Discussion Qualcomm Snapdragon Thread

[mikegg]

Very good catch that the omission of Apple seems to be an oddly specific one.

It's not possible to know exactly how much Apple spends on chip R&D.

 

[moinmoin]

It's not possible to know exactly how much Apple spends on chip R&D.

I doubt you get the exact recipes for any company, "only" the margin of error is different.

 

[mikegg]

I doubt you get the exact recipes for any company, "only" the margin of error is different.

But the other companies are full semiconductor companies. So all of their R&D, in theory, goes into designing and making semiconductors. Meanwhile, Apple's R&D can go into displays, glass, metals, speakers, consumer software, and all sorts of random things.

 

[moinmoin]

But the other companies are full semiconductor companies. So all of their R&D, in theory, goes into designing and making semiconductors.

In theory, indeed. @Doug S posted numbers from 2017 which included Samsung and Toshiba, omitted by "AITechInvesting" in the current numbers (which also omits MediaTek curiously enough, unless its R&D is too small despite Dimensity?). In the thread he wrote he's still trying to work out R&D numbers for further companies. So the more semi- or unrelated stuff companies work on the higher the margin of error. In the end it's all guesstimate.

What would be very informative is foundry R&D being split from chip design R&D. Maybe we get it once IFS is more properly separated.

 

[FlameTail]

Sir, this is the Qualcomm thread. Why are we talking about Apple here?

🔥 Breaking News! 🔥.

Did anybody even read this?

Edit: In all seriousness I have no issue with anyone talking about Apple or any other company is this thread. However I do feel disgruntled if those alternate discussions end up burying messages related to the thread topic- Qualcomm Snapdragon.

 

[moinmoin]

We weren't talking about Apple per se, we were talking about the chart you posted:

View attachment 89947

Wow. I wonder, what did Qualcomm spend that much R&D on?

What are they cooking?

Source:https://x.com/AITechInvesting/status/1732070071566999567?s=20

If Intel were listed without R&D spent on its foundry business (which I'd expect to make up more than half of Intel's total R&D expenses) Qualcomm would likely be ahead.

 

[Nothingness]

If Intel were listed without R&D spent on its foundry business (which I'd expect to make up more than half of Intel's total R&D expenses) Qualcomm would likely be ahead.

If one removes money Qualcomm gets from patents, I'm not sure they'd be ahead.

 

[FlameTail]

Windows Hudson Valley

 

[gdansk]

Windows Hudson Valley

Are you OK?

 

[FlameTail]

Are you OK?

Yeah-no.
Thanks for asking.

 

[FlameTail]

A redditor based on this information believes, Qualcomm isn't designing a ground-up new E-core. Instead it seems we have a Zen4 - Zen4c kind of situation here. If anything, the names "Phoenix-L" and "Phoenix-M" are obvious pointers.

 

[SarahKerrigan]

A redditor based on this information believes, Qualcomm isn't designing a ground-up new E-core. Instead it seems we have a Zen4 - Zen4c kind of situation here. If anything, the names "Phoenix-L" and "Phoenix-M" are obvious pointers.

They did exactly that with Kryo.

 

[FlameTail]

They did exactly that with Kryo.

Hmm. That should be noted.

Kryo Prime, Kryo Gold, Kryo Silver are all using entirely different core designs (Cortex X, A7, A5).

So I guess Phoenix-M could be a ground up new E-core then?

 

[SarahKerrigan]

Hmm. That should be noted.

Kryo Prime, Kryo Gold, Kryo Silver are all using entirely different core designs (Cortex X, A7, A5).

So I guess Phoenix-M could be a ground up new E-core then?

I'm talking about Kryo, the microarchitecture, not "Kryo," the marketing name for Cortex-of-the-week. In the Kryo chips, there was only one uarch, but two clusters, one of which had larger caches and higher clocks.

 

[FlameTail]

So Oryon is not only lacking SVE2, but also SME and SME2 (Scalable Matrix Extension), which ate part of ARMv9.

 

[FlameTail]

https://www.reddit.com/r/hardware/comments/18h5hzu

Reddit post that compiles all the leaks of the Snapdragon 8 gen 4 by Revegnus.

- Will use 2P+6E Oryon CPU
- TSMC N3E fabrication
- GPU does 7200 points in 3Dmark WLE
- CPU hits 10k points in GB6 Multi at 8W

Also solid comment from Vince:

So there are 3 possibilities of what Phoenix-M could be like:

1. Apple-E core; Ground up new design

2. Zen4 -> Zen4C; Same IPC, lower clockspeed, smaller size

3. Cortex X1 -> A78; modified Phoenix-L with reduced area, power and performance.

 

[ikjadoon]

An updated comment from Qualcomm to Digital Trends today: Snapdragon X Elite "outperforms M3 in 1T; outperforms M3 by 21% in nT"

//

That will be interesting to test next year. I presume Geekbench, as that's NUVIA / Qualcomm's benchmark of choice.

