Amlogic S928X Cortex-A76/A55 will soon power some 8K TV boxes running Android TV from SDMC and others, and we now have detailed specifications as a “Quick Reference Manual” dropped in my inbox last night.
While like most Amlogic processors, the S928X is designed for TV boxes, the penta-core SoC may end up in affordable yet powerful single board computers with features such as HDMI 2.1a, Gigabit Ethernet, PCIe 2.0 or USB 3.0.
Amlogic S928X specifications:
- CPU sub-system
- 1x Arm Cortex-A76 core and 4x Arm Cortex-A55 cores in big.LITTLE configuration
- NEON and Crypto extensions
- Private L2 cache and unified system L3 cache
- Build-in RISC-V core for system control processing
- GPUs
- Arm Mali-G57 MC2 3D GPU with support for OpenGL ES 3.2, Vulkan 1.2, and OpenCL 2.0
- 2.5D graphics processor for scaling, alpha, rotation, and color space conversion operations
- VPU
- Amlogic Video Engine (AVE) with dedicated hardware decoders and encoders
- Decoding
- AV1 MP-10 @ L6.1 up to 8Kp60
- VP9 Profile-2 @ 6.1 up to 8Kp60
- H.265 HEVC MP-10 @ L6.1 up to 8Kp60
- AVS3 Phase 1 up to 8Kp60
- AVS2-P2 Profile up to 8Kp60
- H.264 AVC HP @ L5.2 up to 4Kp60
- MPEG-4, WMV/VC-1, AVS, MPEG-2, MPEG-1 up to 1080p60
- MJPEG and JPEG unlimited pixel resolution decoding (ISO/IEC-10918)
- Support multi-video decoder up to 4x 4Kp60
- Encoding
- JPEG image encoding up to 4Kp60
- H.265 video encoding up to 4Kp60 with low latency
- H.264 video encoding up to 4Kp60 with low latency
- 8th Generation Advanced Amlogic TruLife Image Engine with support for Dolby Vision (optional), HDR10+, HDR10, HLG, and HDR Vivid processing
- AI accelerator – Up to 3.2 TOPS Neural Network Accelerator (NNA) with Tensor Processing Unit (TPU) architecture, supports TensorFlow and Caffe
- Memory and Storage Interface
- Up to 8GB RAM DDR4-3200/LPDDR4/4X-4266 SDRAM
- SDSC/SDHC/SDXC card and SDIO interface with 1-bit and 4-bit data bus width supporting spec version 2.x/3.x/4.x DS/HS modes up to UHS-I SDR104
- eMMC and MMC card interface with 1/4/8-bit data bus width fully supporting spec version 5.1 HS400 with HW CQ
- Video Output – HDMI 2.1a transmitter including both controller and PHY supporting CEC, Dynamic HDR, and HDCP 2.2/2.3, up to 8Kp60 max resolution output with support for eARC, VRR, QMS, QFT, ALLM, DSC & SBTM (Source-Based Tone Mapping)
- Audio
- Support for MP3, AAC, WMA, RM, FLAC, Ogg Vorbis, Opus, SRS Truvolume, Dolby Audio (optional), DTS (optional), and programmable with 7.1/5.1 down-mixing
- Low-power VAD
- Built-in SPDIF input/output up to 192KHz 16/24/32bit stereo
- 3x built-in TDM/PCM/I2S ports with TDM/PCM mode up to 48kHz x 32bits x 32ch or 192kHz x 16bits x 16ch and I2S mode up to 384kHz x 32bits x 16ch
- Digital microphone PDM input with programmable CIC, LPF, HPF, support up to 8 DMICs
- Built-in stereo audio DAC
- Digital Television Interface
- 4x serial transport stream (TS) or 2x serial and 1x parallel TS input interface with built-in demux processor for connecting to external digital TV tuner/demodulator
- Built-in PWM, I2C, and SPI interfaces to control tuner and demodulator
- Integrated ISO 7816 smart card controller
Integrated I/O Controllers and Interfaces
- Networking
- Integrated IEEE 802.3 10/100/1000M Ethernet MAC with RGMII interface
- WiFi and Bluetooth via PCIe/SDIO/USB/UART/PCM
- Network interface optimized for mixed WIFI and BT traffic
- USB – 1x USB XHCI OTG 2.0, 2x USB 2.0 host, 1x USB 3.0 (5 Gbps) multiplexed with PCIe 2.0
