He is testing a snapdragon 888 here. But at the time stamp of about 8:30 there is a freq vs power graph. Hope it helps with regards to the a78 performance. There is a perf vs freq graph at 5:50 as well if you want to skip to the relevant parts!
Thank you for sharing. The 6:57 graph shows the less than 1:1 relationship between perf & frequency (on all the cores, actually: A55, A78, and X1), so that may be it, but it's sharper than I realized. One issue is that this is a bit more of a scratch benchmark (computing SHA256 hashes) so unsure we can apply that widely to Arm's GB claims.
Umm. So TLBC (Too Long, Bit Confusing)- what's the IPC gain we can expect from Cortex X5?
Looking at
Arm's own claims (these aren't comparable, FWIW, as each year's IPC calc has different benchmarks),
A76 was
+28% IPC over the A75 (+35% claimed with a +7% clocks, so)
A77 was
+29% IPC over the A76 (estimate from a bar chart with no numbers; geomean of SPEC2006 int & fp, no weighting)
X1 was
+30% IPC over the A77
X2 was
+16% IPC over the X1
X3 was
+11% IPC over the X2
X4 was
+13% IPC over the X3
Of course, as AnandTech showed, Arm's own IPC projections aren't known to be precise, even as AnandTech used the same benchmark. We go into the age-old debate of cache gimping, bad tweaks, etc. And Arm is likely "adjusting" their IPC calc's collection of tests so the end result is at least a double-digit IPC gain YoY (
as they showed in 2023). But luckily, Arm doesn't go wild: it's usually either some SPEC (int or fp; 2006 or 2017) or some version of Geekbench.
To "close the gap"—as the note claims—with Oryon & M3's IPC in GB6.2 1T, the Cortex-X5 needs around
+6-8% IPC in GB6.2 1T, which is not much. Arm's
minimum has been
+11%.
So if it's the largest IPC uplift YoY in five years, that'd imply Arm's Cortex-X5 could have a higher pts / GHz ("IPC") in GB6.2 1T than Oryon and M3. Of course, the X5 will release six to 12 months later (Jan 2025 ish).
- Apple M3 - 764 pts / GHz (4.056 GHz / 3,099 pts) - 119.6%
- Qualcomm SDXE / NUVIA Oryon - 753 pts / GHz (4.30 GHz / 3,236 pts) - 117.8%
- Qualcomm SD8G3 / Arm Cortex-X4 - 706 pts / GHz (3.30 GHz / 2,329 pts) - 110.5%
- Qualcomm SD8G2 / Arm Cortex-X3 - 639 pts / GHz (3.36 GHz / 2,146 pts) - 100.0%
- Intel i9-14900K / Raptor Lake - 522 pts / GHz (6.00 GHz / 3,134 pts) - 81.7%
- AMD 7950X / Zen4 - 512 pts / GHz (5.70 GHz / 2,916 pts) - 80.1
For reference, Alder Lake was
+17% vs Rocket Lake and +29% vs Sky/Coffee/Comet Lake in IPC (of course, ADL launched six years later than Skylake, so 29% isn't wild).
Arm appears to be more accurate with its IPC claims over the years (with the X4 even sandbagging, though likely due to GB6.1 vs 6.0 changes), but +30% still sounds hard to believe, especially a year out with no public simulations. It may be +30% in a specific area? As the A77 launch showed, IPC is quite test-dependent.
So is the A77 ~15% or +35% IPC gain over the A76? Both, right? I mentioned in the Zen thread that quoting IPC without a test is like quoting frames per second without a game, and it'll apply to Arm, too.
But, if we believe Arm to the letter, there
should be some tests where X5 is +30% IPC YoY over the X4. Which is wild.
Interesting proportional difference of Int and FP increases between the 2 cores.
I always had the impression that X1 was dramatically better than A78 in every performance aspect from Int to FP to SIMD.
I wonder if later X series addressed this or it was effectively a new trend in ARM core design.
Yeah, that's what Arm's new X1 branding would have suggested, right? AnandTech did share the X1 was an
"ultra-charged A78" with fewer changes in execution and more in the front-end,
Overall, what’s clear here about the Cortex-X1 microarchitecture is that it’s largely consisting of the same fundamental building blocks as that of the Cortex-A78, but only having bigger and more of the structures. It’s particularly with the front-end and the mid-core where the X1 really supersizes things compared to the A78, being a much wider microarchitecture at heart. The arguments about the low return on investment on some structures here just don’t apply on the X1, and Arm went for the biggest configurations that were feasible and reasonable, even if that grows the size of the core and increases power consumption.
Namu Wiki has
some more background, though unsourced:
The CXC (Cortex-X Custom) project started with the Advanced Development Program (ADP), and has reached the present through the ELP (Enhanced Lead Partner) program in the middle. ADP was a semi-custom method that modified only a very limited part of the underlying core or interconnect. As the ELP program accommodates more requests, extensive changes are reflected, and accordingly, the product name, code name, and CPU ID are treated as separate processors.
According to these program changes, Hercules [A78] was customized through ADP and became Hera [X1], and then rebranded as Cortex-X1 by ELP program.
//
The question is, when ARM says 5 years, do they mean since A77, or do they mean A76?
Because in January 2019 the A77 wasn't to be announced for 5 months yet, so they could be talking about A76.
Do we have any comparative Int/FP numbers on A75 -> A76, and A76 -> A77?
A good Q. I ended up adding the A76 launch to the top list as Arm may well be counting like that, too
--
TL;DR: Cortex-X5
appears to be a strong launch for Arm. Next year should be a hell of a year for CPUs:
Apple M4 vs Arm Cortex X5 vs Qualcomm Oryon vs Intel Lunar Lake vs AMD Zen5? Let's freaking go. Can't remember the last time we genuinely had
five uArch designs competing in PCs.
EDIT: Confused the X3 with the X4 in the IPC. The Cortex-X5 needs hardly 6-8% IPC uplift to match the M3 & Oryon IPC in GB6.2 1T. Fixed now. By that, the Cortex-X5 may hold be
the highest perf / GHz in GB6.2 1T ever of any consumer CPU, depending on the Apple A18 Pro's results late 2024.
Facebook
Twitter