1T
M3 : ~3030 pts
SXE (23W device TDP): ~2760 pts
SXE (80W device TDP): ~2959 pts
SXE (80W device TDP; 100% fans): ~3236 pts

nT
M3 : ~11,694 pts
SXE (23W device TDP): ~13,928 pts
SXE (80W device TDP): ~15,235 pts
SXE (80W device TDP; 100% fans): ~17,387 pts

The "SXE beats M3 on 1T" seems to rely on a much higher power draw, but a win is a win. nT is a major win. Though it is 4+4 vs 12+0, if you want nT perf, the SXE delivers in spades.

 

[FlameTail]

An updated comment from Qualcomm to Digital Trends today: Snapdragon X Elite "outperforms M3 in 1T; outperforms M3 by 21% in nT"

//

That will be interesting to test next year. I presume Geekbench, as that's NUVIA / Qualcomm's benchmark of choice.

1T
M3 : ~3030 pts
SXE (23W device TDP): ~2760 pts
SXE (80W device TDP): ~2959 pts
SXE (80W device TDP; 100% fans): ~3236 pts

nT
M3 : ~11,694 pts
SXE (23W device TDP): ~13,928 pts
SXE (80W device TDP): ~15,235 pts
SXE (80W device TDP; 100% fans): ~17,387 pts

The "SXE beats M3 on 1T" seems to rely on a much higher power draw, but a win is a win. nT is a major win. Though it is 4+4 vs 12+0, if you want nT perf, the SXE delivers in spades.

(1) Some people would argue the X Elite should be compared to the M3 Pro, not M3. However, I argue otherwise. If you look at the die size, M3 (TSMC N3B) is 146 mm² and X Elite (TSMC N4P) is about 172 mm².

The M3 has a smaller die but it's on the more expensive N3B node, which also has lower yields. So I guess when it comes to actual cost, both are similar and the M3 might potentially be more expensive than the X Elite.

(2) These numbers of the X Elite are better than Meteor Lake and Hawk Point. Arrow Lake and Strix Point will beat it, but they are coming in late 2024. So I guess the X Elite will still have some thunder to boast about when it arrives in mid-2024.

 

[FlameTail]

I fear that Qualcomm is overhyping the X Elite chip, and it's going to suffer from something I'd call the Cybertruck Effect.

Tesla announced the Cybertruck 4 years ago. But when it did get released a few weeks ago, the specs were lower and the price was higher compared to what was announced 4 years ago. Now don't get me wrong- the Cybertruck is an amazing vehicle, but it has to be said that it's impact today would be greater if Tesla hadn't overpromised years ago.

 

[FlameTail]

I can imagine Qualcomm making a lower end Snapdragon X chip with a 64 bit memory bus.

Targeted squarely at Chromebooks and low-end Windows laptops, which sell in very high volumes.

= LPDDR5T + 64 bit

= 76 GB/s memory bandwith

Not bad, huh?

That's the equivalent of DDR4-4800 dual channel.

 

[FlameTail]

Snapdragon 8 Gen 3's Adreno GPU is as strong as a GTX 1650

 

[Thibsie]

Snapdragon 8 Gen 3's Adreno GPU is as strong as a GTX 1650

Links please ?
Thx

 

[FlameTail]

Links please ?
Thx

I was seeing a reddit thread which mentioned the GTX 1650 is 50% faster than GTX 1050. Quick Google search confirmed it is indeed the case.

We know the 8 Gen 2's GPU matched the GTX 1050 thanks to Geekerwan:

8 Gen 3's GPU is ~50% more powerful than 8 Gen 2's.

So if GTX 1650 is 50% faster than GTX 1050,

We can conclude 8 Gen 3 and GTX 1650 are in the same ballpark.

 

[FlameTail]

GPU

CPU

The power consumption of the X Elite (compared to Apple Silicon) is concerning.

 

[SpudLobby]

View attachment 90633
View attachment 90634
View attachment 90637
GPU
View attachment 90635View attachment 90636
CPU

The power consumption of the X Elite (compared to Apple Silicon) is concerning.

Not necessarily.

The 13800H performance that the Snapdragon matches in CB24 at about 32W total power (65% less) is about a 976 apparently. Qualcomm’s peak performance is about a 1229 on CB24 for the “80W TDP” which here would at least be similar to the actual peak. And then it’s 1022 in the “23W TDP” config (which tells us nothing about the power though we do know the MT boost clocks there should be 3.4GHz).

Anyway, with some margin of error here it’s about 1/4-1/3 more from the 32W point where they match the 13800H to the peak for QC on performance which would make sense and puts you around the 1229 range more or less (or the 32W point is lower rather).

So say the MT is 950 @ 32W. Yeah that’s worse than the M2 Max, but not by much, and take it down to 20W where the M3 is, it’s steep but should still be within 10-15%. That’s all assuming the calculations here are even roughly accurate and the graph is representative.

I’m not that worried tbh.