Intel Thread Director Virtualization Patches Boost Some Workloads By ~14%

[drakonas777]

I only have data on Alder Lake, but SPECint2017 includes a gcc benchmark and 8 E-cores scored 35.52, while 8 P-cores scored 70.47 (1 thread per core) and 75.10 (2 threads per core). That suggests 16 E cores could nearly equal the P-cores' performance at just over half the area and less power consumption.

If Raptor Lake's E-cores are nearly doubling its compilation performance at only about 50% more die area, I'd say it's a win. Wouldn't you?

I must clarify my statement: in code compilation the practical benefits of E cores are questionable compared to AMD's unified approach. Obviously E cores add benefit in compilation in absolute terms and they make sense for Intel to use, but by comparing Intel and AMD top mainstream desktop SKUs we can't definitely say Intel hybrid approach is better in common compilation workloads.

[coder]

by comparing Intel and AMD top mainstream desktop SKUs we can't definitely say Intel hybrid approach is better in common compilation workloads.

I think it is, if you consider that Raptor Lake is made on a "7 nm"-class node and Zen 4 is made on a "5 nm"-class node.

What Intel did was to build larger, more complex cores with higher single-thread performance, and then offset that with E-cores. Like knights in shining armor, the P-cores' superior single-threaded performance is a win for lightly-threaded workloads, while the E-cores are like hordes of ground infantry that swarm multi-threaded tasks and overcome them by sheer numbers.

[drakonas777]

I think it is, if you consider that Raptor Lake is made on a "7 nm"-class node and Zen 4 is made on a "5 nm"-class node.

Valid point, but to be fair AMD's advantage due better lithography is mostly in the efficiency. They certainly do not have frequency advantage - quite contrary. Higher density allows them to optimize surface area, but ZEN4 has no IPC advantage (it's a little lower in fact) over Raptor Cove. Since we focus on raw compilation performance and not efficiency or semicon economics, Intel's litho disadvantage is not that impactful in this context. Though I agree it's definitely not apples vs apples comparison.

[coder]

Higher density allows them to optimize surface area, but ZEN4 has no IPC advantage (it's a little lower in fact) over Raptor Cove.

This was a conscious decision on AMD's part not to make Zen 4 a bigger, more complex core than it already is. That's what would've been needed to achieve higher IPC, but they didn't go there.

By contrast, Intel could afford to make their P-cores bigger, more complex, and less efficient, since they don't entirely rely on them for their multithreaded performance.

[drakonas777]

This was a conscious decision on AMD's part not to make Zen 4 a bigger, more complex core than it already is. That's what would've been needed to achieve higher IPC, but they didn't go there.

By contrast, Intel could afford to make their P-cores bigger, more complex, and less efficient, since they don't entirely rely on them entirely for their multithreaded performance.

Sure. And if Intel P core had, say, 30%+ IPC advantage over AMD's core I would really like Intel hybrid approach a lot more. Perhaps this will happen in the future by utilizing better lithography. However, as for today, I don't find Intel's approach appealing in the practice, even though it sounds nice on the paper. I eager to see Arrow Lake vs ZEN5 benchmarks where Intel will have a better litho and there won't be "excuses" any more.

[coder]

Sure. And if Intel P core had, say, 30%+ IPC advantage over AMD's core I would really like Intel hybrid approach a lot more.

Yes, but that's not the whole story. IPC isn't clockspeed invariant. Yet, Golden Cove has a pretty impressive ability to keep scaling performance with clock speed, whereas Zen 4 plateaus rather early. I expect a lot of that is due to Golden Cove's larger reorder buffer (512 entries), whereas Zen 4's is just 320 entries. This is also likely helping Intel in server contexts, where memory contention can be a real issue.

https://chipsandcheese.com/2022/11/05/amds-zen-4-part-1-frontend-and-execution-engine/

Perhaps this will happen in the future by utilizing better lithography. However, as for today, I don't find Intel's approach appealing in the practice, even though it sounds nice on the paper. I eager to see Arrow Lake vs ZEN5 benchmarks where Intel will have a better litho and there won't be "excuses" any more.

Oh, well... Meteor Lake has no major P-core improvements. Redwood Cove has about the same IPC as Raptor Cove. In those CPUs, only the Cresmont E-cores have IPC improvements over Gracemont.

Arrow Lake does provide P-core performance improvements, but it's not yet clear (from what I've seen) how much of that is due to clock speed vs. IPC. IPC improvement will probably be in the realm of 10% or less, which is not as good as Zen 5's is rumored to be.