docs/performance.md

Performance

On this page

How close Linux gets to the hardware, measured on one tablet in October 2026, mostly against Android on the same tablet. Single runs: expect a few percent between runs. Linux: Debian 13 with GNOME, Open Device Helper defaults.

Summary#

AreaAndroidLinuxNotes
CPU, Geekbench 7 single / multi3142 / 105203096 / 10121 (98 % / 96 %)with memlat (patches 0120-0127); kernel-t40 without it: 3038 / 9951
Memory bandwidth (copy)25.7 GB/s25.6 GB/ssame
Memory latency, 64 MB random68 ns69 nskernel-t40 without memlat: 127 ns
GPU compute, Geekbench 7 OpenCL1920420167 (105 %)the project's Mesa build; Debian's Mesa 26.1.6: 7559 (39 %), see GPU
GPU raw throughput (clpeak)—FP32 3.05 TFLOPS, 66 GB/sDebian's Mesa 26.1.6; about 83 % of the theoretical 3.7 TFLOPS
Storage, sequential read / write2.7 / 1.8 GB/s4.4 / 1.9 GB/sAndroid's /data is encrypted (inline crypto)
Storage, 4K random read QD1 / 4 jobs13.6k / 459k IOPS13.1k / 401k IOPSwith MCQ (patch 0119); see Storage for how they were measured
Video decode, 4K—H.264 319 fps, HEVC 531 fpshardware decoder (iris); 40 Mbit/s test clips
NPU, ResNet50 INT8 / Llama 3.2 3B Q4_0—1.5 ms / 1732 tok/s prompt, 28 tok/s generationHexagon NPU with Qualcomm's runtimes; see NPU

Against published Android results#

Where the same benchmark runs on both, Linux on this tablet against results published for other Snapdragon 8 Elite Gen 5 devices (Android):

BenchmarkLinuxPublished (Android)
Geekbench 6 CPU single / multi3641 / 11175about 3655–3710 / 10758–11672
Speedometer 3.1 (Chromium 150)22.518.1 (OnePlus 15), 23 (Xiaomi 17), 46 (Galaxy S26 Ultra)
JetStream 2.2 (Chromium 150)389.5about 281–295 (Chrome)
Geekbench 6 Vulkan26977 (the project's Mesa build), 18099 (Debian's Mesa 26.1.6; an earlier run on kernel-t40: 18342)about 27200–29700
Geekbench 6 OpenCL23034 (the project's Mesa build), 9657 (Mesa main)about 24000
GravityMark 1.89 Vulkan, 1920×1080, 200k asteroids5482 (32.8 fps, Mesa 26.1.6)no Adreno 840 entries; Adreno 830: 3530–4568

With Debian's Mesa 26.1.6, Geekbench 6 OpenCL fails one workload's result check (Particle Physics) and scores 2593. Graphics (GravityMark) lands where an Adreno 840 should be, above the Adreno 830 phones. With the distribution's Mesa the loss is in GPU compute (Geekbench's OpenCL and Vulkan compute workloads), not across the board; the project's Mesa build closes it (see GPU).

CPU and memory#

Two kernel changes closed most of the CPU gap:

  • Patch 0117 gives the scheduler the two core sizes. Before it, all eight cores looked the same and a busy single thread could stay on a small core (Geekbench 6 single-core 2288 → 3590).
  • Patches 0120-0127 (Qualcomm's memlat series, not in mainline as of Linux 7.3) start the memory-latency governor in the CPU control firmware. Without it, DDR stays at a low clock while the CPU waits on memory with little bandwidth in use, and random memory access takes twice as long as on Android.
KernelGB7 singleGB7 multi
kernel-t4030389951
kernel-t40 with GNOME stopped30169999
with memlat309610121
Android314210520

Earlier Geekbench 6 runs on Linux: 2288 / 10359 (before 0117), 3590 / 11016 (with 0117), 3572 / 10857 (kernel-t40). Android results published for this tablet are about 3655–3710 / 10758–11672.

Per workload (Geekbench 7, memlat kernel against Android on the same tablet), single-core is uniformly about 97 % of Android. Multi-core is level except three long all-core workloads: Asset Compression 83 %, Clang 89 %, Ray Tracer 91 %. During them all eight cores are busy while the clocks fall within a second or two to 2.0–2.9 GHz (small cores, 3.63 GHz maximum) and 2.7–3.1 GHz (large cores, 4.61 GHz) at about 100 °C; the operating system reports no throttling, so the limit comes from the firmware. Whether Android sustains higher clocks there was not measured.

GPU#

The GPU itself runs as it should: clocks up to 1200 MHz (Android's limit too), FP32 3.05 TFLOPS and 66 GB/s measured with clpeak (Debian's Mesa 26.1.6). Graphics (OpenGL) go through the distribution's Mesa (freedreno). For Vulkan and OpenCL compute the project provides its own Mesa build (Turnip and rusticl for the Adreno 840): installed next to the distribution's Mesa through Open Device Helper (Performance → Graphics Drivers, or tb323fu-ctl mesa update), or downloaded from the project's mesa-* releases. See Graphics drivers.

