Yesterday afternoon, Google presented, among other things, its new best-of-its-kind smartphones, the Pixel 7 and Pixel 7 Pro. One of the advantages of these devices, as in the case of the previous generation, is the presence of a system-on-a-chip or SoC, custom: Google Tensor G2Developed directly by Google in collaboration with Samsung. The SoC, which includes the CPU, GPU and various co-processors, is definitely atypical compared to other proposals on the market, to the point that it is difficult to fit into a specific segment of the sector.
Right from the start, Google made it clear The main focus of the Tensor initiative is artificial intelligence: This allows Google to offer highly advanced features in the field of image processing, text-to-voice translation and vice versa, language recognition in the Google Assistant – in short, all those features that have always distinguished Google’s hardware software.
As one might expect for such an atypical system, its evolutionary path from one generation to the next also proves unpredictable. So it is worth stopping to do The goal of the situation “Hot”, right after the announcement, although some basic details are still missing (in short: field tests) that allow us to get a completely complete and comprehensive picture.
One of the most specific aspects of the original Google Tensor was the CPU configuration, with two “main” cores surrounded by two middle cores from two older generations and four high-efficiency cores (versus the classic “1 + 3 + 4” configuration that is now an industry standard). Google has kept the two extreme groups unchangedWithout going after that to the new designs introduced by Arm in 2021 that we saw on the 2022 flagship chips like the Snapdragon 8 Gen 1, Dimensity 9000 and Exynos 2200, And instead upgrade the intermediate block, and moving from Cortex-A76 to Cortex-A78. These cores belong to the same generation as Cortex-X1, so let’s say at least some uniformity has been achieved at this level. Cortex-A78s also benefit from a slightly higher operating frequency: 2.35 GHz versus 2.25.
Everything else is the same. It remains to be seen how much difference the new cores will make – Obviously such a segment wouldn’t spark benchmarks when compared to the other top of the range, but it’s also clear that Google doesn’t care much. Initially, the Pixel 6s, which featured the first generation Tensor, suffered a bit of lag and lags of various kinds, but time has shown that it’s all about software that hasn’t been improved yet – which is quite understandable given the debut. A critical component such as the SoC.
Surprises are not lacking here either. This time, Google implemented the latest Arm architecture, which was introduced at the same time as the Cortex-X2, Cortex-710 and Cortex-510, which is therefore a smaller generation of CPU, but has Cores reduced dramatically. The Arm Mali-G78 MP20 from the Tensor G1 had 20 cores (designated by the final number after the abbreviation “MP”), while the Mali-G710 MP07 from the Tensor G2 had to settle for 7.
On paper, the new architecture is more refined and advanced, but the massive reduction in cores leaves some doubts about the actual overall performance. exactly the contrary: It seems very likely that they will be worseActually, not a little. Just do some quick math. We know that the Tensor G1 GPU hit about 1.9 TFLOPs overall; The Tensor G2 tests haven’t come out yet, but the GPU is already known to be in many of the SoCs currently on the market. Depending on the operating frequency, the chip returns approximately 73 to 164 GFLOPs per core; Just multiply this data by 7 to get, in the most optimistic scenario where the operating frequency is the maximum allowable, a result of 1.1 TFLOPs.
It’s obviously taken with extreme caution because it is Only in theory and not verified in the field. Google may have a few tricks up its sleeve (much higher frequencies for example?). Or … well, he might conclude that this power is more than enough in a mobile environment. In short, it will be necessary to run tests on a real smartphone to understand the implications of this choice.
As we have seen, in this period, the “flight” of most chipmakers from Samsung foundries to TSMC foundries was observed, due to the apparent superiority of the production processes of the latter compared to the Korean competitor. For Google, the problem is a bit thorny, because Samsung has also collaborated on the development of the same chip. In fact, in the months since the release of the Tensor G1, experts Anand Tech They did not hesitate to call it Exynos Custom. Having said that, it is safe to assume that further reduction of the production node will ensure greater efficiency from an energy and heat point of view – this is called a lot of reality, The battery will last longer and the temperature of the smartphone will decrease under pressure. In this context, and given the above considerations on the GPU, it would be interesting to evaluate how the chip would behave in conditions of prolonged workload.
Evaluation of other ingredients is more sensitive, as there are few criteria. We can report it According to Google, the new generation TPU is 60% stronger and 20% more efficient Compared to before, now your ISP (Image Signal Processor) is able to process HDR content with a color depth of 10 bits (1.07 billion colors) versus 8 bits (16.7 million colors) than before.