- PCIe – 1x PCIE 2.0 up to 5Gbps (multiplexed with USB 3.0)
- Other peripherals
- Multiple PWM, UART, I2C, and SPI interfaces with slave select
- Programmable IR remote input/output controllers
- 10-bit SAR ADC with 4 input channels
- GPIOs
- Misc
- General-purpose timers, counters, DMA controllers
- 24 MHz crystal input
- Embedded debug interface using ICE/JTAG
- Integrated Power On Reset(POR) module
- Security
- Trustzone-based Trusted Execution Environment (TEE)
- Secure boot
- Crypto Engine
- AES/block cipher with 128/256-bit keys, standard 16 bytes block size and streaming ECB, CBC, and CTR modes
- TDES block cipher with ECB and CBC modes supporting 64-bit keys for DES and 192-bit keys for 3DES.
- SM4 block cipher with ECB, CBC, CTR modes
- RSA block cipher with 4K-bit keys
- Hardware crypto key-ladder operation and for transport stream encryption
- Built-in hardware True Random Number Generator (TRNG), CRC, and SHA-1/SHA-2/HMAC SHA engine
- Hardware ECC
- TS descrambler/scrambler and TS demux
- Built-in 4k bits One-Time-Programming memory for key storage
- Pre-region/ID memory security control and electric fence
- Hardware-based Trusted Video Path (TVP), video watermarking, and secured contents (needs SecureOS software)
- Secured IO and secured clock
- Package – 17 x 16 mm, FCBGA
The block diagram is a little confusing with regard to USB 3.0, and the rest of the document makes it clear it’s either USB 3.0 or PCIe 2.1, and there’s no extra dedicated USB 3.0 port. The Amlogic S928X processor also has IP blocks for artificial intelligence and machine learning with a 3.2 TOPS and a VAD. It won’t be suitable for smart displays though since the only video output interface shown in the document is the HDMI 2.1 port, and there isn’t any camera port either.
As far as software support is concerned, we are just told about a standard development environment utilizing a GNU/GCC Android toolchain, but we already know S928X will support Android 13 from the previous announcements. It does not look like Linux will be supported, at least not by Amlogic themselves, since it’s not the target market. For some specific use cases, It could still be a potential Rockchip RK3566 competitor with fewer interfaces, but an additional big Cortex-A76 core, 8K support, and a faster AI accelerator.
Jean-Luc started CNX Software in 2010 as a part-time endeavor, before quitting his job as a software engineering manager, and starting to write daily news, and reviews full time later in 2011.
Support CNX Software! Donate via cryptocurrencies, become a Patron on Patreon, or purchase goods on Amazon or Aliexpress
Typo?
It looks fine to me. Where do you think there’s a typo?
I suspect more than one reader will be surprised by the single A76 core! you basically have more total power on the small cores than on the big one in this CPU. That’s a bit odd.
It’s unusual, but moving from a typical 4x A55 TV box (e.g. RK3566) to this could be really good, with the nice single-threaded speedup.
This SoC could become much more common than those Intel 1+4 CPUs.
> moving from a typical 4x A55 TV box (e.g. RK3566)
I would better compare it with their own S905X4 since asides improved media/video capabilities, more/faster memory and one additional A76 core slapped onto the die S928X seems to be an S905X4 evolution.
Wrt I/O even little RK3566 is a big one compared to any Amlogic offering (had some hopes for more than USB3/PCIe Gen2 multiplexed on S928X but maybe that’s something for an A311D2 successor).
