No Free Lunch?

No Free Lunch?
​

Yesterday Samsung (005930.KS) announced that the new Galaxy S25 flagship line would be based on the new Qualcomm (QCOM) Snapdragon 8 Elite chipset, the first mobile device to use this processor.  This chipset competes directly with Mediatek’s (2454.TT) Dimensity 9400 chipset, which is already the basis for 3 smartphone models from Vivo (pvt) and 3 from Oppo (pvt) that were released in October of last year.  The characteristics of each of those chipsets are listed below, along with a number of benchmarks that compare performance and details on what those benchmarks capture.  However while there are benchmarks that score the AI related performance of smartphones, there does not seem to be a recognized benchmark that rates the performance of the chipsets themselves in terms of AI performance. 
We believe this is due to the integration of the AI into each phone’s hardware and software, which can influence AI performance, along with the fact that the Snapdragon 8 Elite is quite new.  Using just the AI benchmarks for the few models that are based on the Dimensity 9400 (average) and comparing that to the previous Snapdragon 8 Gen 3 chipset, it shows that the Dimensity has just under twice the AI performance of the older Snapdragon, so we have to assume that the 8 Elite Ai benchmark will at least match that once the Galaxy S25 line is benchmarked.  In the interim, the chipset comparisons are the only benchmarks to compare the performance of the processors.

Test Bench Criteria

• GeekBench 6 – Real-world tasks
  • Image Processing – Applying filters, resizing images, etc.
  • Machine Learning – Training & running ML models.
  • Augmented Reality – Processing & rendering AR.
• AnTuTu 10 – Combined Workloads
  • Image Processing – Rendering 3D scenes
  • Video Editing – Processing & encoding video files
  • Multitasking – Switching between apps & performing multiple tasks
• Wild Life Extreme (GPU)
  • Frame Rate – Higher frame rate implies smoother graphics
  • Graphics Quality – Details, texture, lighting, effects
• AI-Benchmark
  • Object Recognition – Object, facial and scene recognition
  • OCR – Character recognition
  • Effects – Deblur, Denoise, Bokeh

While the structure of these two chipsets are different, there performance on the test bench is very close, and that seems to indicate that Mediatek is getting close to closing the gap in terms of phone performance.  However, as AI functions are added to these devices, the need for larger die size, higher densities, and more memory affect the cost and yield as the NPU becomes a larger part of the chipset.  Ideally the NPU should have a dedicated core with an instruction set specifically designed for the NPU but the complexity of the design and production of such a chipset is not economically feasible yet.  For now chipset NPUs share the cores with device processes, but the current drive toward on-device Ai processing intensifies the impact on overall device performance.  This leads to eventual chipset designs with a dedicated NPU core or a hybrid approach with a dedicated NPU core and main core sharing the load.  With each of these improvements comes more manufacturing complexity and cost, but if AI is going to drive a mobile device replacement cycle, these improvements will have to continue with each chipset iteration.
Each of the chipsets below contained an NPU, and while NPU ‘size’ is a function of a number of factors, the NPU affects cost.  We expect it will continue to be an increasingly large component of the overall cost of subsequent chipsets, and, if the Samsung Galaxy line is any indication, smartphone brands have yet to try to recoup the incremental cost.  We expect that, at the smartphone level, consumers will assume that the added functionality that AI adds is already built into the cost of the device.  They have learned over the years that the incremental changes in each smartphone generation (50MP camera to a 100MP camera, or 6.7” display to a 6.8” display) do not seem to add to the device cost, so why should they think that AI is different, especially as AI looks to the user as software, not a visible physical change…  Another slice out of smartphone brand margins.