Pixel 6 Performance with Google Tensor Processor | Gadgets News

When Google announced at the October presentation of the Pixel 6 smartphone that compared to its predecessor, the performance of the CPU and GPU of the new product increased by 80% and 370% (4.7x), respectively, it sounded very encouraging. Among other things, this meant that the new Google Tensor processor provides the highest graphics performance at least among Android smartphones. To what extent these hopes were justified, the review published today from AnandTech.

According to our colleagues, the Google Pixel is equipped with a CPU and GPU with the following configuration (next to top-end Samsung and Qualcomm chips for comparison):

Google Tensor Samsung Exynos 2100 Qualcomm Snapdragon 888
Technical process Samsung 5 нм (5LPE) Samsung 5 нм (5LPE) Samsung 5 нм (5LPE)
CPU 2x Cortex-X1 (2.8 ГГц)
2x Cortex-A76 (2.25 ГГц)
4x Cortex-A55 (1.8 ГГц)
1x Cortex-X1 (2.9 ГГц)
3x Cortex-A78 (2.8 ГГц)
4x Cortex-A55 (2.2 ГГц)
1x Cortex-X1 (2.84 ГГц)
3x Cortex-A78 (2.4 ГГц)
4x Cortex-A55 (1.8 ГГц)
GPU Mali G78 MP20 (848 МГц) Mali G78 MP14 (854 MHz) Adreno 660 (840 MHz)
Modem Samsung Exynos 5123
(external)
Samsung Exynos 5123
(built-in)
Qualcomm Snapdragon X60
(built-in)

And here’s what the test results look like in various benchmarks – let’s start with the CPU:

Pixel 6
Pixel 6 Pro
(Google Tensor)
Pixel 5
(Qualcomm Snapdragon 765G)
Qualcomm
Snapdragon 888
Samsung
Exynos 2100
iPhone 13 / iPhone 13 Pro
(Apple A15)
SPEC CPU 2017 (integer) Cortex-X1 — 3.93 (3.15 Вт)
Cortex-A76 — 2.17 (1.19 Вт)
Cortex-A55 — 0.78 (0.62 Вт)
Cortex-X1 — 4.48 (3.18 Вт)
Cortex-A78 — 3.17 (1.66 Вт)
Cortex-A55 — 0.70 (0.32 Вт)
Cortex-X1 — 4.15 (3.44 Вт)
Cortex-A78 — 3.68 (2.96 Вт)
Cortex-A55 — 0.74 (0.53 Вт)
Avalanche – 7.28 (4.11 watts)
Blizzard — 2.42 (0.44 Вт)
SPEC CPU 2017 (floating point) Cortex-X1 — 6.17 (3.51 Вт)
Cortex-A76 — 3.45 (1.48 Вт)
Cortex-A55 — 0.88 (0.59 Вт)
Cortex-X1 — 6.28 (3.48 Вт)
Cortex-A78 — 4.58 (1.80 Вт)
Cortex-A55 — 0.85 (0.30 Вт)
Cortex-X1 — 6.23 (3.97 Вт)
Cortex-A78 — 5.25 (3.24 Вт)
Cortex-A55 — 0.94 (0.57 Вт)
Avalanche – 10.15 (4.77 watts)
Blizzard — 2.42 (0.44 Вт)
GeekBench 5 1,041
2,910
574
1523
1,071
3,559
1,106
3,584
1,750
5,022

If we measure the CPU performance according to the popular GeekBench 5 benchmark, then we see a 91% increase in speed compared to the Pixel 5. Which even slightly exceeds the 80% promised by Google – although it does not quite reach the top-end Snapdragon 888, not to mention the Apple A15.

And here is the performance of the GPU of the new chip:

Pixel 6 / Pixel 6 Pro
(Google Tensor)
Pixel 5
(Qualcomm Snapdragon 765G)
Qualcomm Snapdragon 888 Samsung Exynos 2100 iPhone 13 / iPhone 13 Pro
(Apple A15)
3DMark Wild Life Unlimited 42.43
17.10
35.19
23.50
34.83
19.60
60.90
42.88
Basemark GPU 1.2 — Medium 1440p (off-screen) 57.25
35.08
21.63
21.63
69.57
64.91
60.72
35.94
125.46
83.21
GFXBench Aztec Ruins — High 1440p (off-screen) 31.48 (7.28 W)
15.46 (2.78 W)
8.35
8.35
31.62
28.21
28.04 (7.69 W)
18.55 (3.73 W)
44.84 (5.92 watts)
31.83 (3.58 W)

This table summarizes peak and sustained performance data. The Pixel 6’s peak speed in two benchmarks is 2.65 to 3.77 times the Pixel 5. This is less than the promised 370% (4.7x), but about the level of top-end Android smartphones. Sustainable performance is much worse – the Pixel 6 has half the peak, so the difference with the Pixel 6 is only 62-85%.

Of particular interest is the performance of the NPU (TPU) – Pixel 6 and Pixel 6 Pro are positioned as smartphones with rich functionality from the field of machine learning (for example, removing unnecessary photographs in the frame and translating spoken speech from 48 languages):

Pixel 6 Pro Galaxy S21 Ultra (Snapdragon 888) Galaxy S21 Ultra (Exynos 2100) iPhone 13 Pro
MLPerf 1.0.1 (image classification) 942.51 1480.22 804.13 612.84
MLPerf 1.0.1 (object definition) 461.02 755.82 385.48 327.30
MLPerf 1.0.1 (image segmentation) 216.14 270.16 184.32 31.44
MLPerf 1.0.1 (Offline Image Classification) 1899.87 2,343.69 2129.40 1405.00
MLPerf 1.0.1 (language processing) 69.81 23.20 14.08 14.01
GeekBench ML 0.5.0 (НПУ/ГПУ) 1,739
1,385
1,200
1,668
718
1,551
2,710
2,246
AI Benchmark 4 – NNAPI (CPU + GPU + NPU) 264.0 197.6 176.6

As you can see, in the tests of the benchmark MLPerf 1.0.1, Snapdragon 888 dominates (except for language processing), in the other two benchmarks – Google Tensor (among Android smartphones).

We would like to thank the author of this write-up for this incredible content

Pixel 6 Performance with Google Tensor Processor | Gadgets News