-up authorcheckmark clipboard combo comment delete discord dots drag-handle dropdown-arrow errorfacebook history inbox instagram issuelink lock markup-bbcode markup-html markup-pcpp markup-cyclingbuilder markup-plain-text markup-reddit menu pin radio-button save search settings share star-empty star-full star-half switch successtag twitch twitter user warningwattage weight youtube