Six weeks after launch New architecture v9 Arm also introduced the first processor core built on them. After this, four new generation graphics chips will be added to it. In both cases, performance and efficiency increase. We will see them in the first products next year.
Cortex-X2
The most powerful processor core of the new arm family will be the Cortex-X2. his X1. Predecessor of It plays the top core in the Snapdragon 888 and Exynos 2100. In novelty, Arm believes in rapidly adopting both top mobile phones and laptops. X2 promises better branching prediction, has a 30% larger reorder buffer and smaller pipeline by one cycle.
According to Arm, its integer power must increase Up 16% compared to x1, Will be demonstrated in machine learning Double. Arv9 also brings with it vector units that help with AI and scientific calculations, a width of 128 bits is used in the case of the X2 SVE2. SVE means extra silicon, but Arm saved it by removing support for 32-bit instructions from the kernel, so it’s a fully 64-bit processor.
New mobiles with Cortex-X2 cores will offer up to 30% more than existing mobiles (Source: Arm)
Cortex-X2 will be used most often in mobile phones In single-core variants, But high-end chips may also offer four cores in combination with the new Cortexy-A710, which will be discussed in the next chapter. After that they will be able to target the laptop as well. Eight core processor Consisting of only one cluster of x2 cores. X2’s single-thread power arm. Is according to Intel i5-1135G7. Up to 40% more than.
And how does IPC compare to other Cortexes? Let’s take the two year old A77 model as the basis, which we still find in many mobile phones today. Cortex-A78 increased performance by 7%. The Cortex-X1 is 20% more powerful than the A78. And the Cortex-X2 mentioned 16% before the X1.
Cortex-a710
The A-core nucleus turns into a three-digit marking. The new Cortex-A710 is the successor to last year’s A7 and will play the role of powerful and effective hard workers. Here too many changes have taken place at the level of microarchitecture. Arm shortened the pipeline and reduced the total to ten per cycle, reducing the number of instructions per cycle from 6 to 5, Arm improved branching prediction, and increased the L1 TLB by half.
Thanks for the modifications, the Cortex-A710 should be A78. 10% more powerful than, Or at the same power 30% more economical. Here also the performance of AI has doubled. Unlike the X2, the A710 still supports 32-bit instructions, but according to Arm, from 2023 all cores will only work with AArch64. Therefore the potential A711 or A720 cores will already be combed.
Cortex-A710 (ZDrose: Arm)
However, the performance enhancement of all reported arm cores needs to be taken with some caution. The new DSU-110 platform, in which the company interconnects different groups of different processor cores, can now take up to 16 MB of shared L3 cache, compared to a maximum of 8 MB for its predecessors. Thus, a large cache can play an important role, and its availability and capacity in end products are not guaranteed. This is optional, but in arm marketing, full size relies on it.
The Cortex-A710 is marked as big in the arm hierarchy, the new Cortex-A510 will be the LITTLE. Four years later We have seen this since the launch of the A55 core New energy saving units, Which can be found in many cheap mobile phones and other electronics even without the support of strong A7x cores.
Due to the long break, Arm produced the most changes from the whole new family here. Instruction processing remains in order, but reprocesses two instead of three instructions per measurement. The data and instruction L1 cache can now contain at least 32 kB (instead of 16 kB), and optional L2 may contain 512 kB (instead of 256 kB). The kernel is 64-bit only and because it is also built on Armav 9, it specifically provides an SVE2 unit with 64- or 128-bit vectors.
Cortex-A510 (zdroj: arm)
However, a change may have a negative impact on performance, but positive for chip size and efficiency. Chipset manufacturers can use the so-called merge-core microarchitecture, where instead of two separate cores, a dual-core complex is used, in which both cores share the L2 TLB, L2 cache, and the NEON / SVE2 unit.
In one way or another, Arm believes the modifications will primarily increase performance. Cortex-A510 should be A55. 35% more powerful than Three times faster in integer operations and machine learning. Performance has already caught up with the five-year-old out-of-order Big Core Cortex-A73, up 10% over the IPC A510, with consumption down to 35%.
Mali-G610, G410, G510 and G310
Arm not only develops processor cores, but also graphics. Its Mali GPUs are used by Samsung, Huawei, MediaTek and others, making them the most used in the world. To the family Will be connected Three new graphics labeled Mali-G410 (successor to G4 in 2020), Mali-G410 (successor to G4 in 2020), Mali-G510 (successor to G52 in 2014), and Mali-G310 (successor to G31 in 2016). 2018).
All will run on the third version of the Wallhall architecture and will certainly vary in performance. It will be on top Mali-G610Which compared to its predecessor Increases gaming performance by 20% with equal consumption, Or at the same power consumption will decrease by 20%.
Mali-G710 (Source: Arm)
The architecture is scalable again, so the Mali-G710 can have 9 to 14 computing units. The predecessor had up to 24, but they were smaller and less complex. Arm also promises redesigned task management to be more suitable for modern APIs such as Vulcan.
However, the Mali-G710 supports Vulcan 1.2, OpenGL ES 3.2 and OpenCL 2.0 Still without directx, So we will not meet him in the world of Windows. In addition, Arm notably increased gross performance, but did not add new technologies, as they are now expanding on PCs and consoles, mainly realtime ray tracing, variable rate shading, etc.