HiSilicon has launched its new high-end mobile processor (or SoC) known as the Kirin 990 at IFA in Berlin.
Kirin 990 is the new chip that is expected to power both the high-end and select premium devices coming from Huawei / Honor for the next 12 months. HiSilicon is a Huawei subsidiary.
Kirin 990 and Kirin 990 5G bring a broad set of improvements, but the most visible to the consumers are going to be the CPU, Graphics and Storage performance and onboard 5G. 5G integration can have a lasting impact on phones’ cost, size and battery life. Variations in price also lead to variations in adoption rate.
Integrated 5G modem (Kirin 990 5G)
When 5G launched earlier this year, phone designs used two modem units: the main processor (SoC) with an integrated 4G LTE modem, and an external 5G modem. That is a typical first step when a new modem technology comes along, but does have downsides that include requiring more internal space and using more power because you need both 4G and 5G units active to hand over the signal if needed.
Integrating the 5G modem in the SoC avoids the above pitfalls, and that’s what HiSilicon has done with the Kirin 990 chip. The 5G unit can now share power and resources with the rest of the chip. Huawei claims that its solution requires ~36% less board space than competing ones from Samsung (Exynos 9825+5100 modem) or Qualcomm (Snapdragon 855+X50 modem).
However, Kirin’s 5G implementation isn’t complete: the mmWave (millimeter wave) technology, one of 5G’s two main radio protocol, is not supported. Competitors like Qualcomm do support it in their products, although not all phone makers have activated mmWave, perhaps because their region doesn’t support it. Depending on your carrier, this may or may not be an issue.
Some carriers rely heavily on mmWave because it lets them reach incredible speeds when the connectivity is excellent. Many carriers use Sub-6Ghz (“Sub-6”) which is a more versatile radio-frequency, but one that isn’t capable of 5G’s fastest performance.
Finally, Sub-6’s coverage is likely to be better than mmWave for the next year or two, because it takes a long time to roll-out a strong mmWave network, so statistically, you are more likely to be on Sub-6 or even 4G LTE, which remains the backdrop for 5G.
Better Image Processor For Better Camera Performance
All mobile chips include one or two Image Signal Processor (ISP) to support the computations required for incoming images from one or multiple cameras. ISP do the heavy lifting when it comes to image processing because they are the only units capable of running some image algorithms at high-speed on tens of millions of pixels.
This year, HiSilicon is promoting a better ISP-based noise-reduction approach called BM3D, which stands for block-matching and 3-D filtering, which is probably the most potent denoising algorithm that can be run in realtime, thanks to specialized units such as ISPs. There are even better techniques, such as WNNM, but they can’t run in real-time for now.
By the nature of the BM3D algorithm, de-noising can be applied very efficiently on single-frame photos or multi-frame photos or videos.
BM3D works by finding, multiple blocks that have similar patterns, either within the same photo, or across multiple frames. Because noise is random by nature, BM3D can find similarities between the blocks to treat certain types of differences (above a certain threshold) as “noise” and remove it (image courtesy of Jongmin Beak, Stanford).
Graphics Performance Back On Top
Kirin 990 features a similar graphics processor (GPU) design to last year’s Kirin 980, with a crucial difference: it contains 60% more computing units, going from 10 to 16.
That’s how GPU can scale performance almost linearly. Because graphics are “computationally unbounded,” throwing more compute units to the problem does accelerate it immediately.
In addition to the extra units, HiSilicon is also adding a share memory system called “Smart cache.” It is a small on-chip ultrafast memory that different sub-units can use to share data, instead of going back to the RAM.
It is intended to reduce RAM access slowdown in some situations, and Huawei says that this boosts graphics power-efficiency by ~20%, inducing a power-savings of ~12%.
This puts Huawei firmly back in the gaming performance race. As we’ve shown in our various smartphone reviews, Kirin 980 was lagging Snapdragon 855’s graphics performance by 14%-27% depending on the test. Huawei says that it now has a lead of 6% over Snapdragon 855, which is in-line with the extra hardware Kirin 990 is getting.
CPU & Storage Performance
To address CPU performance, HiSilicon is increasing the maximum operating frequency of the CPUs of all three CPU islands. The 8 CPU cores of both the Kirin 980 and Kirin 980 are separated in three islands, with different clock speeds (operating frequencies):
- Highest performance: 2x Cortex A76 at 2.86GHz (from 2.60GHz) +10%
- High performance: 2x Cortex A76 at 2.36GHz (from 1.92GHz) +23%
- Power-Efficiency: 4x Cortex A55 at 1.95GHz (from 1.8GHz) +8.3%
The bump is very decent and could help for peak-intensity tasks, but overall keep in mind that increasing frequency isn’t as optimal as improving architecture efficiency because running higher frequencies isn’t as power-efficient, so a performance-per-watt metric might decrease.
Nonetheless, HiSilicon uses the dynamic frequency capabilities of the CPU clusters to seek a sweet spot for every task.
When it comes to storage, Kirin 990 supports UFS 3.0 a flash storage protocol to access the internal storage of the phone. Already adopted by Qualcomm’s Snapdragon 855 and Samsung’s Exynos 9820, UFS 3.0 more than doubles the speed of the UFS 2.0 protocol supported in the current Kirin 980 chip.
That, coupled with the proprietary work that Huawei has done on filesystems, should lead to yet another round of low-latency app launches and general user experience (UX) performance. I guess we’ll know soon enough with the Mate 30 Pro launch, which is imminent.
A.I and Machine Learning Performance
Kirin is the chip that started the NPU war (Neural network Processing Unit) in the mobile space by aggressively marketing a variant of what was previously commonly called DSP (Digital Signal Processor) before. There are differences, but that was the move that put smartphone AI hardware in the news.
From the experience of the past few years, HiSilicon has developed a proprietary NPU design, called DaVinci and can build NPUs of various sizes depending on how many compute units are required.
Kirin 990 has two NPU based on that design, a small low-power and a large high-power one. As one can imagine, the low-power is designed to do things like Face Unlock with as little power as possible, while the other one is tasked to churning AI for games and other high-intensity applications as fast as possible.
In synthetic benchmarks, Huawei says that it can get a 1.88X speed improvements, so the race is still firmly on.
Kirin 990 5G is a beast and raises the prospect of faster adoption of 5G in many parts of the world. 2020 is without a doubt the real year of 5G with that technology fully entering the high-end, premium, and perhaps the high-mid-range markets.
This new processor raises the bar across the board: user experience, application loading, execution speed, better imaging, better gaming: every single Kirin 990 user will enjoy benefits for his or her use case.
Now we’re looking forward to seeing the first wave of handset powered by Kirin 990, starting with the Huawei Mate 30 series.