Dimensity 9000 – MediaTek’s 4nm, 3 GHz octa-core Armv9 processor for flagship 5G smartphones

MediaTek is the top mobile SoC vendor in terms of unit shipped, well ahead of the competition, but the company’s processors were mostly found in entry-level, mid-range, and “premium” smartphones, while flagship smartphones usually feature application processors from  Qualcomm, Samsung, or Apple.

But MediaTek has upped the ante with the introduction of the Dimensity 9000 smartphone processor featuring an octa-core Armv9 CPU subsystem, including one Cortex-X2 core clocked at up to 3.05 GHz, a Mali-G710 MC10 GPU, 5G connectivity, and manufactured with the bleeding edge 4nm TSMC manufacturing process.

Dimensity 9000 specifications:

  • Octa-core CPU subsystem
    • 1x Arm Cortex-X2 “Ultra Core” at 3.05 GHz
    • 3x Arm Cortex-A710 “Super Cores” at up to 2.85 GHz
    • 4x Arm Cortex-A510 “Efficiency Cores” up to 1.80GHz
    • 8MB L3 cache
  • GPU – Mali-G710 MC10 @ 850 MHz with support for Vulkan, (software) raytracing
  • AI Accelerator – 5th generation APU (AI processing unit) with up to 4x higher power efficiency
  • Memory I/IF – LPDRR5x 7500 Mbps (60 GB/s), LPDRR5 up to 51.2GB/s; 6MB System Cache
  • Storage I/F – TBD
  • Display – Support for FullHD+ display up to 180Hz
  • Camera – 18-bit HDR Imagiq790  ISP with support for up to three 32MP cameras, 320MP single camera, up to 9Gpixels/s
  • Video
    • Decode
      • 8K60 H.265/HEVC, H.264, VP9
      • 8K30 AV1
    • Encode – 8K30 & 4K120 H.265/HEVC, H.264, VP9 (TBC)
  • Wireless connectivity
    • 3GPP Release-16 5G Smartphone Modem
      • 3CC Carrier Aggregation (300MHz)
      • 7Gbps downlink peak performance (using sub-6GHz band)
      • Support for R16 UL Tx Switching for both SUL and NR UL-CA based connections
      • MediaTek 5G UltraSave 2.0 power-saving enhancement suite
    • WiFI 6E 2×2 MIMO
    • Bluetooth 5.3 (Dimensity 9000 is apparently the first SoC to support the latest Bluetooth standard), LE audio-ready
    • GNSS – Beidou III-B1C support (and I suppose the usual GPS, Galileo, GLONASS, but those are not listed)
  • Manufacturing process – TSMC N4 4nm-class process

I mixed the information from MediaTek’s product page and Anandtech to get more complete specifications, but there are still some unknown and points to be confirmed.  With regards to performance, Arm announcement of their Armv8 cores, Arm stated that the Arm Cortex-X2 flagship core would deliver 30% performance improvements over Cortex-X1, the Arm Cortex-A710 “big” CPU core  a 30% energy efficiency gain and more modest 10% uplift in performance compared to Cortex-A78, while the Arm Cortex-A510, high efficiency “LITTLE” Armv9 core would bring up to 35% performance improvements and over 3x uplift in ML performance compared to Cortex-A55.

Supposedly Dimensity 9000 vs Bionic A15 vs Bionic A12 vs Snapdragon 888

We also have several comparisons against other flagships devices, courtesy of the Anandtech post, with MediaTek claiming phones based on Dimensity 9000 will have about the same multi-core performance as Apple Bionic A15 based iPhone 13, per GeekBench 5 results, and much better performance than Snapdragon 888 smartphones. MediaTek also claims gaming performance is similar, or even better over a 24-minute period of gameplay, compared to the A15, and power efficiency has been increased by 25% (gaming) to 65% (video playback) against Snapdragon 888. You’ll find more slides about the APU, ISP, etc… shared by MediaTek on Anandtech.

If MediaTek Dimensity 9000 performance and efficiency improvement claims can be eventually verified that means high-performance, yet relatively low power, Arm chips may soon be found in SBCs, mini PCs, systems-on-module, and other embedded hardware. That means Apple may not have an insurmountable lead in the space. This is especially encouraging, as the way I understand it, MediaTek uses Arm’s own Armv9 core without additional customization/optimizations, and others may soon follow suit.

We’re told that Dimensity 9000 is currently shipping, and the first commercial devices are expected by Q1 2022.

Share this:

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

62 Replies to “Dimensity 9000 – MediaTek’s 4nm, 3 GHz octa-core Armv9 processor for flagship 5G smartphones”

    1. There is always hope, my brother.

      Mediatek Posts 8k Lines Of New Linux Kernel Driver Code For AI Processing Unit Support
      Mediatek MT8167 DRM Driver Support Coming For Linux 5.15
      Panfrost Open-Source Mali Driver Adding Mediatek MT8183 Support With Linux 5.14

      1. MT81xx have gotten a great boost from Chromebooks, they mostly work out of the box by the time hardware gets sold now. If Google decides to put the same amount of pressure on mainlining drivers for phone chips that they have for laptop chips, it won’t even be that hard to do: Dimensity-series MT89xx/MT89xx are closely related to MT81xx, and the overall system design isn’t as complex as the corresponding Qualcomm chips with their many firmware layers.

