Today, Samsung LSI took the virtual stage to announce the new Exynos 2100 flagship SoC. The new chip is quite special for Samsung’s chip division as it marks a departure from past iteration designs – being the first SoC not using Samsung’s own in-house performance CPU microarchitecture, but instead relying on Arm’s Cortex cores such as the new X1.

It’s a big change for Samsung not just technologically, but also in terms of general tone and messaging – with a much more upbeat enthusiasm from the company and also the first ever actually presented launch event for a Samsung mobile SoC, more closely mimicking what we would see from Qualcomm.

The new SoC promises some very large technical upgrades – the aforementioned new Cortex-X1 performance CPU, a very large GPU performance boost, but also very large gains in things like AI performance, a re-integrated leading 5G modem, camera support for up to 200MP sensors, AV1 video decoding, all on a new 5nm process node.

Samsung Exynos SoCs Specifications
SoC

Exynos 990

Exynos 2100

CPU 2x Exynos M5 @ 2.73GHz
2MB sL2
3MB sL3
+
2x Cortex-A76 @ 2.50GHz
2x 256KB pL2
4x Cortex-A55 @ 2.00GHz
4x 64KB pL2

1MB sL3
1x Cortex-X1 @ 2.9GHz
3x Cortex-A78 @ 2.8GHz
4x Cortex A55 @ 2.2GHz

?MB sL3
GPU Mali G77MP11 @ 800 MHz Mali G78MP14 @ ? MHz
Memory
Controller
LPDDR5 @ 2750MHz

2MB System Cache
LPDDR5 @ ?MHz

?MB System Cache
ISP Single: 108MP
Dual: 24.8MP+24.8MP
Single: 200MP
Dual: 32MP+32MP

(Up to quad simultaneous camera)
NPU Dual NPU + DSP
>10 TOPs
Triple NPU + DSP
26 TOPs
Media 8K30 & 4K120 encode & decode
H.265/HEVC, H.264, VP9
8K30 & 4K120 encode &
8K60 decode

H.265/HEVC, H.264, VP9
AV1 Decode
Modem Exynos Modem External 

(LTE Category 24/22)
DL = 3000 Mbps
8x20MHz CA, 1024-QAM
UL = 422 Mbps
?x20MHz CA, 256-QAM

(5G NR Sub-6)
DL = 5100 Mbps

(5G NR mmWave)
DL = 7350 Mbps
Exynos Modem Integrated

(LTE Category 24/18)
DL = 3000 Mbps
8x20MHz CA, 1024-QAM
UL = 422 Mbps
4x20MHz CA, 256-QAM

(5G NR Sub-6)
DL = 5100 Mbps
UL = 1920Mbps


(5G NR mmWave)
DL = 7350 Mbps
UL = 3670 Mbps
Mfc. Process Samsung
7nm LPP
Samsung
5nm LPE

Starting off with the big-ticket item today, it surrounds Samsung’s usage of the new Arm Cortex-X1 CPU cores. This follows a 5 year journey that had started off with the Exynos 8890 back in 2016 and the Galaxy S7 series of devices- with Samsung using their own designed CPU microarchitecture dubbed the M1 through to the M5 CPU in last year’s Exynos 990 used in the Galaxy S20 series.

Unfortunately, Samsung’s own designs were never really successful, and actually brought the opposite of what the company had hoped for – instead of positive differentiation in the SoC market the usage of custom cores was actually more of a negative, bringing with them reduced performance and quite worse power efficiency compared to the Arm Cortex counterparts. While Samsung had been patient with the SARC CPU design team, too many failures in a row, particularly big blunders such as the Exynos 9810 and the Exynos 990 resulted in the termination of the project, with Samsung now choosing to simply just use Arm Cortex cores.

In last year’s Cortex-X1 announcement from Arm, Samsung was actually one of the leading partners which publicly acknowledged their use of the new high-performance CPU IP, and today’s announcement of the Exynos 2100 unveils the fruits of this collaboration.

