Amlogic A311D2 octa-core Arm processor supports up to 16GB RAM

Amlogic A311D2 octa-core processor is the successor to Amlogic A311D hexa-core SoC with four Cortex-A73 cores, four Cortex-A53 cores, a more powerful Mali-G52 MP8 (8EE) GPU, support for eDP and LVDS video interface, 8Kp24 AV1 video decoding, and support for a whopping 16GB LPDDR4/X memory.

But it’s not all! The chip also includes supports for three HDMI 2.1 inputs up to 4Kp60, a 16MP ISP, 4K video encoding, and the more powerful GPU and higher memory bandwidth enable 4K user interfaces which were not feasible on most other Arm hardware barring the NVIDIA Shield.

Comparison of Amlogic A311D2 and A311D specifications.

Model
Amlogic A311D
Amlogic A311D2
CPU
4x Cortex-A73 at 2.2GHz, 2x Cortex-A53 at 1.8GHz4x Cortex-A73 at 2.2GHz, 4x Cortex-A53 at 2.0GHz
GPUArm Mali G52 MP4 (6EE) at 800MHzArm Mali-G52 MP8 (8EE)  
NPU5 TOPSTBC, extra license required
RAM Up to 4GB 32-bit LPDDR4/X @ 1608MHzUp to 16GB 64-bit LPDDR4/X @ 2133 MHz
Display I/F1x HDMI 2.0 up to 4Kp60
1x MIPI DSI
1x HDMI 2.1 port up to 4Kp60
with EMP, ALLM, VRR, QMS, eARC, etc...
2x MIPI DSI up to 1920x1080
LVDS
2x 8-lane V-by-One up to 4Kp60
2x 4-lane eDP up to 2560x1600
Camera I/F1x 4-lane MIPI-CSI with 8MP ISP2x 4-lane MIPI-CSI with 16MP ISP
HDMI InputNo3x HDMI 2.1 up to 4Kp60  
4K UINoYes
Video Decoding
4Kx2Kp60 + 1x1080p60
Main Video Codecs4K H.265, VP9, and AVS28Kp24/4Kp60 AV1, H.265, VP9, 4Kp60 AVS2
Video EncodingH.265 & H.264 at 1080p60H.265 & H.264 at 4Kp50
EthernetGigabit Ethernet MAC
Fast Ethernet PHY
Gigabit Ethernet MAC
Fast Ethernet PHY
USB*1x USB 3.0 multiplexed with PCIe
1x USB 2.0 with OTG support
1x USB 3.0
1x USB 2.0 with OTG support
PCIe*1-lane PCIe multiplexed with USB 3.01-lane PCIe
Process
12 nm

* Amlogic A311D multiplexed PCIe and USB 3.0  interfaces means two configurations are possible: 1x USB 2.0 Host + 1 PCIe, OR 1x USB3.0 (No PCIe), but Amlogic A311D2 PCIe and USB 3.0 interfaces are independent, so we are not losing USB 3.0 when using PCIe (5 Gbps).

The quick reference manual does not include any specific information about an AI accelerator, except for an optional dual HiFi 5 audio DSP with Neural Network Extension for audio/voice processing, so there’s no indication about an NPU for computer vision application, but we’ll have to see. At first, I thought Amlogic A311D2 may be a variant of the long-expected Amlogic S908X, just like A311D and S922X-B are practically the same processor, but while 8K video decoding is supported, 8Kp60 video output is not.

I would only expect Amlogic A311D2 to outperform A331D in terms of CPU power in highly parallel tasks, but the real improvements are with the GPU, 4K video encoding, HDMI 2.1 features support, separate USB 3.0 and PCIe interfaces, HDMI inputs, and obviously the extra RAM, which would allow new applications such as multiple camera NVR’s, more powerful and versatile Arm Linux mini PC’s, SBC’s with additional high-speed interfaces, etc…

Share this:

Support CNX Software! Donate via PayPal or cryptocurrencies, become a Patron on Patreon, or buy review samples

38 Replies to “Amlogic A311D2 octa-core Arm processor supports up to 16GB RAM”

  1. Nice. I wonder if this is a stop gap whilst e wait for the S908X SOC.

    AV1 support will give it a longer life.

    I suspect an Odroid N2++ soon.

  2. Nice, we’ll probably see it on an updated VIM3 or Odroid-N2+. I don’t think the 2 extra cortex A53 will bring any value here, however it allows them to play in the “octacore” court yard for very little extra silicon and likely a very little extra licensing cost, if any at all.

    1. > I don’t think the 2 extra cortex A53 will bring any value here

      First VIM4 sbc-bench result flew in few hours ago. A kitchen-sink benchmark like 7-zip shows almost 25% more performance compared to A311D/S922X devices even with the A73 in A311D2 clocked only at 2.2GHz right now. Memory performance has greatly improved…

      1. OpenSSL scores reflect CPU clockspeeds perfectly: A53 @ 2.0GHz show the usual ~940800 aes-256-cbc score, A73’s difference (1252070 on A311D2 vs. 1365895 on A311D) is exactly the 9% clockspeed difference: 2.2 vs. 2.4 GHz.

