Discussion Qualcomm Snapdragon Thread

[Covfefe]

The M4 Max achieves 2000+ points with a power consumption of 57-60 watts.

Yeah, but the M4 Max is over twice the size as the X2 Elite and has a 512bit memory bus. It's great for what it is, but it's not remotely in the same class.

 

[jdubs03]

Yeah, but the M4 Max is over twice the size as the X2 Elite and has a 512bit memory bus. It's great for what it is, but it's not remotely in the same class.

Comparing to the X2 Elite Extreme (12P+6E)*. The M4 Max has 16 Cores (12P+4E). It’s as close a comparison as can be performed.

 

[Covfefe]

Comparing to the X2 Elite Extreme (12P+6E)*. The M4 Max has 16 Cores (12P+4E). It’s as close a comparison as can be performed.

The core count being the same doesn't make it a fair comparison. The M4 Max Macbooks start at $3200. Cache and memory bandwidth can affect performance too.

 

[jdubs03]

The core count being the same doesn't make it a fair comparison. The M4 Max Macbooks start at $3200. Cache and memory bandwidth can affect performance too.

That’s on Qualcomm and their partners. I wouldn’t be surprised if these X2 Extreme Elites go into $2000+ laptops.

What would be a fair comparison otherwise?
It’s highest end vs highest end, same p core count, similar overall core count.

Plus by the time their out the M5 Max will be out and the direct comparison might be the M5 Pro if you’re focusing strictly on product price point.

 

[Covfefe]

That’s on Qualcomm and their partners. I wouldn’t be surprised if these X2 Extreme Elites go into $2000+ laptops.

What would be a fair comparison otherwise?
It’s highest end vs highest end, same p core count, similar overall core count.

Plus by the time their out the M5 Max will be out and the direct comparison might be the M5 Pro if you’re focusing strictly on product price point.

There is no totally fair comparison, but the closest one would be the M4 Pro vs X2 Elite Extreme. Qualcomm's cores are smaller than Apple's P-cores. Qualcomm can literally fit 18 cores in the same area as Apple's 10+4 cores. If anything that comparison favors Apple because the M4 Pro has more LLC and memory channels.

If you want to compare highest end vs highest end then don't let me stop you. I'm just pointing out that they are wildly different in production cost and end product price point.

 

[jdubs03]

There is no totally fair comparison, but the closest one would be the M4 Pro vs X2 Elite Extreme. Qualcomm's cores are smaller than Apple's P-cores. Qualcomm can literally fit 18 cores in the same area as Apple's 10+4 cores. If anything that comparison favors Apple because the M4 Pro has more LLC and memory channels.

If you want to compare highest end vs highest end then don't let me stop you. I'm just pointing out that they are wildly different in production cost and end product price point.

You bring up an interesting point about the core size. I’d imagine Apples’ p-cores are bigger because of additional functionality (power gating, other reasons, etc.). I have wondered how Qualcomms’ p-core is as performant as it is with being a decent bit smaller. What are the likeliest reasons for that? That to me looks to be an advantage.

 

[ToTTenTranz]

2048 FP32 ALUs at 1.85GHz means 7.6 TFLOPs. That's a lot.

 

[Covfefe]

You bring up an interesting point about the core size. I’d imagine Apples’ p-cores are bigger because of additional functionality (power gating, other reasons, etc.). I have wondered how Qualcomms’ p-core is as performant as it is with being a decent bit smaller. What are the likeliest reasons for that? That to me looks to be an advantage.

I think of a big part of it is Apple has wider P cores with higher IPC. They have a huge lead in single core performance in a lot of benchmarks. Other than that I'm not sure. More/better power gating seems likely too.

 

[poke01]

Best to wait for M5 Pro to compare and to check in non bandwidth benchmarks.

Still I would say a future M5 Pro would use 60 watts to achieve the score the X2 Elite is scoring.

 

[mvprod123]

Best to wait for M5 Pro to compare and to check in non bandwidth benchmarks.

Still I would say a future M5 Pro would use 60 watts to achieve the score the X2 Elite is scoring.

*X2 Elite Extreme

 

[Tup3x]

I guess translation/emulation was never really the issue.

 

[Jan Olšan]

When games are GPU bound, obviously translation doesn't matter much. The GPU driver itself is native, even. GPU operation is not afflicted at all. All shaders obviously run in native mode. Even shader compiler would be native code.

The only problems are the availability of DRM/anticheat tools for the platform and (a harder problem) the GPU driver stack and ecosystem. Where it's a long road to even catch up with Intel, not to mention AMD and Nvidia.

