Qualcomm Snapdragon 8 Gen 2 performance overview

Benchmarks are of little or no use in real experience, especially if we think about systems that use less and less CPU and GPU to redistribute part of the hard work to different NPUs or accelerators.

The value expressed by the number does not always reflect the real experience, or rather, it almost never reflects it because we have reached a point where the power of the smartphone’s processor far exceeds what may be the needs of the software itself.

However, it is interesting to understand what evolution is in relation to the production process (and its improvement), and above all what kind of jump can be expected from generation to generation.

We’ve had a Snapdragon 8 Gen 2 reference design in our hands for a few hours, with Qualcomm’s new SoC for flagship phones fitted inside what’s more of a reference platform intended for testing, and we’ve done some quick measurements.

Let’s see the results.

The CPU is getting better, but it’s still chasing

Qualcomm has changed the CPU architecture from the classic 4+3+1 combination, where 1 is the most powerful core, to a new 3+4+1 combination.

The node has not changed compared to Snapdragon 8+ Gen 1, we are still in front of TSMC N4 (5 nm) but the cores have changed, instead of Cortex-x2 there is Cortex-x3, and the clock frequency is .

Since we didn’t have access to more sophisticated solutions during the Snapdragon Summit for benchmarking / watts, we limited ourselves to the classic GeekBench, with single-core and multi-core benchmarking.

As can be seen from the graph, the score shows a fairly balanced increase compared to the 8+ Gen 1 version, where we’re around 12% for a single core. It should be noted that only the most powerful core, Cortex-x3, is scaled into a single core, and ARM has dropped many of the legacy 32-bit instructions from that core making it smaller.

The result in multicore shows a more pronounced increase, but it is clearly a result affected by the change of the structure: there are two additional high-performance cores. Qualcomm, using Cortex-x3 and 64-bit Cortex A715, freed up some space to fit another high-performance core in place of one high-efficiency core.

However, if we look at the competition, we can see that the faster core of Snapdragon 8 Gen 2 is still slower than the Icestorm core of Apple A14, the core released in 2020 for the iPhone 12. And much slower than the current one. A 16.

The reason Qualcomm bought Nuvia is clear: the company is well aware of ARM’s Cortex cores They have no way of competing in performance with those built in Cupertino.

The demo also comes from performance benchmarks for in-browser javascript engines, essential in a world made up of hybrid applications based strictly on a JS framework: the difference is massive.

The biggest jump is in the neural processor, but the next Exynos will line up

Qualcomm has put its hand forward: the biggest jump in performance will be visible above all in machine learning. In this case, the only standard available is MLPerf, which is the standard developed by a consortium of manufacturers and the only reliable reference for comparative comparisons.

Ironically, when we were looking at the results in MLPerf from the previous Snapdragon for a comparison, we discovered that Samsung had sent the results of the Exynos 2300 in development: behind an existing S22 processor and has its initials. In fact, it also has a performance that seems to be in line with that of the new Qualcomm SoC in some segments, that is, calculating that we are comparing a final processor with closed drivers from a processor that has not yet been announced, and that today was only the subject of rumors.

Here is the data.

Light and shadows on the GPU: results for verification

The most important novelty of the new Snapdragon 8 Gen 2 is the presence of a GPU, Adreno 740, compatible with hardware ray tracing.

However, in the paper Qualcomm gave us, 3D Mark and GFX Bench were present as the only graphics benchmarking applications. In comparisons with the previous generation, the jump is there, but it’s not at all exaggerated especially with GFXBench: Normal increase in perspective 10%.both in terms of OpenGL and Vulkan APIs.

We did some GPU temperature analysis and everything was normal

However, we didn’t find ourselves unprepared: a ray-tracing modular program exists and it was created by the developer of the ray-tracing rendering engine, Tellusim Engine.

This software works on all operating systems, first looking for any hardware acceleration, then if it isn’t there, try software acceleration and finally raster conversion.

We tried running this program, GravityMark, on an iPhone 14, on a Galaxy Fold with Snapdragon 8 Gen 1 and on a Snapdragon 8 Gen 2. The settings chosen were for all phones 1080p, 50,000 asteroids.

In the case of the Snapdragon 8 Gen 1, with its Adreno 730, the phone struggled a lot during testing, scoring around 10.1fps.

If we switch to Snapdragon 8 Gen 2 the situation improves, the graphics drivers are Vulkan engines and the frame rate goes up to around 17fps. The iPhone 14 Pro is stable, running Metal at 33fps.

Desktop solutions, using this same benchmark, are still light years away, proving that in the end we’re still talking about GPUs that consume a few watts (if all goes well) and can’t compete, despite the fact that we’re still We talk about it “console like games”, with a laptop or desktop GPU.

Qualcomm managers present during the tests told us that they completely ignored the existence of this benchmark and were very surprised by the “low” result, and reserve the right to understand better once they return to the San Diego labs.

To this we also add the rather curious fact that Qualcomm decided to delete the Geekbench results for the “Compute” part of its reference sheet, that is, those related to GPU load. Results that more or less align with what was shown by GravityMark.

So we are facing results that need to be checked carefully, partly because in some cases the Adreno GPU appears to be much faster than the one in Apple’s A14 (GFX Bench Offscreen) and other 3DMark where the result is completely inverted instead.

Below are the values ​​provided by Qualcomm.

Standards Snapdragon 8 Gen 2 reference design results
Geekbench5 Single-core: 1.485 – 1.495
Multi-core: 5,050 – 5,200
Antoto 1,270,000 – 1,280,000
PCMark 18500 – 18900
jet streams 167 – 170
Speedometer 144 – 146
WebXPRT3 219 – 220
GFXBench Manhattan 3.0 off screen: 329 – 332
T-Rex off screen: 481 – 484
Manhattan 3.1 off screen: 224 – 226
Off-screen car chase: 129-130
Aztec Ruins Vulkan High Tier Offscreen: 65
Aztec Ruins OpenGL High Tier Offscreen: 60
Aztec Ruins Vulkan Normal Offscreen Tier: 178
3DMark Wildlife Unlimited: 82 fps
Wildlife Extreme Unlimited: 23 fps
MLPerf (AI) Image rating: 3,915 – 3,920
Body detection: 1,765 – 1,800
v2.0 Image Fragmentation: 945 – 950
Language comprehension: 185
Offline photo rating: 4,980 – 5,020

No miracle, novia needed. Don’t worry, it is still the fastest in the Android world

The architecture hasn’t changed, the kernels haven’t changed much: it was impossible to work miracles, but we didn’t expect miracles either. Basically compared to the Snapdragon 8+ Gen 1, and we took the ‘+’ because it came out so well, the new model’s performance is about 10-15% higher in every area, with the advantage in terms of machine learning being the most important aspect. It’s still by far the best Android SoC and it’s also supposed to drop in consumption this year.

Compared to Apple SoCs, in areas where a reasonable comparison can be made, the new Snapdragon lags behind, but Qualcomm is so aware of this, that they’ve made us understand over and over again that they can expect to close the gap when they do. Start abandoning the cores made by ARM to switch to the Oryon core, which will be used first on PCs and then on all other platforms (smartphones included).

Next year will be interesting, between 3nm and Oryon.

Source link

Leave a Reply

Your email address will not be published. Required fields are marked *