The new Exynos 2100 follows a CPU configuration that had been first introduced by Qualcomm in the Snapdragon 855 in that it’s a 1+3+4 CPU design, featuring one new high-performance Cortex-X1 core clocking up to 2.9GHz, three Cortex-A78 cores up to 2.8GHz, and four Cortex-A55 cores at up to 2.2GHz.

What’s interesting here about this setup is the clock frequencies, and how they contrast to Qualcomm’s new Snapdragon 888 which was released a few weeks ago and is starting to ship in the first new flagship devices of 2021.

Particularly on the middle and on the efficiency cores, the Exynos chip shows a 400MHz advantage in clock frequencies over the Snapdragon chip, which should result in a notable advantage when it comes to multi-threaded CPU workloads.

On the part of the X1 cores at 2.9GHz – well it’s a shame we again just missed the mark on the symbolic 3GHz mark this generation, but it’s still 60MHz higher than the Snapdragon counterpart so it should result in a small performance boost, although I doubt it’ll be noticeable.

Samsung actually did have some performance figures shown during the presentation, including the expected 33% multi-threaded performance boost thanks to the now much higher performance middle cores.

I was actually delighted to also see Samsung post a separate 19% single-threaded performance figure – however this number is a bit odd to me as the jump from a 2.7GHz Exynos M5 to a 2.9GHz Cortex-X1 should be larger than that. I don’t know what Samsung uses in terms of coming to these figures, but I’m optimistic that things could end up higher than just 19%.

As part of the launch event presentation Samsung also claimed to have included “better cache memory” – this could be a reference to a possible 8MB L3 cache in the CPU cluster, but could also be related to a better system-cache structure. They had also made mention of a more efficient voltage control implementation – what exactly this means in practice is open to interpretation but one area Samsung could differentiate itself from the Snapdragon 888 is to give the Cortex-X1 cores their own dedicated voltage rail. We’ll have to see how the actual SoC behaves as Samsung doesn’t go into such technical depth for their chip launches.

A 40% GPU Performance Uplift?!

On the GPU side, we’ve seen the move from a Mali-G77MP11 at up to 800MHz to a new Mali-G78MP14 at an undisclosed frequency. That’s generally an expected implementation boost in terms of number of cores – 27% more cores to be exact.

What’s interesting here is the fact that Samsung is proclaiming a massive 40% better graphics performance figure. This actually really gives me pause and I question as to how Samsung has achieved such a figure. The Mali-G78 was meant to be a rather small generational upgrade with only 10% better perf/W improvements. The process node of the Exynos 2100 is advertised as bringing a 20% power efficiency improvement. In theory and on paper, the new process and GPU IP should results in a 32% better perf/W improvement but it’s very rare that such theory-crafting actually results in accurate estimates for actual silicon designs, particularly on the GPU side.

So while 40% performance improvement sounds fantastic, it’ll all depend on how much power and how efficient the new Exynos 2100 is – and how well it can sustain those performance figures in prolonged gaming sessions.

A More Powerful NPU

Another big upgrade is found on the part of the new generation NPU. Samsung has evolved their IP from 2-core design to a new 3-core design, and promises new architectural changes such as pruning and weight compression to improve the power efficiency of the NPU by 2x.

Samsung today proclaimed AI performance figure of 26 TOPS, but this figure seems to be an aggregate figure across all the SoC computational blocks, including CPU, GPU, NPU and DSPs, comparing itself to a 15 TOPS figure of the Exynos 990. The predecessor chipset was actually announced with a 10 TOPS figure last year, so there’s a bit of marketing revisionism going on here, probably wanting to match the figures that Qualcomm is putting out for their SoCs. It’s unfortunate that both companies are resorting to such blurring of the actual IP block performances.

More Powerful ISP - Up to 4 Simultaneous Sensors

In terms of multi-media features, Samsung says to have made large improvements to its camera ISP and DSP’s. The new chip supports single-sensors up to 200MP, while also connecting up to a total of 6 camera sensors, and simultaneously running 4 of them in parallel. At least in this regard, the new chip flexes higher capabilities than the Snapdragon 888 which only supports up to 3 simultaneous cameras. Edit: Qualcomm had reached out to clarify that it supports 5 simultaneous cameras in the form of 3 RGB sensors + 2 monochrome sensors - it's not clear if the Exynos also has such a distinction between it's 4 concurrent sensors.

