A reproducible, chip-to-chip comparison of what unprivileged user code can
actually do for AI on two flagship mobile SoCs: MediaTek's Dimensity 9500s
(mt6991, POCO X8 Pro Max) and Qualcomm's Snapdragon 8 Elite (SM8750, QRD
reference device on the Qualcomm Device Cloud). Same workloads on both sides:
NPU reachability + int8/fp16 throughput, LLM inference on CPU and GPU
(llama.cpp, Gemma-3 1B Q4_K_M), raw GPU matmul, and on-device training
(ggml, MLP + AdamW). Every number was measured on real silicon; every raw
output is committed under each chip's results/.
| Test | Snapdragon 8 Elite | Dimensity 9500s |
|---|---|---|
| NPU | Hexagon HTP V79 | MDLA (NPU 890) |
| NPU runtime access | QNN 2.48 from Maven Central + public FastRPC — no SDK gate | NeuroPilot 8.2.26 preinstalled as public system libs |
| int8 8-layer FC (×ばつ2048) | 4307 GOPS, 0.50 ms | 1100–1470 GOPS, 1.46 ms |
| NPU fp16 inference | ✅ 1089 GFLOP/s | ✅ runs (unquantified) |
| NPU vs its own CPU (same graph) | ×ばつ | n/a (CPU int8 ref not measured) |
| NPU training | ❌ architecturally impossible | ❌ architecturally impossible |
| CPU | ×ばつ Oryon @4.09 + ×ばつ @2.78 GHz | ×ばつ X925 @3.73 + ×ばつ X4 + ×ばつ A720 |
| Gemma-3 1B Q4 prefill / decode (8T) | 171–173 / 62–64 tok/s | 58–74 / 22–31 tok/s |
| GPU | Adreno 830 | Immortalis-G925 MC11 |
| Working GPU compute API | OpenCL (public lib; Vulkan broken for quantized inference) | Vulkan (OpenCL namespace-blocked) |
| Gemma-3 1B Q4 prefill / decode (GPU) | 717–723 / 41.5–42.5 tok/s (OpenCL) | ~39 / ~27 tok/s (Vulkan) |
| matmul 40963 f32 | 479 GFLOP/s = ×ばつ CPU | 67 GFLOP/s = ×ばつ CPU (coopmat+f32 acc only) |
| transformer-layer training step (GPU vs CPU) | 403 vs 62.5 GFLOP/s = ×ばつ (OpenCL) | ~50 vs 37 GFLOP/s = ×ばつ |
| GPU training (ggml MLP+AdamW) | ✅ Vulkan, no flags, converges identically to CPU | ✅ Vulkan with GGML_VK_DISABLE_COOPMAT=1, converges worse |
| GPU training via the "fast" API | ❌ OpenCL lacks SUM/OPT_STEP_ADAMW kernels (patch started) |
❌ OpenCL blocked entirely |
| Best local-LLM recipe | GPU-OpenCL prefill + CPU decode | CPU prefill + GPU-Vulkan decode |
Both chips tell the same top-level story — the NPUs are real, reachable, and inference-only; the gaps are software — but they invert on every axis of openness: on MediaTek the GPU compute door is Vulkan and OpenCL is walled off; on Qualcomm OpenCL is public and tuned while Vulkan quantized inference is broken. The Snapdragon platform is the more open of the two (public OpenCL, public FastRPC, redistributable QNN), and its silicon leads every measured workload: ×ばつ on NPU int8, ×ばつ on CPU LLM throughput, ×ばつ on GPU matmul.
Full evidence and per-chip narratives:
.
├── README.md # you are here — the comparison
├── LICENSE # MIT
├── dimensity-9500s/ # MediaTek side (POCO X8 Pro Max, Termux, no root)
│ ├── README.md, REPORT.md
│ ├── docs/ scripts/ # provisioning + on-phone toolchain setup
│ ├── npu/ # NeuroPilot probes + MDLA benches
│ ├── inference/ # llama.cpp build+bench, LiteRT-LM status
│ ├── gpu-matmul/ # Vulkan matmul + backward-pass investigation
│ ├── training/ # CPU (PyTorch) + GPU (ggml Vulkan) training
│ └── results/ # raw captured outputs (12 files)
├── snapdragon-8-elite/ # Qualcomm side (QDC QRD, adb, cross-compiled)
│ ├── README.md, REPORT.md
│ ├── docs/ scripts/ # QDC provisioning + one-shot build/run scripts
│ ├── npu/ # QNN/HTP probes + ONNX-Runtime NPU bench
│ ├── training/ # ggml training + matmul + trainstep benches
│ ├── patches/ # ggml OpenCL memset_tensor kernel
│ └── results/ # raw captured outputs (14 files)
└── third_party/ # dependency checkouts (gitignored; scripts populate)
Each side is reproducible with its own harness, matching how each device is reachable:
- Dimensity 9500s (on-phone, Termux): provision per
dimensity-9500s/docs/PROVISIONING.md,
run
scripts/setup_phone.sh, then the per-experiment commands in each directory README (npu/,inference/llama-cpp/,gpu-matmul/,training/). - Snapdragon 8 Elite (cross-compiled, adb): provision per
snapdragon-8-elite/docs/PROVISIONING.md,
then two commands:
export ANDROID_NDK=/path/to/android-ndk-r27d snapdragon-8-elite/scripts/build_host.sh # deps + all cross-builds snapdragon-8-elite/scripts/run_device.sh # push + run the full suite
The Dimensity phone was re-benchmarked with HyperOS Ultimate mode + charging (dimensity-9500s/results-ultimate/): CPU gains modestly (prefill 74 → 79 t/s, retrain steps ~25% faster when cool), but GPU throughput regresses under a 17-minute back-to-back suite (matmul 67 → 45 GFLOP/s) — passive cooling throttles regardless of the mode toggle, and charging adds heat. The Snapdragon QRD reproduced its numbers within ~4–7% across a comparable suite. For sustained on-device AI, thermal headroom matters more than peak-clock modes.
- Shared workloads: NPU = 8 chained int8 FC layers, M=32 K=N=2048 (+ a
single fp16 FC); LLM = llama.cpp
llama-bench -p 128 -n 64on Gemma-3 1B Q4_K_M; matmul = ggml f32mul_matat 1024/2048/40963; training = the same 385-param MLP, AdamW, MSE, 60 epochs. Benchmarks are run twice and reported as ranges. - Different access paths by necessity: retail phone via Termux (MediaTek) vs QRD via adb/NDK (Qualcomm). Both execute as unprivileged user code; the QDC shell happens to be root but nothing measured here uses root-only facilities (the loaded libraries are on each device's public-libraries list).
- Different NPU drivers by necessity: MediaTek's Neuron Adapter is called raw (NNAPI-like C API); Qualcomm's QNN is driven through ONNX Runtime's QNN EP (with CPU fallback disabled) because raw QNN graph composition needs the gated SDK headers. Both measure the same thing: a quantized graph fully resident on the NPU.
- Different llama.cpp builds: MediaTek numbers were taken 2026年07月13日 on
the then-current llama.cpp; Snapdragon numbers on commit
8e8681e(2026年07月27日, pinned inbuild_host.sh). Tokens/s at this model size are dominated by hardware, but treat small cross-chip deltas accordingly. - Thermals: the phone was benched warm-and-cold (ranges reported); the QDC QRD idles cool and varied <4% across runs.
MIT — see LICENSE.