It does everything most people would want for a TV box, that’s how I’d analyze it.
8 GB RAM which should be overkill, AV1/H.265/VP9 8K60 decode, and more responsive with the A76 core.
Exactly. I think it will be good for a TV Box but won’t be anything great for a SBC or Retro Gaming SoC.
As we seen with the RK3399, even having some high performance cores doesn’t mean it will be a powerful machine. You really need a quad core of A7X cores to see some decent performance.
I get your point, but let’s face it, when is the last time you saw a workload that required more than one core AND whose execution would not have benefitted from a bigger one instead ? What I mean is that IF you have one A76, you just need one A55 to run the low-importance tasks and be done with it. Multi-core CPUs are made to increase total performance by parallelizing work on multiple cores. When you have enough work to keep multiple cores busy, you will always want to see all this work done on the faster core. Look at RK3399: it’s 2xA72+4xA53, in practice it’s 2xA72, period. The A53s are totally useless on it. For me a 2xA76+1xA55 has was more value than 1+4, which is equivalent to 1+1.
Nowadays cores are mostly sold by marketing bullshit claiming “run more tasks in parallel” because the vast majority of customers have no idea what an operating system scheduler is nor that their tasks are really run on the same CPU naturally, so they believe that they can run as many applications as they have cores. But I’d like to see at least one instance of an A76 being used and 4 A55 used at more than 66% each so that the 4 are really needed. In practice you’ll see all of them being used at just a few percent and one use moderately, meaning that all A55 tasks can be moved to a single A55.
For me big.LITTLE should be: “put as many big cores as needed, add a single little one for occasional background tasks and avoid touching the big power-hungry once if not strictly necessary”.
OK some will say “run some specific software tasks in the background on dedicated cores”, but we’re not talking about building radars with ultra-precise time measurements etc, just TV boxes to display movies and receive orders over IR from a remote.
Big little actually was for better battery use the sales team spun it when they noticed how it confuses people. The S928X are for TV, TV Box, Signage that interface to other home, work, media devices.
Will see how it goes.
I know and it makes a lot of sense when you have several big cores and a little one, to run mostly idle tasks without waking up the big ones. I’m just saying that *for me* there’s no point having more than one little core as soon as there’s a least one big core.
I understand where you come from. Nvidia Tegra Note was a tablet that had a SoC with one low power core, year 2013.
I have two here (on Android 5 ) one is broken.
Sales will be the judge and time will tell. 🤔
I understand Android devices may require a core with peak performance, e.g. for web browsing, and several smaller cores for background tasks or low-power tasks (audio, video playback).
Recent smartphones SoC are designed in a similar way as the S928X, but with three clusters: 1x Cortex-X2/3 premium core, 3x Cortex-A710 high-performance cores, 4x Cortex-A510 efficient cores.
For instance, check out the Mediatek Dimensity 9200.
> I understand Android devices may require a core with peak performance
Or one of Amlogic’s core customers like Amazon simply defined a minimum value for ‘single-threaded peak CPU performance’ and by adding an old A76 onto the die requirements were met.
Curious whether it’s like on A311D2 where the A76 will be cpu0 (where most of generic stuff ends up w/o proper settings) or cpu4.
> minimum value for ‘single-threaded peak CPU performance’ and by adding an old A76 onto the die requirements were met
That’s extremely possible indeed. “we want the device to react very fast to a screen touch” => “ok, let’s make sure most of the work is being done fast when this happens and that the CPU quickly leaves idle and runs at full speed for a few seconds”.
I know but that’s marketing BS. All their so-called “background tasks” can trivially run on a single core. Regarding the 9200 I’d have done it with 1 big, 3 medium and 1 small only, so that the small one is used only when the device is mostly idle. By the way the myth of the “single-threaded performance task” doesn’t hold nowadays for most use cases as we all know that it’s essentially browsers that are called like this, and again, marketing claims that “one browser=one task” except that running “top” on our browsers while a page is loading quickly shows that more than 100% is being used, hence multiple high-performance cores can be useful for short-duration peak activities like this.