        1. What does this mean from an end user’s perspective? Taking MT8192/Kompanio 820 as an example. Next year Chromebooks based on this SoC will be available. How much time and efforts needed to operate such a thing with an off-the-shelve Linux distribution?

    2. Sure, shortly after Apple’s M1 is fully supported. Rule of thumb wrt mainline Linux on ARM: it’s ready once the hardware is obsolete…

        1. There’s also the law that Rockchip, Amlogic and Allwinner only launch processors with new Cortex-A cores at least 5 years after the initial announcement, followed by Kaiser’s law means ten years before affordable platforms become supported.

          1. We’re still waiting for Rockchip’s RK3588 with A76 when Mediatek is advertising a Cortex X2…

          2. Now would be the time to focus all mainlining efforts on this dimensity 9000 so we would have quite contemporary hardware supported.

            Mybe would be worth it, as I currently see what a piece of crap QCA drivers are and all the other cheap manufacturers just warm up their old stuff again and again.

          3. The problem is that MediaTek may not sell that chip to anybody, only larger phone manufacturers. If they could make another variant for laptops, mini PCs, or the embedded market, then the motivation would probably be much higher.

            If they can provide something that can compete on performance against some of the Apple processors, offer a similar or lower price, plus more openness (in terms of supplies), then it would be a winner.

          4. I wouldn’t even mind abusing a phone. After all it would be an alternative to the pinephone.

            But yepp mtk is not known for its openness even for the ex ralink line of WiFi SOCs its not easy to get the required (calibration) data to get the WiFi going

          5. Agreed. When I talked about “openness (in terms of supplies)” in the context above, I meant the ability to purchase the processor or a module to create custom hardware.

            It’s impossible with Apple, and I don’t think it will happen anytime soon. But I’ve seen designs with MediaTek processors even from smaller companies, and not only phones.

          6. If they could make another variant for laptops, mini PCs, or the embedded market, then the motivation would probably be much higher.

            If they can provide something that can compete on performance against some of the Apple processors, offer a similar or lower price, plus more openness (in terms of supplies), then it would be a winner.

            Latest one
            https://www.xda-developers.com/mediatek-teases-kompanio-1200-chromebooks-not-ready/

            But performace is more at smartphone level.
            Since the kompanio line based on their phone SOCs, it is unlikely i can or will exceed M1 which is design for laptops/Mini-PC power envelop

      1. >Rule of thumb wrt mainline Linux on ARM: it’s ready once
        >the hardware is obsolete…

        And it’ll be actively maintained for a whole two weeks after it’s finally mainlined as no one is interested anymore. (Allwinner stuff being the weird exception).

  1. Everyone is still trying to get 2019 silicon out the door. The next three years will be yesterday’s technology at tomorrow’s prices for SBCs.

  2. “that means high-performance, yet relatively low-performance, Arm chips” … what does that mean? Or is it a typo?

    1. That means my mental abilities are declining by the day.
      I’ve corrected my sentence. I meant low power in the second part.

  3. Hey, frankly, rk3588 while not top of the line, will be extremely affordable and powerful for the price.. even if it takes some years to get fully supported on mainline linux, it’s far better than expensive and heavily packed with blobs like any of the high end stuff.

    Also, for apple silicon, the all attempt to run linux arm there is useless. No only we lack of basic drivers, macos performance rely on every accelerator inside the soc that will require a huge and mostly impossible effort to reproduce on linux. We will never be competitive there. BUT, as desktop, we are the kings on mostly any cheap arm sbc, even if we lack of vpu on most platforms. A realistic approach is always to focus on where we can win, apple silicon isn’t.

    Going back to rk3588, it’s the first time we have something close to recent years flagships at affordable prices. The gap is quite close now. I don’t care about dimensity 9000 or M1 super pro 2×4 bc they are useless on mainline linux and they always will be. Also, not affordable products…

    1. >  rk3588, it’s the first time we have something

      Did I miss something? Vaporware turned hardware already?

      I really wonder about the excitement… RK3588 was initially announced to be available Q1/2020, then ‘Q3/Q4 2020’ and the last time for ‘Q2/Q3 2021’ which is not that realistic any more, isn’t it?

        1. Why would I even care? If keen on ARM we try to focus on inexpensive and/or energy efficient and/or performant (two of them, not all three at the same time of course). Most importantly: hardware we’re actually able to buy if/when needed (including correct invoice / warranty handling).

          Ended up with Apple or even stuff from RPi Trading Ltd.

          1. Hahaha you ended up buying rpi stuff.. you ended badly then. That’s not the Tkaiser I was used to hear. Nice toys they are.