Samsung has updated its MFC block (Multi-format Codec) video encoder and decoder to support up to 8K60 video decoding, although encoding has remained at 8K30/4K120. What’s interesting here is that Samsung does now support AV1 decoding, marking it the second chipset in the mobile market after MediaTek’s Dimensity 1000 to support the new feature.

5G modem integrated

Finally, the new platform is again a new monolithic design, reintegrating the 5G modem on the same silicon die, promising some gains in power efficiency and allowing for better PCB space usage. The actual modem itself doesn’t seem to have seen notable changes in features compared to the discrete Exynos 5123 modem – however that’s probably because it already was fully featured in terms of supporting all new technologies such as 5G sub-6GHz and mmWave connectivity.

In the announcement video, Samsung made an emphasis that mmWave deployments are growing in 2021, and that we’ll be seeing more Exynos powered devices using mmWave. This is going to be interesting as we haven’t yet seen Samsung’s own mmWave antennas in the market, although the company had let out in 2019 that they had multiple generations of such designs.

5nm Process Node

Finally, Samsung reveals that the new Exynos 2100 is manufactured on Samsung’s 5nm process node, which brings either 10% boost in clock frequencies, or a 20% drop in power at the same frequency. Last year we had suspected that Samsung’s 7nm process node was anywhere between 20 and 25% less power efficient than TSMC’s latest N7P node, so essentially we’re expecting the new process to just catch up in terms of power efficiency, with newer TSMC 5nm SoCs such as the Apple A14 and the Kirin 9000 to still have a notable process node lead.

Positive Messaging, Good Outlook

Today’s launch event was extremely positive in terms of tone and general enthusiasm on the part of Samsung, proclaiming catchphrases such as “Exynos is back” and during the presentation acknowledging that the previous generation SoCs had issues. I think that’s a major change in messaging from the company, and at least on paper the new Exynos 2100 looks to be extremely competitive. I do hope this actual also translates onto the actual products, with the upcoming new Galaxy S21 series in a few weeks’ time sporting the new Exynos chipset for the first time.

Related Reading:

POST A COMMENT

34 Comments

View All Comments

  • Rudde - Tuesday, January 12, 2021 - link

    "Samsung will only be able to use AMD’s GPU IP to counter Qualcomm’s Adreno GPUs in the mobile market"
    From https://www.anandtech.com/show/14492/samsung-amds-...
    Reply
  • yeeeeman - Tuesday, January 12, 2021 - link

    Getting a bit bored of this whole RDNA GPU hype. Are the RDNA desktop cards so...amazing? are they so efficient? what is exactly so amazing about AMD cards that makes you think they could put that architecture in a mobile GPU and fare better than the existing and much better tuned for mobile GPUs?
    I think only the 5 year olds here are getting wet about the RDNA gpu in a mobile SoC. I don't expect much from it.
    Reply
  • phoenix_rizzen - Tuesday, January 12, 2021 - link

    Considering the Adreno GPU tech originally came from AMD, everyone is hoping lightning will strike twice, and AMD will have something that can compete with (or even beat) Qualcomm's GPU. Reply
  • Kangal - Wednesday, January 13, 2021 - link

    ...maybe.
    Though the primary reason is that of fanboys. Since the revolution of Ryzen 1, and also RDNA-1, there has been a lot of people liking AMD's products. So the hope is that perhaps they can revolutionise the iGPU segment in ARM SoC's. I don't see that happening with GCN, Vega, or RDNA-1 honestly. If, and that's a big IF, they can miniaturise RDNA-2 (or something better) then they possibly MIGHT have a winner on their hands. AMD's GPU technology is the second best, but they don't have any devices in the ultra-portable segment, apart from the yet-to-be-seen Samsung designs.