> Where do you think there’s a typo?
Maybe s/big.LITTLE/DynamIQ/ ?
Yes, it could be, but the document is still using “big.LITTLE”.
But at what price?
That’s the issue?
It has novel features but do they under cut the price of the S905X 3/4 etc ?
To much and you hit RK3588S prices. Use and software that talks to other home hardware will decided, I suggest.
PS will you be reviewing the broadcon idea3588 * Processor: Rockchip RK3588 octa-core processor, 4x Cortex-A76 cores @ 2.6GHz, 4x Cortex-A55 cores @ 1.8 GHz
* RAM: 4GB LPDDR4
* Storage: 8GB
* Interfaces: 4x USB2.0 Host, 1x USB3.0 Host, 1x USB3.1 Type-C, 2x UART, RS485, CAN, Ethernet, HDMI OUT, HDMI IN, DisplayPort, MIPI CSI, MIPI DSI, PCIe2.0, PCIe3.0, SATA3.0, SD+SIM, Audio I/O, RTC, etc.
* Operating system: Android 12, Debian 11, buildroot
* Board size: 165mm x 120mm
and software?
It will be more expensive than S905X because of the extra features, but fairly lower than RK3588S.
Boardcon does not typically send boards for review, I’ve never seen software from them (AFAIK, they only send it to customers), and I think I’ve reviewed enough RK3588 devices, so if I’m asked to review another one it would better have some special features.
Yes its getting reviews of the same old venders, never bite the hardware supplier.
It has one native and one muxed USB 3.0 port, no?
I think not. All USB 3.0 pins read like “USB3.0/PCIE”.
All the anticipation for meh.
It will need to b cheap to compete.
If their target is TV, Android TVs and stuff like that, i dont know why use those small cores.
Low power consumption isnt needed because you will be always connected to a power cable. You also arent doing lots of things at the same time in an Android TV, stuff like background tasks are minimal to a A76 CPU.
It would make sense having a single A55 for small tasks and to be on while the Android TV is off, so it can update while sleeping and at least two A76 to make sure that the performance is there when opening an App or processing something.
Thank you I feel less alone 🙂
Most probably the reason for 4 little cores is as easy as in ‘DynamIQ big.LITTLE only allowing for certain cluster designs like 1+3, 1+4, 1+7, 2+4, 4+4, …’.
Less little cores than 3 can’t be chosen and 1+4 looks less weird than 1+3. Something like that.
Amlogic does not design CPUs but simply relies on what ARM offers to combine it with Amlogic’s own and 3rd party IP blocks.
That would make lots of sense. I thought everyone would just buy the license of the IPs and do their own configurations, but this explain why all smaller SoC companies have similar CPU settings
As for the reasons why ARM engineers prefer many little over one big core I could imagine at least the following:
Die space (if we can trust into this die shot and the area labelings then an A55 is almost twice as space efficient compared to an A76), DVFS or in general energy efficiency even if this doesn’t matter to some people or use cases (power cable).
Indeed, that could be an explanation.
Both of us should also keep in mind that nobody at ARM or $SoC maker wastes a single thought on the use cases we’re interested in 🙂
It’s all about the typical Android e-waste scenarios and if the SoCs become feature monsters then sophisticated stuff like ‘smart vehicle’ and similar stuff.
Will the Android scheduler be aware of all cores in Amlogics, Android 12. Seems Arm patch is in the Linux Kernel. Have Amlogic embraced dynamic IQ.
“The reason we haven’t seen big.LITTLE MP to date is because updating the Android scheduler to be aware of all four cores simultaneously has taken time; the functionality isn’t baked into Android 4.3.”
2013 Extremetech