Samsung announces 8nm Exynos 9820 with CPU Tri-Group design

Morning Morning Morning Morning, Samsung LSI announced the new generation of SoC, the Exynos 9820.

The Exynos 9820 succeeds this year in the Exynos 9810, which we covered extensively both in the Galaxy S9 review and in separate tracks ending with an official microarchitectural disclosure of the core M3 on the HotChip 2018.

The Exynos 9810 was not a very successful chip, as the M3 microarchitecture and the application had some serious flaws, with real chip-based devices that were largely delayed by the opposite parts of Snapdragon, had to sacrifice many performance, keep up with battery performance, and even then, do not do this most of the time.

The Exynos 9820 wants to fix this situation with new design and new CPU configuration, but first to overcome the specification:

Technical Specifications Samsung Exynos SoCs
SoC	Exynos 9820	Exynos 9810
Processor	2x M4; @? GHz It has L2 2x Cortex A75 @; GHz Probably L2 4x Cortex A55 @; GHz Let's hope he has L2 Shared Composite L3 @ => 4MB	4x M3 @ 1c2.7 / 2c2.3 / 4c1.8 GHz 4x 512KB L2 4096KB L3 4x Cortex A55 @ 1.8 GHz not L2 512KB L3
GPU	Mali G76MP12	Mali G72MP18
Memory Controller	4 x 16-bit CH LPDDR4x @MHz	4 x 16-bit CH LPDDR4x @ 1794MHz
Media	8K30 & 4K150 encoding & decoding H.265 / HENO, H.264, VP9	10bit 4K120 encoding & decoding H.265 / HENO, H.264, VP9
Modem	Shannon Built-in LTE (Category 20/13) DL = 2000 Mbps 8x20MHz CA, 256-QAM UL = 316 Mbps 3x20MHz CA, 256-QAM	Shannon Built-in LTE (Category 18/13) DL = 1200 Mbps 6x20MHz CA, 256-QAM UL = 200 Mbps 2x20MHz CA, 256-QAM
ISP	Back: 22MP Front: 22MP Dual: 16MP + 16MP	Back: 24MP Front: 24MP Dual: 16MP + 16MP
Mfc. Process, process	Samsung 8nm LPP	Samsung 10nm LPP

The new 9820 radically changes the CPU configuration from a 4 + 4 installation to a 2 + 2 + 4 installation where we now have two fourth-generation Samsung processors as top units, two Cortex A75 CPUs as mid-tier performance cores and four processors Cortex A55 for maximum power output.

First of all, the new CPU configuration corrects one of the Exynos 9810 issues: Here, the A55 cores were in a separate cluster block in the SoC, and so the cache consistency in the performance cores had to go through the SCI / interconnect, which undoubtedly affected efficiency. In the Exynos 9820, all processors, including Cortex processors, are now integrated into the same cluster and the new L3 is a cache hierarchy across all CPUs, undoubtedly it not only improves the power efficiency but also the performance of the smaller cores now.

The new L3 is very similar to the DynamIQ DSU of Arm, but it is designed by Samsung, as Arm does not allow third-party cores to connect to the DSU.

Choosing a Cortex A75 as a middle class CPU is obvious: Having four large cores, like the M3 or obviously M4, is a lot of space waste because in a mobile power folder, the four cores will never be able to operate at the maximum frequency / performance. So instead of using the same large microarchitecture in a lower timed configuration, it makes sense to use only a smaller CPU and retrieve some of the uterine space. This also allows Samsung to further optimize the large customized kernels for peak frequencies and performance, as they no longer have to estimate the threshold of the power curve – which is now covered by the A75.

From the point of view of the new fourth-generation custom CPU, M4 (if it does so) is advertised to have "improved memory accessibility." As we covered M3's microarchitectural revelation, Samsung recognized that the cache hierarchy was one of the last pieces designed for the IP and it seemed obvious that there were some compromises. The memory hierarchy and memory subsystem of M3 appear to be the Achilles heel of microarchitecture, as it seems to prevent the kernel from reaching the higher levels of performance that one would expect from such a large nucleus.

The new M4, according to marketing materials, is now promising a 20% or a 40% increase in energy savings compared to the Exynos 9810. I'm still working on the revision of Mate 20 and the new Cortex-A76 athletic Kirin 980, which should be published shortly, but these Samsung improvement requirements make do not seem sufficient to be able to remain competitive and reach the Cortex A76. Here, Samsung will have to perform a 20% improvement in performance, while improving energy efficiency by 50% to cover the Kirin 980. Unfortunately, the current gap seems too long for 9820 to be overcome unless of course it hits the marketing promotions.

One thing that S.LSI did not announce today was any frequencies. For the Exynos 9810 they had chosen to advertise some unrealistically high watches "up to 2.9GHz", but these were not feasible without some absurd power. Along with IPC upgrades of the new kernel, I expect the M4 to remain in the 2.5-2.7GHz range, with Samsung avoiding going too high at the end of the trend curve. The Cortex A75 should be in the lower 2GHz range to take advantage of the best performance points.

Larger GPU – Mali G76MP12 Arm

On the GPU side of things, the Exynos 9820 now offers the new Mali G76 Arm in MP12 format. This represents a 33% increase in performance machines compared to the latest generation of G72MP18, and Samsung's materials advertise a 40% increase in performance or an increase in efficiency by 35% in the same performance. In the Kirin 980 the new G76 made some great improvements, but it lost the mark in terms of my projections. However, S.LSI has always been able to achieve much better results than HiSilicon, so the 9820's new G76 could end up being respectable, although I still do not expect it to be better suited to Apple's best and the next generation of Qualcomm.

A new dedicated NPU, along with connectivity and media improvements

The new 9820 also features a new first for S.LSI: The new chip now has NPU (Neural Processing Module) for the first time. The details of the new IP are missing and all we know is that Samsung promises to be used to enhance camera shots or use cases like AR and be "7 times faster" than Exynos 9810 in its tasks.

In terms of media capabilities, the new SoC IP encoder now promises 8K30 capability. Now I really do not expect new devices to be able to record in this way as it will also require camera sensor support – but having a powerful media encoder like this can also mark the improved write power performance.

The new modem now has LTE connectivity up to category 20 in the download and upload, achieving up to 2Gbps at 8CA downstream and 316Mbps at 3CA upstream.

One thing Samsung did not emphasize in its announcement but cited in the specification is the fact that SoC now supports the storage of UFS 3.0. The new standard will be able to double the transfer speeds compared to UFS 2.1, achieving a theoretical maximum of 2.9 GB / s.

8nm LPP Manufacturing node

Finally, but equally important, the new Exynos 9820 is built on a Samsung LPP 8nm process node. 8LPP is an evolution of 10LPP and promises an improved power consumption of up to 10%. Here the improvements that are being advertised are relatively poor and some people will wonder if this will be enough to compete with the 7nm TSMC node on today's Kirin 980 as well as the upcoming Snapdragon. Although it looks like Samsung will have a density disadvantage, I do not expect the performance and the power disadvantage in everything that is dramatic. However, the Exynos 9820 will have a tough time this year, both on the microarchitecture and on the process node side.

Overall, the new Exynos 9820 should bring the required improvements compared to its predecessor, but competition this year will be extremely difficult. Here I hope Samsung makes the best of the silicon and is also able to have a good stack of software, as the new configuration of a three-tier CPU will need a good scheduler. The new Exynos 9820 is planned to reach mass production by the end of the year, and we will see it this spring in the new Galaxy S10.

Relevant reading