Discussion Qualcomm Snapdragon Thread

[FlameTail]

Both employ former Anandtech writers. But I still am hopeful Qualcomm will talk about this chip at HotChips.

The similarity is uncanny.

 

[FlameTail]

This is a good naming scheme.
Take note AMD: generation means generation not release year.

Snake oil and core truths...

Intel accuses AMD of selling snake oil CPUs in now deleted attack

AMD rebranded its CPUs last year and Intel isn’t happy.

www.theverge.com

 

[FlameTail]

What do you guys think about this launch speed advantage?

Look how the X Elite pulls higher than the X Plus at the top of the curve.

 

[FlameTail]

X Plus' GPU consumes only 20W, about 33% less than X Elite's.

X Plus = 20W = 3.8 TFLOPS
X Elite = 30W = 4.6 TFLOPS

Clearly there has been a large frequency reduction to cut the power consumption by 1/3rd, despite only reducing performance by ~17%

 

[FlameTail]

I am yet to confirm the authenticity of this slide, but it is showing the X Elite consuming almost 70W in Cinebench 2024 MT, which is far higher than the 40W in Geekbench 6 MT.

 

[FlameTail]

Chips and Cheese just dropped an article reviewing the GPU in the Snapdragon 8cx Gen 3.

Sizing up Qualcomm’s 8cx Gen 3 iGPU

Qualcomm has long harbored ambitions to move up into higher performance segments. Their current Snapdragon 8cx Gen 3 tries to make Qualcomm’s laptop dreams come true, and features in products…

chipsandcheese.com

This is the first time they are reviewing a Snapdragon PC SoC. Enjoy!

PS: I hope they review the Snapdragon X Elite too, when it does arrive!

PS2: Damn, we really had a lot of Qualcomm news today!

 

[DrMrLordX]

I am yet to confirm the authenticity of this slide, but it is showing the X Elite consuming almost 70W in Cinebench 2024 MT, which is far higher than the 40W in Geekbench 6 MT.

GB6 MT is kind of a joke by this point, isn't it? I wouldn't expect it to draw much power.

 

[FlameTail]

@adroc_thurston

It is rumoured that Microsoft will raise the NPU requirement 100 TOPS in 2026. Is this correct? Do you have any more info?

 

[FlameTail]

Looks like Qualcomm's marketing department is well funded, atleast;

 

[Henry swagger]

Qualcomm will be exposed by independent reviews

 

[ikjadoon]

Unsure if I missed the discussion on this, but Qualcomm interestingly ever so slightly lowered their memory speeds for the X Elite.

Compared to their October demos, however, there are a couple of important points to point out – areas where the chip specs have been downgraded. First and foremost, the peak dual core clockspeed on the chip (what Qualcomm calls Dual Core Boost) will only be 4.2GHz, instead of the 4.3GHz clockspeeds we saw in Qualcomm’s early demos. The LPDDR5X memory frequency has also taken an odd hit, with this chip topping out at LPDDR5X-8448, rather than the LPDDR5X-8533 data rate we saw last year. This is all of 85MHz or 1GB/second of memory bandwidth, but it’s an unexpected shift since 8448 isn’t a normal LPDDR5X speed grade.

The GB6.2 1T perf guidance has also lowered and Qualcomm is no longer mentioning their previous 3.2K runs. These are all presumably in the 80W device TDP config.

October 23, 2023:
QCOM says Oryon hits 3227 pts (100%) - ? clocks on ?

October 30, 2023:
QCOM says SXE hits 2939 to 2979 pts (91% to 92%) - 4.3 GHz boost on Windows
QCOM device shows SXE hitting 3236 (100%) - 4.3 GHz boost on Linux

April 24, 2024:
QCOM says SXE hits 2850 to 2900 (88% to 90%) - 4.2 GHz boost on ?

I'm all for limiting boost clocks, but funny Qualcomm promoted this 3,227 score beating its competitors, when Qualcomm's final shipping SoC will either be barely faster or may well lose:

Maybe only matching M2 Max, actually.

Maybe actually quite a bit slower than a heavily boosted i9 Raptor Lake

Of course, the headline (much lower power at equal perf) is great, but that bottom chart is becoming more suspicious, particularly when as the launch date moves closer, Qualcomm's internal benchmark scores keep getting lowered.

EDIT: 2024 typo

 

[igor_kavinski]

Maybe tweaking the SoC v/f curve before launch for better battery life?

 

[FlameTail]

Maybe tweaking the SoC v/f curve before launch for better battery life?

Yeah, how did the they reduce the power consumption by 10W...

October 2023


March 2024


April 2024

 

[SpudLobby]

And this is why I have been saying all along that X Elite G2 and it's Pegasus core should not increase the clock speed, and should arguably decrease it. Pushing the clocks up so hard is antithetical to a philosophy of efficient computing. In such a philosophy, performance gains should come from IPC increase, and not clock speed increases.

Clocks going up to follow node gains is fine, pushing further not so much.

 

[SpudLobby]

Yeah, how did the they reduce the power consumption by 10W...

October 2023
View attachment 97766

March 2024
View attachment 97768

April 2024
View attachment 97769

You’re reading too far into this without thinking through it. They didn’t, they’re just not showing it at those points in
the new graphs.

 

[SpudLobby]

This is a good naming scheme.
Take note AMD: generation means generation not release year.

Yeah I agree. AMD’s is a mess.

 

[FlameTail]

View attachment 97750
I am yet to confirm the authenticity of this slide, but it is showing the X Elite consuming almost 70W in Cinebench 2024 MT, which is far higher than the 40W in Geekbench 6 MT.

This slide is causing some controversy in certain circles.

