Question ARM 2023 IP CPU/GPU news - Cortex X4/A720/A520 + Immortalis G720/Mali G720/G620

[Abwx]

What is missing on the graph above is the process, the A15 use a 5nm process, what about for the comparisons..?..

 

[Lodix]

@Abwx You can do "software simulations" of performance that pretty much represents what an actual chip would deliver. No need to fab both cores just to compare them to have an idea of their characteristics. This is how all CPU designers work.

Again ARM IP is Process agnostic. The client then chose the fab and node, cache, core counts, etc.

The first graph is just an IPC comparison at ISO frequency and configuration but 2MB L2 cache for the X4. The process is irrelevant for an IPC comparison.

The X4 will be used in 4nm SoC next year, the same as the X3 this year is on 4nm. ARM disclosed that X4 is less than a 10% increase in size (no counting cache). And if their efficiency claims are to be believed ( 40% less power for the same performance), there should be no problem if they use the same node.

 

[ikjadoon]

In Dr ian video not much difference between A55 and A510, but someone told me in Andrei Graph we should look at watts instead of joules.

These seem like more reliable numbers. The A510 here is more efficient than the A55, so it is contradicting Geekerwan's & Golden Reviewer's numbers (which, as I understand it, are platform power, not CPU power, so it seems quite variable: at low power & energy, anything on the platform could introduce big swings). I'll be honest, though: I don't know how Andrei / Ian measure CPU power & energy consumption.

But it just might as well be the node. But how come these other tests didn't catch the efficiency improvements?

8G2: A510 merged core (2021 version), 2.0 GHz, TSMC N4
D1200: A55, 2.0 GHz, TSMC N6
888: A55, 1.8 GHz, Samsung 5LPE

But, I agree with @BorisTheBlade82: Joules are usually the best here because it measures how much energy was consumed → direct relationship to battery consumption (and battery life). Power (W) can fluctuate and tests run for very different lengths of time.

1 Watt = 1 Joule / second
1 Watt-Hour = 3600 Joules

So most phones today have ~10 WHr, i.e., 36 KJ of energy.

All benchmarks take different lengths of time on different devices (faster CPUs finish the test faster), so we can't directly connect an average of 4W x 500 seconds = 2000 J.

//

However, I still like to review Watts, because Watts can share how hot a device gets (as mobile phones dissipate ~5-10W sustained).

In energy, 15W for 0.5 seconds = 1.5W for 5 seconds = 7.5 Joules consumed. But the 15W device will feel a lot warmer in the hand. That's exaggerated for example, but the same idea.

 

[ikjadoon]

The X4 will be used in 4nm SoC next year, the same as the X3 this year is on 4nm. ARM disclosed that X4 is less than a 10% increase in size (no counting cache). And if their efficiency claims are to be believed ( 40% less power for the same performance), there should be no problem if they use the same node.

Quite interestingly, Arm taped out Arm v9.2 cores also at TSMC N3E. So perhaps someone like MediaTek might release X4 on TSMC N3E (either in January-ish or a July-ish mid-cycle refresh).

Source.

“In this new generation of CPU designs, we are taking our long-standing partnership with TSMC a step further through taping out the Cortex-X4 on the TSMC N3E process – an industry first,” says Bergey, “this ensures that our ecosystem is ready to maximise the PPA benefits of our processor technologies once they are taped out

 

[soresu]

1 Watt = 1 Joule / second
1 Watt-Hour = 3600 Joules

Sooo, is a GigaWatt 1 billion Joule's / second?

It gets so confusing when people refer to femtosecond lasers running at PetaWatt power levels.

I could only assume that would be 10¹⁵ Joules / Second multiplied by 10¯¹⁵ seconds?

....which ends up being just 1 joule / femtosecond pulse again, lords give me strength 😂