 

[Raqia]

When games are GPU bound, obviously translation doesn't matter much. The GPU driver itself is native, even. GPU operation is not afflicted at all. All shaders obviously run in native mode. Even shader compiler would be native code.

The only problems are the availability of DRM/anticheat tools for the platform and (a harder problem) the GPU driver stack and ecosystem. Where it's a long road to even catch up with Intel, not to mention AMD and Nvidia.

That their driver weighs in at ~130MB vs the ~800MB for nVidia ones seems to indicate MUCH less game specific support in the form of game specific shader replacement.

They are improving GPGPU support and moving to a monthly cadence for GPU updates starting with this gen:

Inside Snapdragon X2 Elite: Qualcomm’s Next-Gen PC Chip Promises Big Gains

The Qualcomm Snapdragon X2 Elite improves upon its predecessor in almost every way, with improved performance and efficiency virtually across the

hothardware.com

In terms of GPGPU workloads, the Adreno X2 has native driver support for WCR (Windows Copilot Runtime), BF16 support, local memory broadcast, and general shuffles, which allows full shuffling between the threads.

On the software side, Qualcomm states it is planning to move to a monthly release cadence, although it didn’t announce when that would begin. As it stands now, the Adreno X2 driver has native support for DX 12.2, OpenCL 3.0 and SYCL, and a Native Vulkan 1.4 driver is coming in Q1 '26. Qualcomm is also working toward full kernel-level anti-cheat compatibility and hopes that by the launch of Snapdragon Elite X2 all major anti-cheat technologies will be natively supported.

Seems like they didn't see it fit to dump resources into driver development for the prior gen. I don't think there was much they could do at the driver level to help the aging, more tile oriented 7xx Adreno architecture in the XE1 gen (which had been delayed by the ARM lawsuit...). GPGPU performance was abysmal and though they bolted on an additional layer of cache in the hierarchy and boosted sizes, performance was poor compared to competing integrated and low end discrete solutions.

Also interesting is the big emphasis on industrial grade security for this part such as the dedicated 4G feed for security functionality. It seems to hint that this may be part of the build out for enabling automotive usage:

Qualcomm and BMW Group Unveil Groundbreaking Automated Driving System with Jointly Developed Software Stack | Qualcomm

Discover Qualcomm & BMW's groundbreaking Snapdragon Ride Pilot, a scalable AI-enabled automated driving system debuting in the BMW iX3. Validated in 60 countries, it's set to expand to over 100 by 2026.

www.qualcomm.com

BMW's iX3 Just Kicked Off Qualcomm’s Big "Flywheel" Bet on Automated Driving

Chipmaker’s Ride Pilot system that debuted on the new BMW iX3 will collect data from different manufacturers and a variety of vehicles.

www.motortrend.com

Given the compute requirements and timing, I would guess a derivative of this chip is being used in the iX3...

 

[adroc_thurston]

I guess translation/emulation was never really the issue.
View attachment 134182

30% perf delta with STX1 despite N3X and a near-Halo levels of membw is lol

 

[Covfefe]

I guess translation/emulation was never really the issue.

AVX2 support in Prism recently came out of beta. It's safe to say that these X2 benchmarks have it enabled. That could make a huge difference.

 

[adroc_thurston]

AVX2 support in Prism recently came out of beta. It's safe to say that these X2 benchmarks have it enabled. That could make a huge difference.

Most games aren't particularly SIMD-heavy.

 

[Covfefe]

Most games aren't particularly SIMD-heavy.

Some of the games on that slide won't run at all without AVX2.

 

[adroc_thurston]

Some of the games on that slide won't run at all without AVX2.

Indeed but SIMD is not the performance limiter there.

 

[jdubs03]

Best to wait for M5 Pro to compare and to check in non bandwidth benchmarks.

Still I would say a future M5 Pro would use 60 watts to achieve the score the X2 Elite is scoring.

I think less, but since the base M5 is less efficient than the M4, it isn’t impossible. Hopefully we don’t have to wait too long.

 

[CouncilorIrissa]

OK so the 128KB L1D is just Geekerwan being wrong apparently. L1D is still 96KB.

 

[Gideon]

Didn't see the computerbase.de benchmarks listed here:

Snapdragon X2 Elite im Benchmark: Alle drei SoCs im Vergleich mit Apple, AMD und Intel

Nach Benchmarks mit dem Snapdragon X2 Elite Extreme konnte ComputerBase jetzt die Ableger X2E-88-100 und X2E-80-100 bei 22 Watt TDP testen.

www.computerbase.de

Qualcomm's reference designs tend to always perform better than actual laptops, but still relatively impressive