The 70W power consumption is indeed excessive.

 

[Tup3x]

Nope...

 

[FlameTail]

https://www.reddit.com/r/macbookpro/comments/19aikto/m3_pro_vs_m3_max_power_usage

.

During a Cinebench 2024 test, the M3 Pro on normal power mode used a max of 29w on the CPU & 17w on the GPU. While the 16-core M3 Max used a max of 54w on the CPU and 33w on the GPU.

Apple M3 Pro & M3 Max analysis - Apple has significantly upgraded its Max CPU

Notebookcheck performance and efficiency analysis of Apple's new SoCs, the M3 Pro and the M3 Max, compared with AMD, Intel and Qualcomm.

www.notebookcheck.net

The small M3 Pro with 11 CPU cores consumes around 24 watts (M2 Pro with 10 cores: around 27 watts); the M3 Pro with 12 CPU cores consumes about 27 watts (M2 Pro with 12 cores: around 34 watts). More performance with lower consumption naturally means significantly better efficiency, and the new M3 Pro is clearly the leader in this respect.

We have so far only been able to test the top-of-the-range version of the M3 Max with 16 CPU cores and its performance advantage over the M3 Pro and the old M2 Max is a whopping 54-56 %, meaning Apple has finally gotten rid of a major point of criticism that the old models had. Although its power consumption has also increased significantly (M3 Max up to 56 watts; M2 Max up to 36 watts), its efficiency is still slightly better than the M2 Pro/M2 Max — but the M3 Max didn't come close to the M3 Pro's efficiency in our multi-core tests.

So X Elite consumes more power than M3 Max in CB2024, while also delivering less performance.

 

[igor_kavinski]

So X Elite consumes more power than M3 Max in CB2024, while also delivering less performance.

Time to change out of the QC T-shirt and start wearing the Apple cheerleading costume

 

[SpudLobby]

Unsure if I missed the discussion on this, but Qualcomm interestingly ever so slightly lowered their memory speeds for the X Elite.



The GB6.2 1T perf guidance has also lowered and Qualcomm is no longer mentioning their previous 3.2K runs. These are all presumably in the 80W device TDP config.

October 23, 2023:
QCOM says Oryon hits 3227 pts (100%) - ? clocks on ?

October 30, 2023:
QCOM says SXE hits 2939 to 2979 pts (91% to 92%) - 4.3 GHz boost on Windows
QCOM device shows SXE hitting 3236 (100%) - 4.3 GHz boost on Linux

April 24, 2023:
QCOM says SXE hits 2850 to 2900 (88% to 90%) - 4.2 GHz boost on ?

I'm all for limiting boost clocks, but funny Qualcomm promoted this 3,227 score beating its competitors, when Qualcomm's final shipping SoC will either be barely faster or may well lose:

Maybe only matching M2 Max, actually.

Maybe actually quite a bit slower than a heavily boosted i9 Raptor Lake

Of course, the headline (much lower power at equal perf) is great, but that bottom chart is becoming more suspicious, particularly when as the launch date moves closer, Qualcomm's internal benchmark scores keep getting lowered.

https://www.reddit.com/r/macbookpro/comments/19aikto/m3_pro_vs_m3_max_power_usage

.

Apple M3 Pro & M3 Max analysis - Apple has significantly upgraded its Max CPU

Notebookcheck performance and efficiency analysis of Apple's new SoCs, the M3 Pro and the M3 Max, compared with AMD, Intel and Qualcomm.

www.notebookcheck.net

So X Elite consumes more power than M3 Max in CB2024, while also delivering less performance.

It’s the M3 Max. It has 4 more cores and is on N3.

Also I would expect QC to do more poorly on CB24 power curves vs Intel and AMD than GB6, it makes sense. Cinebench is designed perfectly to benefit from HT. GB5 MT would be more interesting.

 

[coercitiv]

Of course, the headline (much lower power at equal perf) is great, but that bottom chart is becoming more suspicious, particularly when as the launch date moves closer, Qualcomm's internal benchmark scores keep getting lowered.

Making claims when the product is in the lab is easy. It gets increasingly more complicated as the launch window narrows down and the chips must match the specs in the wild. This is the reason people begin raising eyebrows as QC continue their barrage of marketing slides that we're supposed to take at face value.

 

[FlameTail]

Also I would expect QC to do more poorly on CB24 power curves vs Intel and AMD than GB6, it makes sense. Cinebench is designed perfectly to benefit from HT.

Apple does pretty well in CB2024;

https://www.cpu-monkey.com/en/cpu_benchmark-cinebench_2024_multi_core

The M3 Max (12P+4E) rivals the Ryzen 7945HX (16P/32T). Both do about ~1600 points.

GB5 MT would be more interesting.

Isn't GB5 deprecated?

 

[SpudLobby]

Apple does pretty well in CB2024;

https://www.cpu-monkey.com/en/cpu_benchmark-cinebench_2024_multi_core

The M3 Max (12P+4E) rivals the Ryzen 7945HX (16P/32T). Both do about ~1600 points.

If you would read, I addressed that. They have a CPU on a better node and with superior perf/W still, and even Qualcomm beats AMD on a per-core basis by a ton, so that’s not surprising. However HT still gives them a significant boost on this app. Not hard.

Isn't GB5 deprecated?

Yeah, but the way they do MT in GB5 is different.

 

[SpudLobby]

IMO the X Plus still looks fine. That 2400 GB6 ST can be had for much less power than Hawk Point.

So responsiveness IRL can compete for the likely segments involved and with vastly better battery life.

That said, it was still stupid to do this RE: ST, but pricing and battery life will determine how much it hurts.