Microprocessor Report (MPR) Subscribe

Exynos 2100 Delivers Big GPU Gain

Galaxy S21 Is First mmWave Phone Without Qualcomm Modem

February 1, 2021

By Aakash Jani


Samsung’s new flagship smartphone, the Galaxy S21, sports the Exynos 2100, which delivers industry-leading CPU performance and an integrated 5G modem. The company manufactures the 2100 in its 5nm technology, reducing power. It reused from the preceding Exynos 990 the 5123 modem but added new Samsung RF chips for millimeter-wave (mmWave) 5G, creating the first mmWave phone using non-Qualcomm components.

Like the Galaxy S20, the S21 will also come in a version based on a Qualcomm chipset: the Snapdragon 888. The Exynos 2100 aims to match Snapdragon by boosting AI throughput to 26 trillion INT8 operations per second (TOPS) and raising its GPU throughput by 40% compared with its predecessor, although it still falls short in the latter category. It features an expanded image processor as well, supporting a single 200MP camera. The 2100-powered S21 competes against other premium Android phones based on the 888 and HiSilicon (Huawei) Kirin 9000.

The new Exynos is shipping in three Galaxy models, all of which feature mmWave: the S21, S21+, and S21 Ultra. Versions that sell in the US and China employ the 888 and Qualcomm mmWave chips, whereas the rest employ the 2100 with Samsung’s mmWave. Both versions sell for the same price. The entry-level S21 carries a $799 price tag.

X1 Displaces Mongoose CPUs

For its new flagship processor, the company eschewed its in-house CPU for stock Arm designs, as Cortex-X1 replaced its Mongoose M5 CPUs. The Mongoose project survived five iterations of attempting to match the performance of Apple’s custom Arm CPUs, but never did. When the X1 debuted, the company saw an opportunity to offload its CPU efforts. Unfortunately for the laid-off Austin design team, the program ended in 2019 with the M5.

As Samsung retires the project, more details about the M5 microarchitecture are emerging, allowing comparisons with the company’s new Exynos. By switching to the X1, the 2100 features a longer pipeline and narrower front-end unit. That CPU has five decode units, one less than the M5. It employs four 128-bit Neon units in the execution engine while the M5 employs only three, aiding in JavaScript and security processing (see MPR 6/1/20, “Cortex-X1: A Bigger Arm”).

Adopting the X1 makes the 2100 more area efficient. In our CPU-die analysis from last year, the M5 placed last in performance per square millimeter (see MPR 7/27/20, “CPU Size Leads Apple to Drop Intel”). We estimate the X1’s area at 2.1mm2, which is 30% smaller than the M5’s (after accounting for the transistor shrink). With its new CPU cluster, the 2100 thus achieves better performance in a smaller area, leaving more room for other heterogeneous-compute blocks.

As Figure 1 shows, the 2100 has one prime core operating at a blazing 2.9GHz, three “big” application cores running at a slightly lower 2.8GHz, and four “little” cores running at a comfortable 2.2GHz. Samsung claims a 19% single-core-performance boost over the 990, but initial Geekbench 5 results indicate an enormous 37% jump. The company probably underestimated the performance gains to add marketing buzz when phones outperform expectations.

Figure 1. Exynos 2100 platform. Samsung upgraded its AI accelerator by expanding the core count from two to three. Its new AI engine delivers an estimated 21 TOPS, hitting 26 TOPS when combined with the CPU and GPU.

Whereas the 2100 has a 1+3+4 CPU configuration, the previous 990 implements a 2+2+4 configuration of Samsung M5, Cortex-A76, and Cortex-A55 cores, respectively (see MPR 12/2/19, “Exynos 990 Enables 5G Smartphones”). Moving to a single prime core reduces die area while maintaining high single-thread performance. We estimate the 2100’s multicore throughput surpasses that of its predecessor by a sizable 39%, thanks to greater IPC from the Cortex-X1 and Cortex-A78 cores as well as an average 11% frequency boost.

