Discussion David Huang's SPEC results Intel64 vs. AMD64 vs. Apple M-series

[igor_kavinski]

Hope I don't find out that there's an easier way to sort these results without cutting pasting in MS Paint!

 

[gdansk]

285K with proper memory would be a bit higher total..

It'll be interesting to see 9950X3D.

 

[DavidC1]

His results are little strange with the HX 370 too. The 5c cores are slightly slower per clock, not faster.

 

[moinmoin]

Hope I don't find out that there's an easier way to sort these results without cutting pasting in MS Paint!

I used www.table-reader.com/image-to-excel to get this file: https://www.mediafire.com/file/heyitxxq2hdmz8a/imagetoexcel.xlsx/file

 

[igor_kavinski]

285K with proper memory would be a bit higher total..

I think Huang goes for JEDEC speeds to isolate the core performance except where the RAM is soldered so don't think he's gonna be doing 285K with DDR5-9600 anytime soon

His possible reasoning may be that the CPU's cache subsystem should shoulder the burden for keeping the cores fed, instead of overclocking RAM.

 

[Meteor Late]

LPDDR is slower than non LP variants, even higher bandwidth LPDDR vs lower bandwidth DDR, close to 10%, for example 6900HS vs 7735U in laptops.
So something to keep in mind when comparing M3 variants vs x86 desktop counterparts.
On another note, Skymont looking really good here.

 

[StefanR5R]

His results are little strange with the HX 370 too. The 5c cores are slightly slower per clock, not faster.

"Per clock" results vary depending on the clock speeds at which they are obtained. "Iso clock" comparisons would paint another picture.

 

[CouncilorIrissa]

"Per clock" results vary by the clock speeds at which they are obtained. "Iso clock" comparisons would paint another picture.

yeah, the 5c cores are clocked a whole 1.8 GHz lower than the classic ones. Lowering clocks typically results in PPC increase.

 

[DavidC1]

yeah, the 5c cores are clocked a whole 1.8 GHz lower than the classic ones. Lowering clocks typically results in PPC increase.

So then explain Oryon-L performing 20% lower. SpecInt isn't affected that badly by clock scaling.

 

[adroc_thurston]

So then explain Oryon-L performing 20% lower.

You can just read his twitter. It throttles heavily in some subtests.

 

[DavidC1]

You can just read his twitter. It throttles heavily in some subtests.

Right, then it shouldn't be in the perf/GHz list.

 

[poke01]

You can just read his twitter. It throttles heavily in some subtests.

Why? Thermal issues?

 

[adroc_thurston]

Why? Thermal issues?

Well yeah you stick a 4.4G core into a phone you expect it to throttle heavily in some workloads.

 

[poke01]

Well yeah you stick a 4.4G core into a phone you expect it to throttle heavily in some workloads.

lol and Xiaomi 15 has the 2nd best cooling.
Pathetic marketing hype by Qualcomm and the audacity to compare with a laptop sku.

 

[adroc_thurston]

Pathetic marketing hype by Qualcomm and the audacity to compare with a laptop sku.

It's not the worst comparison point but it's also not indicative of steady-state phone perf either.
But yeah, aggressive boost is no longer an x86-only thing.

 

[DavidC1]

lol and Xiaomi 15 has the 2nd best cooling.
Pathetic marketing hype by Qualcomm and the audacity to compare with a laptop sku.

Also makes it look much better in power comparisons.

Gen 2 vs Gen 1 is also phone vs laptop. After what they did with original Oryon it should be taken with a heap of salt.

 

[coercitiv]

So then explain Oryon-L performing 20% lower. SpecInt isn't affected that badly by clock scaling.

Before I go read Huang's take on it, here's some other piece of information:

Our first hands-on with a real-world Snapdragon 8 Elite phone reveals a hot mess

We've got our hands on one of the first Snapdragon 8 Elite phones, and benchmarks reveal some very concerning results.

www.androidauthority.com

This behavior made me wonder about our CPU benchmarks as well. So we installed a spoofed version of Geekbench 6 to determine if Realme was indeed turning up the wick when it detects benchmark apps. The results show that single-core and multi-core CPU performance drops by a massive 63% and 47%, respectively, compared to the standard Geekbench 6 app.

Single-core scores actually drop below the Pixel 9 Pro XL, while multi-core scores are in the same territory as the Tensor chip. That’s an alarming performance downgrade for apps that aren’t on Realme’s performance optimization list.

If one phone maker has thermal issues, it may be their fault. However, if we see more of them blowing steam...

 

[igor_kavinski]

Spliced in scores of M4 and 9800X3D.

 

[igor_kavinski]

Comparison of P vs. E vs. LP E-cores (removed entries of less exciting stuff)

 

[igor_kavinski]

I love this post here:

https://www.reddit.com/r/hardware/comments/jvq3do/comment/gd0no7q

Games are great ways of getting a very broad measure of a chip's performance, because they push everything at once: local memory, shared memory, large codebases, pointer-heavy abstractions, tight vectorized loops, locks, bandwidth and latency, etc. etc. They tend to have lots of subsystems, all of which are simultaneously performance critical and budgeted for, and they give you a nice variety of performance targets.

The only problems are that 1) there's a heavy dependence on non-CPU functionality, like the GPU and storage, so they aren't nice abstract and pure CPU-only numbers, and 2) most non-game software isn't written the same way. Imagine if Slack, Word, or Photoshop or whatever was written with the same sort of attention to performance that the best games have. There's lots of general purpose software with targeted optimizations in specific areas, but very little with the holistic attitude of ‘our render budget is 16ms on this fixed piece of hardware, every frame’.

Anyone know who Veedrac is in the real world?

Also liked this:

https://twitter.com/x/status/1341660453009846272