    Hence this hope is not completely unfounded.
    For instance, when Nvidia made the switch from their "unified architecture" (Kepler) to their "tile-rasterisation architecture" (Maxwell), they made a pretty big improvement in performance per watt, 2014-2016 GTX 750/Ti - GTX 960 - 980Ti, especially at the low-power segment. So much so that, they ported the Maxwell GPU to an ARM SoC, and that's how we ended up with the Nvidia Shield TV and the Nintendo Switch (albeit severely handicapped on 22nm-Planar instead of 14nm-FinFet). Nvidia does have the cutting-edge best technology when it comes to GPUs, however, outside of the above examples they really haven't done much in terms of GPUs in ARM SoCs in the recent years.

    With that said, there's actually not much competition for iGPUs in the SoC market anyway. Apple is the current market leader, but they don't have to compete as iOS devices are distinct to Android devices. Still, Apple only makes medio improvements yearly, thanks to no competition and their access/customisation of PowerVR GPU technology. ImaginationTechnologies used to be the market leader in Android, but they've sunk in past 6 years or so, and recently have had an aggressive buyout from China. Qualcomm uses Adreno tech which they acquired from Radeon/ATi/AMD long ago. And they're good but they also only introduce modest improvements each year.... they do not have a competitor. The next best thing is Intel with their Iris Pro HD iGPUs. They've always been decent from a performance per watt scenario, and lately with their new "Xe architecture" in the 11th-gen Tiger Lake APUs they're pretty competitive. The only thing is, they're solely in x86 designs, as they do not make ARM SoCs, nor do they have any Atom-SoC for phones anymore. Last but not least is ARM's GPU division "MALI" from Norway produces GPU designs which they license for free or really cheaply, and thus have a big market share. However, the MALI technology is immature compared to the likes of the latest Xe, Adreno, PowerVR, RDNA, or CUDA.
    Reply
  • Unashamed_unoriginal_username_x86 - Tuesday, January 12, 2021 - link

    Here's a comparison of the two flagship processors, just copied and pasted:
    https://docs.google.com/spreadsheets/d/1ciPXd5loz1...
    A lot of details are missing, but it seems Samsung's trying to edge QC out where they can, QC probably has a good lead in DPU, ISP and GPU.
    Reply
  • Unashamed_unoriginal_username_x86 - Tuesday, January 12, 2021 - link

    Ignore that last bit I have no clue what I'm talking about there Reply
  • dragosmp - Tuesday, January 12, 2021 - link

    Andrei, this is the most hopeful and optimistic I remember seing you about Samsung Exinos proucts in years. Hope it's well placed, though the claims vs expectations do seem strain mathematics and common sense quite a bit. One can hope it's at least comparable to Qualcomm's flagship.

    My feeling is the lack of competition made Q slow down quite a bit, had to have Apple's M1 to see what Arm can actually become.

    Looking forward for a deep dive
    Reply
  • Rudde - Tuesday, January 12, 2021 - link

    Isn't Qualcomm just as far from Apple as they have been the last few years? I as well hope Samsung manages to catch up to Qualcomm with this new processor. Reply
  • abufrejoval - Tuesday, January 12, 2021 - link

    As a consumer I take a bit of offence at a statement like this: "While Samsung had been patient with the SARC CPU design team..."

    It's not Samsung's management that was tried, it was the consumers who had a sub-par part forced onto them for a premium price, simply and only because they happend to be living in the wrong part of the World.

    The sheer arrogance of that decision speaks volumes about one facet of a company that bribes heads of state as another means of getting what it wants for itsself.
    Reply
  • s.yu - Wednesday, January 13, 2021 - link

    There's very little to no barrier of importing from another region, so don't blame it on the manufacturer for your own reluctance to buy from those sources.
    And since prices are different for other brands/models from other regions the playing field is unique to each region, or else similarly you could say Xiaomi "forces a premium price" on every market beyond China.
    You're probably living in Europe but face it most things are just expensive there, but on the "up side" you could say Chinese consumers all have "sub-par parts forced onto them for a premium price" by European car manufacturers?
    Reply

Log in

Don't have an account? Sign up now