Expanding AI for 26 TOPS

Samsung upgraded the 2100 with Arm’s newest GPU, the Mali-G78 (see MPR 6/1/20, “Mali-G78 Raises Premium Performance”). Combining the new architecture with an increase in the shader count from 11 to 14 units leads to a 40% generational graphics improvement. We estimate the new GPU consumes 25% more area than the previous one, an expansion mitigated by removing the oversize Mongoose cores. The earlier Exynos 990 lagged far behind competitors in GPU performance, so even this improvement is insufficient to catch up to Qualcomm.

The G78 also enables new capabilities, such as faster screen-refresh rates. It supports 120Hz displays with 8K resolution. The multiformat codec can decode 8K video at 60fps but can encode at only 30fps. By decreasing the resolution to 4K, users can achieve the full display rate at 120 frames per second.

The Exynos 2100 ups AI performance by integrating three deep-learning-accelerator (DLA) cores, one more than the 990. The AI engine, CPU, and GPU combined deliver 26 TOPS, exceeding the earlier chip by 73%. We estimate the DLA alone contributes 21 TOPS, a 40% increase per core. While expanding the DLA, Samsung halved the power consumption per core by employing weight compression.

The S21 Ultra features five cameras: the main 108MP sensor, one 12MP ultrawide, two 10MP telephoto (3x and 10x magnification), and one 40MP front facing. It also has a laser sensor to measure distance. The image processor (ISP) supports all four rear cameras simultaneously, reducing latency when switching between modes. Additionally, it overlays multiple images to improve zoom quality and enhance details for ultrawide shots. For example, the image processor uses the ultrawide image for the background while stitching together two telephoto images for a crisp foreground.

The 2100 introduces a computational-photography variation that Samsung calls semantic segmentation. The ISP and DLA stencil an image pixel by pixel, assigning a label to each data structure. The processor segregates labels by foreground and background objects and, on the basis of these labels, renders pixels with varying detail to make salient foreground objects pop.

Samsung continues to rely on its leading-edge LPDDR5 chips. Last year’s Galaxy S20 was the first smartphone to introduce LPDDR5 memory. The 2100 also supports LPDDR5, but it boosts the memory bandwidth by 16% to 51.2GB/s. That additional throughput comes with a power penalty, however.

Integrating mmWave

Using design advancements and a greater transistor budget thanks to 5nm manufacturing, Samsung integrated the 5123 modem with the Exynos 2100. The Exynos 990 also used the 5123, but kept it separate from the main SoC. The modem supports both mmWave and sub-6GHz bands, but the company didn’t implement mmWave capability in previous phones.  

For mmWave, the 5123 delivers peak downlinks of 7.4Gbps and peak uplinks of 3.7Gbps. Teardowns confirm the 2100 has two Samsung mmWave RF modules, making it only the second company (behind Qualcomm) to ship a complete mmWave solution. For sub-6GHz, it enables two-carrier aggregation at up to 5.1Gbps down and 1.9Gbps up. The modem supports up to QAM-256 for sub-6GHz and QAM-64 for mmWave. Operators in the US and abroad handle both modulation schemes, but they still trail Qualcomm’s QAM-1,024.

Like its predecessor, the 2100 lacks integrated Wi-Fi and Bluetooth. Instead, it requires an external combo chip such as Broadcom’s BCM4389. The S21 Ultra features Wi-Fi 6E, providing access to the 6GHz band to reduces network congestion; both the Qualcomm- and Broadcom-enabled versions have this capability.

Leading CPU but Lagging GPU

Table 1 shows 2100 and 888 both use Cortex-X1 as their prime core and achieve similar peak frequencies, resulting in a virtual tie for single-core performance (see MPR 12/21/20, “Snapdragon 888 Delivers 26 TOPS”). But because of its faster Cortex-A78 and Cortex-A55 cores, the 2100 outscores that competitor by 16% in multicore performance.

Table 1. Premium Android smartphone processors. ST=single thread; MT=multithread. The faster Cortex-X1, Cortex-A78, and Cortex-A55 CPUs in Exynos deliver best-in-class CPU performance. *Relative to a 1.0GHz Cortex-A9 across a set of mobile benchmarks; †GFXBench Aztec Normal Offscreen test. (Source: vendors, except †GfxBench.com and AI-Benchmark.com, and ‡The Linley Group estimate)

Although it lags the Kirin 9000 slightly in prime-core frequency, the 2100 offers a resounding 20% greater single-core performance (see MPR 12/14/20, “Kirin 9000 Is Huawei’s First 5nm Chip”). The X1 delivers much greater IPC than the older Cortex-A77, compensating for the frequency difference. The 2100 leads in multicore performance by 17%, also because of more-advanced IP and greater frequencies across the board.

