Samsung announces the Exynos 9820 of 8 nm with the design of a tri-group processor

Early this morning, at the time of Korea, Samsung LSI announced the launch of its flagship SoC product of the latest generation, the Exynos 9820.

The Exynos 9820 is the successor to this year's Exynos 9810, which was thoroughly analyzed in our review of the Galaxy S9, and was also investigated separately, resulting in disclosure. official microarchitectural of the M3 core at HotChip 2018.

The Exynos 9810 was not very successful because the M3 microarchitecture and implementation had serious flaws, with chip-based devices lagging far behind their Snapdragon counterparts, who had to sacrifice a lot of performance to stay on top of the battery efficiency, and even then, do not do it most of the time.

The Exynos 9820 seeks to remedy this situation with a new CPU design and configuration, but first look at the specifications:

The new 9820 now fundamentally changes the processor configuration from a 4 + 4 configuration to a 2 + 2 + 4 configuration, where we now have two fourth-generation Samsung processors as the most powerful dual-processor units. Cortex A75 as intermediate level performance. and four Cortex A55 processors for maximum energy efficiency.

First of all, the new configuration of the CPU corrects one of the problems of the Exynos 9810: here, the A55 cores were on a separate cluster block on the SoC, and the consistency of the memory cache performance cores had to go through the SCI / interconnection. , which has undoubtedly had an impact on efficiency. On the Exynos 9820, all processors, including Cortex processors, are now integrated in the same cluster and the new L3 is a cache hierarchy for all processors, which not only improves energy efficiency , but also the performance of smaller hearts.

The new L3 cluster is very similar to Arm's DynamIQ DSU, but it was designed in the same way as Samsung, since Arm does not allow third-party kernels to connect to the DSU.

The choice of a Cortex A75 as a group of middle-level processors is obvious: having four large cores, such as the M3 or the M4, is a considerable waste of space, because in a mobile power envelope, the four hearts can never be replaced. operate at their frequency / maximum performance. So, instead of using the same large microarchitecture in a less rhythmic configuration, it makes sense to simply use a smaller processor and recover some of the available space. This also allows Samsung to further optimize the custom cores for peak frequencies and performance, as they no longer have to evaluate as much as the bottom of the power curve, which is now covered by the A75.

Regarding the new generation of fourth-generation custom processors, the M4 (if they end up naming it that way) is advertised as having "improved memory access capability". As we had covered the microarchitectural disclosure of M3, Samsung admitted that the hierarchy of caches was one of the last elements designed for intellectual property and it seemed obvious that some compromises had been made. The memory hierarchy and the memory subsystem of the M3 seemed to be the Achilles heel of the microarchitecture, as they seemed to prevent the kernel from achieving higher performance levels than you expected from such a large kernel .

The new M4, according to marketing materials, now promises a 20% increase in performance, or an energy efficiency increase of 40% compared to the Exynos 9810. I am still working on the revision of the Mate 20 and the new Cortin-A76 sportsman Kirin 980, which should be released soon enough, but these improvements made by Samsung do not seems not enough to stay competitive and catch up with the A76 Cortex. In this case, Samsung should show a performance improvement of 20% while improving its energy efficiency by 50% to catch the Kirin 980. Unfortunately, the gap currently seems too important for the 9820 to overcome, unless of course to beat the advertised marketing numbers.

One thing S.LSI did not announce today is a kind of frequency. For the Exynos 9810, they had chosen to announce irresistible clocks of "up to 2.9 GHz", but these were not achievable without absurd power figures. Along with the new kernel IPC upgrades, I expect the M4 to remain in the range of 2.5-2.7 GHz, Samsung abstaining if possible to climb too high in the game extreme of the voltage curve. We hope that the Cortex A75 will be in the lower range of 2 GHz to take advantage of the best points of efficiency.

A larger GPU configuration – Mali G76MP12 of Arms

On the GPU side, the Exynos 9820 now offers the new Armorer Mali G76 in an MP12 configuration. This represents a 33 percent increase in the number of runtime engines compared to the next-generation G72MP18, and Samsung materials report a 40 percent increase in performance at equal power, or a 35 percent increase in performance at equal performance. On the Kirin 980, the new G76 has made some very important improvements, but it just missed the goal in terms of my projections. However, S.LSI has always managed to achieve much better results than HiSilicon. The new G76 of the 9820 could well be respectable, even if I do not expect it to match again the best Apple and the next generation of Qualcomm.

New dedicated nuclear power plant, coupled with connectivity and media improvements

The new 9820 also offers a new first for S.LSI: the new chip integrates for the first time a neural processing unit (NPU). The details of the new IP address are sorely lacking, and all we know is that Samsung promises that it will be used to enhance photo captures or use cases such as RA , and that it will be "7 times faster" than the Exynos 9810 in its tasks.

In terms of multimedia capabilities, the new SoC IP Encoder now promises 8K30 capacity. Now I'm not really expecting new cameras to capture in this mode because it will also need support for the camera sensor – but have such a media encoder can also indicate increased efficiency of the recording power.

The new modem now offers LTE connectivity up to category 20 for download and download, reaching 2Gbps at 8CA downstream and 316Mbps at 3CA upstream.

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

8 nm LPP manufacturing node

Finally, the new Exynos 9820 is manufactured on an 8 nm LPP process node from Samsung. 8LPP is an evolution of 10LPP and promises an improved energy consumption up to 10%. Here, the announced improvements are relatively meager, and some people will wonder if it will be enough to compete with the TSMC 7 nm node in the current Kirin 980 as well as in the next Snapdragon. Although it seems that Samsung will have a disadvantage in terms of density, I do not expect such a disadvantage in terms of performance and power. Nevertheless, the Exynos 9820 will go through a difficult period this year, both on the microarchitectural side and the process.

Overall, the new Exynos 9820 is expected to provide much needed improvements over its predecessor. However, the competition will be extremely tough this year. I'm hoping here that Samsung makes the most of silicon and may also have a good software stack because the new tri-group processor configuration will require a good planner. The new Exynos 9820 is expected to enter mass production by the end of the year. We will see it next spring, of course, in the new Galaxy S10.

Related reading