Tensor is the company's first smartphone chip, and in the practical use of the Pixel 6, it impresses even when competing with Qualcomm's flagship chip. However, there are a lot of details about the Google tensor that the company hasn't delved into, but new in-depth research shows the Exynos roots of the product, gives better data compared to other chips, and so on.
The folks at AnandTech published an in-depth study of the Google tensor on the Pixel 6/Pro this week, providing plenty of technical details about the chipset and what it brings.
The first interesting detail in this report is to see the Exynos root of the Google tensor. As others have discovered before, Tensor's chip ID is almost identical to Samsung's format. The ID of Tensor is S5P9845, while the ID of Samsung's latest self-developed chip Exynos 2100 is S5E9840. Our Dylan Roussel also observed several direct references to "Exynos" in the files on The Pixel 6, which helps with a little backup of these Exynos roots.
On the silicon chip side, the chip has another model, SoC's fusion chip logo following Samsung's Exynos naming scheme. Here we find that the chip has an ID "0x09845000", which corresponds to S5E9845(edit: it is S5P9845). Samsung Electronics' latest LSI SoC model is Exynos 2100 and labeled S5E9840.
The report also explained that Tensor uses "the same clock management and power management architecture" that Samsung Electronics uses on Exynos CPUs, while other chip components such as memory controllers have more in common. Previously, we also found evidence that appeared to support Samsung making chips and led us to believe that Tensor, then known as Whitechapel, would share some software components with Exynos Design.
The Samsung-supplied modem is particularly notable because it makes the Pixel 6 Pro one of the first U.S. phones to offer mmWave connectivity on a non-Qualcomm modem. Interestingly, the modem is not directly integrated into the tensor's chip, as it is on the similar Exynos 2100.
However, there are also many differences between tensors and typical Exynos chips. Google uses two media encoder IP blocks, one provided by Samsung and one provided by Google, primarily for processing AV1. Also, Google uses a different processor core setup than "equal" Exynos, with tensors having a 2+2+4 arrangement. At this point, Google's use of aging and much less efficient A76 core is described as "pointless" because of the huge difference in performance efficiency.
One of the biggest differences between tensors and Exynos comes from "edge PU" developed by Google. Interestingly, the TPU version of the tensor seems to get 5W power, a considerable jump from the 2W power of the namesake chip Google announced in 2018.
TPU also had different results in benchmarking compared to solutions on the Exynos and Snapdragon chips. Tensor outperforms Exynos in image classification/segmentation and target detection tasks but lags Snapdragon, but beats rival chips in language processing, as well as Apple's A15.
The Google Tensor is not Google's chip, at least because of Samsung's Exynos design. At the same time, Google has made a lot of changes that are truly their own, and the description of "semi-custom" is a good description of the product.
The report next delves into memory, saying that "tensor DRAM latency is not very good" compared to the Exynos 2100 and Snapdragon 888, and that overall, Google's chip handles some things differently than those of more standard chipsets. This caused the tensor to lag in tests involving memory delays. The main effect is likely to be lower power consumption, as tensors end up "wasting" clock cycles on the kernel while waiting for memory, thus losing some power.
Compared to Snapdragon 888, the Google tensor is approximately 12.2% less energy overall, while doing the same task at a slower pace consumes approximately 14% more energy. According to AnandTech's reasoning, the size of the gap seems to have something to do with how Google handles memory.
Meanwhile, when it comes to GPUs, AnandTech describes them as "beasts".The Mali G78 GPU has 42% more cores than the same GPU on the Exynos 2100 and runs at frequencies up to 1Ghz. However, these powerful specifications do not translate directly into quite powerful results. The Pixel 6 Pro performed about 21 percent better on the benchmark app 3DMark Wild Life, actually beating the Exynos 2100 and Snapdragon 888 in peak performance and only trailing Apple's iPhone.However, the throttling keeps the phone much lower, behind both chips and only slightly better than the Snapdragon 865+. Peak power output exceeds 9W, but the duration frequency is low, around 3W. We noticed in our review that the game was heating up relatively quickly. The Aztec High benchmark showed similar results, while the Basemark GPU kept the Pixel 6 from overtaking the Exynos 2100 for unknown reasons.
