Geekerwan’s Review of the Snapdragon 8 Gen 3 (Xiaomi 14).

[Jimmyjames]

Just saw this on Youtube and thought I’d post it hear if anyone wants to discuss. Have only skimmed it and will probably edit as I look deeper, but so far:

A17 Pro still leads in Spec, and single core Geekbench performance, and I think GB 6 multi core.

Cortex X4 even wider decode stage than A17 (10 vs 9). Bigger ROB than A17 (384 vs 321). I think this was already known.
Cortex X4 has a 18% Int increase vs X3 but a 39% power increase. +7% FP, +24% power increase!!

Spec tests reveal
A17 P core has a 28% Int lead over the X4 while using just 2% more power.
A17 P core has a 50% Fp lead over X4 while using 5% more power.
X4 doesn’t match A15 in either Int or Fp performance.

GPU
8 Gen 3 seems faster on GB 5 and 6 multi core performance. Gotta say I think they are underselling the A17 GB 6 multi core score, I’ve taken mine a bunch of times and it’s usually quite a bit higher than thier figure.
8 Gen 3 is more efficient per their figures. How accurate are their A17 figures? I seem to recall some doubt they are correct in their 14 watt figure for normal use.
8 Gen 3 wins big on their gpu benchmarks (Wildlife Extreme, gfxbench), according to them at lower power
8 Gen 3 wins on ray tracing. Iirc, 15-20% higher at only 75% of the power of the A17.
I had forgotten, or didn’t know the Gen 3’s gpu has 8 cores compared to the A17’s 6.

More equal in actual games. They claim Genshin is highly optimised for Apple devices.

 

[Yoused]

Cortex X4 even wider decode stage than A17 (10 vs 9). Bigger ROB than A17 (384 vs 321). I think this was already known.

What? Even Firestorm (M1 P-core) has a ROB of 630. Are they comparing X4 (a P-core) to an A17 E-core?

 

[Jimmyjames]

What? Even Firestorm (M1 P-core) has a ROB of 630. Are they comparing X4 (a P-core) to an A17 E-core?

Hmmm not sure. I assume they mean the P core. It’s mentioned at 2:57.

 

[Yoused]

old info

 

[dada_dave]

What? Even Firestorm (M1 P-core) has a ROB of 630. Are they comparing X4 (a P-core) to an A17 E-core?

Good catch, @Jimmyjames is it worth hitting up Dougall on mastodon?

Edit: here is his post on mastodon

Dougall (@dougall@mastodon.social)

A15 and A16 saw some shrinks of out-of-order structures, but A17 seems to mostly be growing these as well: * ~320 coalesced-entry ROB (vs A15: ~293, A16: ~270) * ~163 entry integer scheduler, with two 12-entry non-scheduling queues * 60 entry load/store scheduler, with 10-entry non-scheduling...

mastodon.social

Did Apple massively shrink the ROB from the A14 or is something else going on here?

CC: @Cmaier @leman

 

[leman]

Regarding ROB: there is a bit of confusion here because folks are using Dougall Johnson’s tool without paying attention to terminology. He explains that Apple uses a “coalesced ROB”, where each entry contains bundles with up to seven operations. So ROB size of 320 means up to 2'240 uops. On average, it will still be over 1000uops. This dwarfs pretty much anything on the market.

What I find interesting is that X4 is wider than A17 and yet doesn’t even reach the IPC of Firestorm…

 

[dada_dave]

Regarding ROB: there is a bit of confusion here because folks are using Dougall Johnson’s tool without paying attention to terminology. He explains that Apple uses a “coalesced ROB”, where each entry contains bundles with up to seven operations. So ROB size of 320 means up to 2'240 uops. On average, it will still be over 1000uops. This dwarfs pretty much anything on the market.

What I find interesting is that X4 is wider than A17 and yet doesn’t even reach the IPC of Firestorm…

Thanks I was wondering if it’s a terminology issue. So Anandtech using vreedak’s tool is measuring ROB size differently than Dougall’s tool? So Geekerwan may be comparing Apples to Oranges depending on how they’re measuring the X4? Or is the confusion just on our end and geekerwan is fine?

 

[leman]

Thanks I was wondering if it’s a terminology issue. So Anandtech using vreedak’s tool is measuring ROB size differently than Dougall’s tool? So Geekerwan may be comparing Apples to Oranges depending on how they’re measuring the X4? Or is the confusion just on our end and geekerwan is fine?

I think Anandtech was reporting what Dougall refers to as the "rename retire queue", which tracks register use. On Apple architectures these are two different structures. As Douglas explains, some instructions do not require a register, while some require multiple registers.

 

[Jimmyjames]

Geekbench Compute for the Adreno 750 (8 Gen 3 Gpu)

12017. Equal to the…A13. Hmmmm

Xiaomi 23116PN5BC - Geekbench

Benchmark results for a Xiaomi 23116PN5BC with an ARM ARMv8 processor.

browser.geekbench.com

Seems under Vulcan it matches an A14. Good job.

Xiaomi 23116PN5BC - Geekbench

Benchmark results for a Xiaomi 23116PN5BC with an ARM ARMv8 processor.

browser.geekbench.com

 

[Jimmyjames]

Weirdly I’m a little suspicious of Geekerwan’s Geekbench 5 multi core score. They quote 6782 but that doesn’t exist on the database. I realise scores can be withheld, but they usually don’t. Most other scores for that device are quite a bit lower. There is one at 6534 but that is also much higher than most of the others, which seem to be around the 6000 to 6200.

This is the current list. As of this post, there are 33 entries.

Geekbench Search - Geekbench

 

[dada_dave]

Weirdly I’m a little suspicious of Geekerwan’s Geekbench 5 multi core score. They quote 6782 but that doesn’t exist on the database. I realise scores can be withheld, but they usually don’t. Most other scores for that device are quite a bit lower. There is one at 6534 but that is also much higher than most of the others, which seem to be around the 6000 to 6200.

This is the current list. As of this post, there are 33 entries.

Geekbench Search - Geekbench

I can’t recall: Did they use their cooler on it? Like they did with the iPhone?

 

[Jimmyjames]

I can’t recall: Did they use their cooler on it? Like they did with the iPhone?

Not that I saw. That was my first thought as well.

 

[leman]

Weirdly I’m a little suspicious of Geekerwan’s Geekbench 5 multi core score. They quote 6782 but that doesn’t exist on the database. I realise scores can be withheld, but they usually don’t. Most other scores for that device are quite a bit lower. There is one at 6534 but that is also much higher than most of the others, which seem to be around the 6000 to 6200.

This is the current list. As of this post, there are 33 entries.

Geekbench Search - Geekbench

Yeah, I am also confused by this. Looking at publicly available GB results (either GB5 or 6), I see the Snapdragon MC at best being on par with A17, but not better.