        Once access to RAM starts to matter (e.g. 7-zip use case) with the A53 nothing changes but the A73 even when clocked slightly lower perform almost 15% better. So it’s more about an overall improved SoC design than those two extra A53 🙂

  3. Am I reading that right? The memory interface width doubled? And it’s clocked 25% higher? There’s your improvement right there.

    1. Look at the VIM4: two H9HCNNNCPMMLXR-NEE next to each other (LPDDR4X, 2133 MHz, 32Gb). Curios how this will affect performance. I mean it’s still just lame/slow A73 here…

      1. I partially disagree. A73 isn’t *that* slow, it’s uneven. On some workloads it can be 25% slower than A72 and on others 50% faster. It’s less easy to optimize the code for it, but code optimized for A53 may run more efficiently on it than on A72. I found memory-bound workloads like compilation to work signifcantly faster on A73 clock-for-clock, but others like software crypto to be slower. However the S922X in Odroid-N2+ can run at 2.4 GHz and this will clearly compensate for the losses compared to A72. I guess the VIM4 could improve things on memory-bound workloads, because clearly, 32-bit DDR was not enough to feed 4 big cores. However one must keep in mind that ARM processors in the top of the range tend to result in boards that cost more than equivalent x86 PCs.

  4. What would be a even a bigger killer would be if the made a s922x + without the LCD etc parts and gave it a G52 12mp bigger MP , keeping the memory bandwidth 64 Android and 16GB memory. Way beyond the Toaster 4 desktop Linux.

  5. It’s nice and all that but I was hoping for at least A76 and even maybe some A78 at one point.
    Apple is miles ahead and we can’t even get some A76 SOC for some decent boards and TV boxes.
    Kinda makes me appreciate the low end Intel stuff, at least those work out of the box in Linux.

    1. > Kinda makes me appreciate the low end Intel stuff, at least those work out of the box in Linux.

      That’s exactly the problem most SBCs suffer from: most of the time they’re not trivial to install, and once they stop responding you have to use a screwdriver and figure how to wire a serial console with a USB-TTL UART board to fix them. On a PC, you just connect a keyboard and screen, you enter the BIOS or boot loader and fix your boot.

      This trouble is acceptable when there is a significant cost saving, i.e. when using a $30 ARM board instead of a $200 x86 one. But ARM boards that are approaching x86 prices really need to be careful about this, and the serial console MUST always be present and externally accessible without having to use a screwdriver. For now it seems that only SolidRun has understood this, and I had to spend lots of time manually installing one in most of my other boards, drilling enclosures etc 🙁

      1. Indeed. We are on a terrain that we are reaching x86 prices… from 2 years ago. I am looking for to all those x86 mini pcs that where 200 bucks(larkbox, etc).. they are gone. There is nothing on the table that you can actually buy. Thats why the odroid x86 sbcs is gone too.

        The main cost of the newer arm sbcs arent the socs mostly, it’s the ram. Even the rpi now have 8 gb models.. you can expect an 8gb sbc to cost 35 bucks… that’s the same on arm and x86.

    1. Well, staring at Geekbench scores (especially v4) is kinda stupid but anyway: https://browser.geekbench.com/v4/cpu/compare/16291496?baseline=16426661

      Given that A311D2 is said to cost as much as RK3588 anyone who’s limiting himself to ARM and just after ‘CPU horsepower’ is better served by something based on RK3588S. Almost twice as fast, probably less expensive and wrt software support let’s see what happens faster: A311D2 or RK3588(S) running without BSP kernels.

        1. The A311D2 is designed for signage but has other uses, software is not the problem, it is getting Amlogic to allow the SoC designers to work with software writers.

        2. Talking about VIM4? If so please include dmesg output already in the ‘unboxing’ blog post or at least early on when running the thing first time with Linux. Seems Khadas has a Jammy image based on 5.4.125 BSP kernel ready…

          1. By the way, I had to disable load check in sbc-bench.sh as the load average is about 2.0 at idle. I can see it looks the same in other VIM4 results. It starts at 2.0 and goes to almost 10. It’s probably something Khadas needs to fix.

          2. Well, average load somewhere above 1 is a well known symptom of running BSP kernels (not limited to Amlogic). Since on Linux ‘average load’ is not CPU utilization it might be worth a try to find processes that are stuck in D state:

          3. There’s no output from the command:

            Here’s the output of the command without filteting:

          4. Ah, it’s called ‘D (disk sleep)’ then. ps manual page tells about states: ‘D  uninterruptible sleep (usually IO)’.

            PIDs 53 and 178 are the affected ones. What processes are these? As a starting point:

          5. “cmdline” appears to be empty

            Those must be part of the Amlogic BSP, because I can’t find any info on the web about those processes.

          6. Thank you! I reported it to Khadas via their forum. Let’s see what happens (most probably nothing).

          7. > I had to disable load check in sbc-bench.sh

            Fixed with v0.9.5 (now also taking CPU utilization into account when average load remains too high). But would be still interesting which processes are to blame for loadavg being that high…

Leave a Reply

Your email address will not be published.

Advertisement
Advertisement