BenchmarkDebian's Mesa 26.1.6the project's Mesa (2026.10.06.3)Android
Geekbench 7 OpenCL75592016719204
Geekbench 6 OpenCL2593 (fails one result check)23034about 24000 (published, other devices)
Geekbench 6 Vulkan18342 (kernel-t40; a later run: 18099)26977about 27200–29700 (published, other devices)

Default GPU clock settings (performance profile), no clock pinning. Measured on one tablet (the author's): for orientation only, results on other units may differ. With Debian's Mesa, OpenCL was the weak spot (39 % of Android in Geekbench 7); the cause was in Mesa's OpenCL and compute paths, not in the kernel. The project's build is checked against OpenCL-CTS and the Vulkan CTS (dEQP-VK) groups that cover compute and synchronization, with the same results as upstream Mesa.

NPU#

The Hexagon NPU (HTP v81) in the compute DSP runs on the mainline kernel as it is (FastRPC; no project patch) with the open fastrpc library, in an unsigned protection domain. The runtimes are Qualcomm's and are not part of the project: the QNN SDK (QAIRT) for neural networks, and llama.cpp's Hexagon backend for language models. Measured on kernel-t47, on one tablet (the author's): for orientation only. Power was not measured, and there is no Android comparison.

Model (QNN, qnn-net-run)Per inference, with the RPCAccelerator time
MobileNetV2 FP161.35 ms1.06 ms
MobileNetV2 INT81.31 ms1.03 ms
ResNet50 FP162.19 ms1.85 ms
ResNet50 INT81.52 ms1.23 ms

Average of 208 back-to-back runs; backend settings vtcm_mb: 0 and graph optimization level 3 (INT8) or 2 (FP16).

Language model, Q4_0 (llama.cpp, Hexagon backend)Prompt (pp512)Generation (tg64)
Llama 3.2 3B Instruct1732 tok/s28.4 tok/s
Qwen3 4B Instruct 25071408 tok/s24.0 tok/s
Qwen3 8B824 tok/s13.6 tok/s

llama.cpp needs GGML_HEXAGON_DMA64=0: with 64-bit DMA it maps the weights with a fastrpc flag that the open library did not have when this was measured.

Storage#

UFS 4.1, fio with io_uring, an 8 GB file, direct I/O, 20 s per test (the commands are under Running it yourself). Linux: kernel-t42, the median of four runs. Android: measured earlier with the same engine and file size, but options that were not recorded; not measured again.

TestAndroidLinux
Sequential read, 1 MB, QD322686 MB/s4421 MB/s
Sequential write, 1 MB, QD321819 MB/s1906 MB/s (one run of four: 2481)
4K random read, QD1 / QD32 / 4 jobs × QD3213.6k / 208k / 459k13.1k / 231k / 401k
4K random write, QD1 / QD32 / 4 jobs × QD3225.9k / 50.8k / 48.2k28.4k / 38.6k / 48.9k

Patch 0119 turns on the controller's multi-queue mode (MCQ). With the same commands, kernel-t40 (without 0119 and the memlat patches 0120-0127) does 8.5k random reads at QD1 and 178k with 4 jobs. In kernel-t42 the MCQ state was lost over suspend: after each resume the controller went through an error reset, ran on without its per-queue interrupts (about 254k with 4 jobs) and now and then failed to resume. Patch 0128 (from kernel-t43) keeps that state, so the numbers above hold after suspend too. The I/O engine matters at low queue depth: libaio is slower than io_uring at QD1. An earlier version of this page gave 16.1k and 448k for Linux; those came from fio options that were not recorded, and the kernel they were measured on gives the numbers above with these commands.

Running it yourself#

./geekbench7 --cpu                                   # Geekbench 7 Linux AArch64 preview
sudo apt install ocl-icd-libopencl1 clinfo clpeak    # the OpenCL loader; the driver comes from the project's Mesa
tb323fu-mesa run ./geekbench7 --gpu OpenCL          # with the project's Mesa (or log in again after `tb323fu-ctl mesa on`)
tb323fu-mesa run clpeak
sudo apt install tinymembench fio
tinymembench                                         # copy bandwidth; latency in the 67108864 row
F=/var/tmp/fio.dat                                   # storage: 8 GB test file, written first
fio --name=prep --filename=$F --size=8G --rw=write --bs=1M --direct=1 --ioengine=io_uring --iodepth=32
fio --name=t --filename=$F --size=8G --direct=1 --ioengine=io_uring --runtime=20 --time_based \
    --group_reporting --rw=randread --bs=4k --iodepth=1 --numjobs=1
#   sequential: --rw=read|write --bs=1M --iodepth=32; random: --rw=randread|randwrite --bs=4k
#   with --iodepth=1, --iodepth=32, or --iodepth=32 --numjobs=4
rm $F

The Geekbench preview builds need an Internet connection and upload every result to the Geekbench Browser. Close other programs and keep the tablet on a charger with Bypass charging off.

View this page's Markdown on GitHub