Discussion Qualcomm Snapdragon Thread

[Nothingness]

Why L0?

Reduced latency.

 

[rgba]

Spoiler: Cache config from CPU-Z

 

[FlameTail]

Spoiler: Cache config from CPU-Z

View attachment 99669

WOOOW

So the L1d difference vs M3-P is that Oryon has 96 kb (vs 128 kb)

 

[FlameTail]

Someone should run GPU-Z next...

 

[SpudLobby]

Idk about Acer. My impression is that their products are generally lower quality and less reliable.

The Lenovo Yoga on the other hand, is a premium and established line.

The Lenovo Yoga and the Asus Vivobook 15 especially are actually pretty great.

Reduced latency.

didn’t make any sense for this. They have a huge L2. AMD and Intel have a need for that with a larger L1 that QC wouldn’t with this setup imo

 

[SarahKerrigan]

Since we're doing wild guessing, here's something that matches 2.8 MB: 2x128 KB L1 + 2×16 KB L0. Or 32 KB uop cache. Anyway 2.8 MB = 10x288 KB, not 10x280 KB.

I still have my doubts about that 2.8 MB figure.

I think this is plausible, given Qualcomm's (though not Nuvia's) history of doing L0 caches on Krait, Kryo, and Falkor.

Could also be 192+96 L1. Not terribly far off from Apple's 192+128.

EDIT: Just saw the posts above. I guess I made a good guess, lol.

 

[Gerard Williams]

I think this is plausible, given Qualcomm's (though not Nuvia's) history of doing L0 caches on Krait, Kryo, and Falkor.

Could also be 192+96 L1. Not terribly far off from Apple's 192+128.

EDIT: Just saw the posts above. I guess I made a good guess, lol.

L0 caches are a waste of energy.

 

[SarahKerrigan]

L0 caches are a waste of energy.

Hmm, why? I've at least seen simulations where a 4K+4K direct-mapped or 2-way 1-cyc-load-to-use L0, with a larger L1 in the 5 cyc range, comes out looking pretty good. (Heck, even Itanium2's tradeoff - 16+16 4-way L1 @ 1 cyc, 256K+1M L2 @ 5-7 cyc - isn't terribly far off from being an L0+L1, though Intel never called them that.)

 

[POWER4]

Hmm, why? I've at least seen simulations where a 4K+4K direct-mapped or 2-way 1-cyc-load-to-use L0, with a larger L1 in the 5 cyc range, comes out looking pretty good. (Heck, even Itanium2's tradeoff - 16+16 4-way L1 @ 1 cyc, 256K+1M L2 @ 5-7 cyc - isn't terribly far off from being an L0+L1, though Intel never called them that.)

Correct me if I'm wrong, but would a small (say 4K or even 16K) albeit fast L0 cache really make a difference to help keep the EUs of a big core such as Apple M or X Elite fed? My guess is that, given its small size, the infrequent hit does not compensate for the penalty of misses.

 

[FlameTail]

L0 caches are a waste of energy.

IT IS HIM!

L0 caches are a waste of energy.

And this is true for micro-op caches as well, right?

Wide decoder is better than [narrow decoder + microOP cache]

 

[FlameTail]

Snapdragon X Elite shows 2X higher multi-threaded performance than Apple M2 in new benchmarks — Apple M2 is faster in single-threaded performance, though

A Snapdragon X Elite-powered Microsoft Surface Laptop 7 scored 7,923.5 points in multi-threaded benchmark results, beating the Apple M2's 3,822.3 points.

www.tomshardware.com

Snapdragon X Elite shows 2X higher multi-threaded performance than Apple M2 in new benchmarks — Apple M2 is faster in single-threaded performance, though

There wouldn't be headlines like this if Qualcomm had not to made the X1E-78-100 without Core Boost.

 

[adroc_thurston]

There wouldn't be headlines like this if Qualcomm had not to made the X1E-78-100 without Core Boost.

It's not that they intentionally made it, it's just that binning is hard.
You really should stop taking Intel/AMD style of speedbinning for granted.

 

[gdansk]

You really should stop taking Intel/AMD style of speedbinning for granted.

Really, you don't think most of these could go over 3.4GHz?
But I guess it serves as a waste basket anyway

 

[adroc_thurston]

Really, you don't think most of these could go over 3.4GHz?

Not at acceptable power at least.

But I guess it serves as a waste basket anyway

That's the SKU they actually sell at large lmao.

Again, the way x86 boys bin their CPUs is truly magical.
Shipping what used to be strictly "IBM z/stuff clocks" in laptops is cuckoo bananas.