Both the Exynos 2100 and Kirin 9000 implement Arm’s newest GPU. But the former has only 14 cores, while the latter has 24. Thus, Exynos lags both its competitors in graphics performance by about 19%. Even though Samsung made large GPU gains, they were insufficient to close the gap. The company ceded its even footing with HiSilicon and yielded further ground to Qualcomm.

Samsung devoted lots of die area to its AI accelerator, increasing the chip’s total TOPS to 26 and thereby matching the 888. However, TOPS doesn’t tell the entire story—early AI-Benchmark scores show the 2100 lagging the 888 by 35%. We also estimate the 2100 will trail the 9000 by 29%.

All three chips feature an integrated modem that works with sub-6GHz and mmWave bands, but only the Exynos 2100 and Snapdragon 888 implement the front-end RF technology for the latter. The 2100 matches the 888 in mmWave downlink rates but surpasses it by 700Mbps on the uplink. For sub-6GHz, the 2100 and 888 lead the Kirin 9000 by 11% for downloads, while the 2100 trails its competitors by 600Mbps for uploads.

Both the 2100 and 888 lack integrated Wi-Fi, whereas the 9000 implements Wi-Fi 6. But these two processors can offer Wi-Fi 6E by integrating external connectivity chips. Although the theoretical speeds differ marginally, access to the 6GHz band reduces latency in dense urban environments. We estimate the 2100’s BCM4389 matches the 888’s FastConnect chip by delivering a theoretical maximum link speed of 3.6Gbps.

AMD GPUs to the Rescue

The Exynos 990 marked the end of Samsung’s venture into custom CPUs. Those designs allowed Exynos to supply greater single-core performance than Arm’s Cortex-A CPUs, albeit with poor area efficiency. At the time of its release, Arm’s Cortex-X family matched Samsung’s Mongoose but was more area efficient; it did so by adopting the central theme of the Mongoose project and improving on it. Switching to the X1 enabled Samsung to decrease the relative area of its CPU, helping it reintegrate the modem without drastically increasing die cost.

As the Galaxy S21 rolls out with both Exynos and Snapdragon chips, comparisons are only natural. Although Exynos differentiates itself by providing better multicore CPU performance, its weak graphics throughput continues to be a sore spot. The Snapdragon 888 adds elite-gaming features along with an industry-leading GPU, while the Exynos 2100 features a modest shader-count increase with limited hardware accelerators. To close this gap, Samsung confirmed that its next-generation Exynos flagship will integrate an AMD RDNA graphics unit. We expect the new technology will be sufficient to that end.

The Galaxy S21 made Samsung the first vendor other than Qualcomm to deliver a complete mmWave solution, including its own RF module. By achieving this landmark, the company paints a bright and diverse future for mmWave suppliers and brings options to a once monopolistic market. We expect HiSilicon and others to eventually provide similar capabilities.

The Exynos 2100 provides generational improvements but lacks unique features. Flagship customers make their purchases on the basis of features as well as benchmarks. They want built-in deep-fake detection. They want GPUs optimized for cloud gaming. They want more than the standard smartphone chip. The 2100 lacks these standout capabilities and fails to emerge from Snapdragon’s vast shadow.

Price and Availability

The Exynos 2100 is in production, and initial phones are shipping now. Samsung doesn’t disclose its chip prices. Further details are at news.samsung.com/global/samsung-sets-new-standard-for-flagship-mobile-processors-with-exynos-2100.

Free Newsletter

Linley Newsletter
Analysis of new developments in microprocessors and other semiconductor products

Events

Linley Spring Processor Conference 2022
Conference Dates: April 20-21, 2022
Hyatt Regency Hotel, Santa Clara, CA
Linley Fall Processor Conference 2021
Held October 20-21, 2021
Proceedings available
Linley Spring Processor Conference 2021
April 19 - 23, 2021
Proceedings Available
More Events »