 

[FlameTail]

Again, the way x86 boys bin their CPUs is truly magical.
Shipping what used to be strictly "IBM z/stuff clocks" in laptops is cuckoo ba

Apple is shipping 4.5 GHz in an iPad

 

[FlameTail]

Really, you don't think most of these could go over 3.4GHz?
But I guess it serves as a waste basket anyway

If you look at Qualcomm's official curve, the power increases by 50% when going from 3.8 -> 4.2 GHz [~9W -> ~14W]

 

[adroc_thurston]

Apple is shipping 4.5 GHz in an iPad

Chefe that's N3e.
If I put Z5 on N3e, you would weep at how fast the thing runs.

If you look at Qualcomm's official curve, the power increases by 50% when going from 3.8 -> 4.2 GHz

Those curves are overly optimistic as is.

 

[eek2121]

Apple never claimed to invent Tandem OLED. Don’t people know by know who invents new display tech?

I don’t think anyone ever claimed Apple invented anything. They are a premium device maker, however, and they should be more on the ball with this stuff.

No.

It doesn't make sense. Running distributed AI workloads on CPU/GPU/NPU is like gaming with SLI. And besides the GPU consumes a ton of power, so it's not efficient run it that way.

https://www.reddit.com/r/hardware/comments/1c70h6n/comment/l04l8zz

NPUs will be used for local tasks like video effects, photo and sound editing, etc. These things aren’t doing anything remotely.

Just because GPUs CAN use hundreds of watts, doesn’t mean they always do. Not every task makes a GPU go bonkers.

Microsoft uses the GPU when virus scanning and other parts of Windows are beginning to use it as well.

X Elite X1E-78-100 CPU-Z

https://twitter.com/x/status/1794202815369490738

Nice, natively, 1T performance is right up there with my 7950X.

Emulated performance is…well a 27%

If you look at Qualcomm's official curve, the power increases by 50% when going from 3.8 -> 4.2 GHz [~9W -> ~14W]
View attachment 99706

Yep, high clocks increase power consumption. I really do question how they are going to compete with Zen 5 considering they are already behind Zen 4 in terms of performance.

 

[FlameTail]

I don’t think anyone ever claimed Apple invented anything. They are a premium device maker, however, and they should be more on the ball with this stuff.

Tandem OLED is the industry name. Apple is branding it as "Ultra Retina XDR".

Nice, natively, 1T performance is right up there with my 7950X.

Haha. That's not even the fastest X Elite version. The X1E-78-100 maxes out at 3.4 GHz. The X1E-80-100 goes to 3.8 GHz, and the X1E-84-100 to 4.2 GHz.
There's also the special X1E-84-1DE, which is exclusively found in the Snapdragon Dev Kit mini-pc, and it goes up to 4.3 GHz.

 

[igor_kavinski]

Haha.

Don't act so smug. QC is making big mistakes. You don't conquer a market with premium devices. You do so with a broad range of price points from $300 to $999. More than that is for professionals or gamers. Microsoft/QC are dumb to think that people will max out their credit cards to afford AI stuff. The average person does not have to lie so much like salespeople/sales engineers/marketing folks/scummy businessmen that they need to keep a database of their lies in their head to avoid their lies getting caught so there isn't enough space in their heads for the mundane stuff like remembering appointments or retaining important details from their emails that they read already. Only these people have a use case for Microsoft Recall at a $1000+ price point. Or maybe imposters who will try to act like executives who know what they are doing.

 

[eek2121]

I was referring to the Apple chip. My 7950X scores 750 for ST, putting it just ahead, for MT it destroys the chip.

 

[igor_kavinski]

My 7950X scores 750 for ST

I think CPU-Z is a bit partial to Intel. I'm getting 840 with my 12700K @ 5.3 GHz on air, with DDR5-6000 CL28. Maybe you should investigate tuning your RAM a bit?

 

[adroc_thurston]

Tandem OLED is the industry name

Dual layer actually.

You don't conquer a market with premium devices. You do so with a broad range of price points from $300 to $999.

well there's Purva but it's a meme.

 

[SarahKerrigan]

Dual layer actually.

well there's Purva but it's a meme.

What is with pro-AMD folks from saying "lol it's a meme" about literally everything non-AMD? It's annoying and does not contribute to discourse.

 

[adroc_thurston]

What is with pro-AMD folks from saying "lol it's a meme" about literally everything non-AMD?

Because it's a meme.
Not viable performance levels for 2025 with nothing (hint: slap a modem innit) to compensate for it.

It's annoying and does not contribute to discourse.

Some parts are a meme.