          2. RPi5 will probably use Cortex-A75, which is the last one they can put on 28nm. No guarantee of any other aspect of it being good at all, like the GPU.

          3. The real transition with RPi 4 happened on the main architecture: VideoCore 4 to 6. If Broadcom’s STB (or multimedia or whatever) division will ever do a VC6 on a smaller node, a future RPi will be on this node too with whatever guest processors Broadcom will be choosing.

            It’s all about availability of 3rd party’s chips. The 8GB RPi 4 appeared after Micron was able to ship 64Gb LPDDR4 modules. We’ll see a RPi 4 with 16GB once 128Gb LPDDR4 chips are available.

          4. Also, you don’t understand where Salvador’s deep disdain of RPi is coming from. But maybe he can explain it.

  4. I get scorned for saying it, but if you want affordable powerful CPU and GPU, then get more people playing games, on cheap handhelds, Smartphones and medium spec TV box / media player.

  5. I used to love rpi3, that was my first arm sbc, later on, if you are really interested on arm linux desktop not just retroconsoles like many rpi owners, you find out that rpi lacks of mostly everything you need. I/O, ram bandwidth, gpu, nothing is there. Rk3399 is the closest thing we have today while RK3588 is near to be unveiled.

    Also, rpi4 is quite not open source friendly thanks to ThreadX. And the 5v only thingy…. why, why not PD. It would resolve sooo many headaches.

    1. Let’s be honest, RPi4 is their first decent board. I seldom use mine to test code, and build times are bearable. The USB-C connector addressed the issues the micro-USB connector faced. The RAM and I/O bandwidths are low by they made a lot of progress since RPi3. And thanks to modern full-metal enclosures that serve as huge heatsinks, throttling is a thing of the past. I’m not saying the product is awesome, but that it works pretty well and is particularly inexpensive. For sure, when you compare it to an Armada8040 running 4xA72 at the same frequency, you understand that something is clearly missing in RPi. But for 1/5 of the price or so, that’s OK.

      RK3399 isn’t that good unfortunately. There are some bus arbitration issues between the A53 and A72 cores to reach L3 and RAM. A single A53 core cannot use the full L3 bandwidth, which indicates that they probably throttle them to work around some internal limitations, or maybe that the use aggressive power savings to limit thermal dissipation. I indeed noticed that an A53 core becomes faster when another one works in parallel!

      A 2GHz RPi4 gives me exactly the same build times as a stock RK3399 with 6 cores (2*1.8+4*1.4). In such conditions I still favor less but faster cores.

      Odroid N2+ with its S922X at stock frequency is 50% faster than RK3399 for me, which is far from being negligible, and it supports being massively overclocked (mine runs at 2.4 GHz). But for *some* use cases, the A73 is not as good as A72. YMMV.

      New ARM cores like A55 show impressive improvements on RAM bandwidth, sometimes 2x even compared to A72, and sometimes manage to significantly shorten the gap with the older A72. At least for this reason A76 will be interesting to have a look at.

      1. > The USB-C connector addressed the issues the micro-USB connector faced

        Way less contact resistance but especially if you want to power 5V consumers like USB disks by an RPi you still need to be aware that there’s also cable resistance (and their USB-charger a pretty good choice here – also rather inexpensive unlike anything USB PD).

        Wrt RK3399 vs. BCM2711 the latter is also more energy efficient than RK3399 fully loaded. Of course RPi 4’s idle consumption could be lower but I guess that’s a challenge when doing USB3 via PCIe.

        1. > their USB-charger a pretty good choice here – also rather inexpensive unlike anything USB PD

          I noticed this as well and even considered ordering a few more, before I remembered that buying several high-current A-to-C cables instead to connect them to my 5V PSU was even more convenient 🙂

          > Wrt RK3399 vs. BCM2711 the latter is also more energy efficient than RK3399 fully loaded

          That’s true, my NanoPI-M4 with roughly the same size heat sink (i.e. a thick 1.5cm one taking all the surface of the board) gets hotter than the RPi4 at 2 GHz.

          1. > NanoPI-M4 … gets hotter than the RPi4

            Since thermals can be misleading in several ways I’m currently in the process of adding support for Netio powermeters to sbc-bench. Manufactured in EU (Prague based) and meant to switch/measure bigger AC equipment it looks promising so far even with rather low wattages: https://transfer.sh/jIQNtm/odroid-hc2.pdf

          2. I completely agree with you on the RPi4.
            I’ve gotten myself two of those (at the same time as I got two 3B boards).
            I’ve only barely used the 3B boards.

            For power, I purchased a premium (low resistance) 3m USB-C-to-USB-C cable, cut it in two, and connected them to a MeanWell IRM-60-5ST PSU.
            This MeanWell was chosen because it can handle voltage spikes on the AC input (as well as voltage drops) while still delivering a ‘safe’ and stable output and at the same time comply with all EU regulations (including stricter danish regulations!).

            What I miss most of all, is the ability to connect some PCIe cards (even old PCI would have been acceptable).

            I also agree on the Solid-Run part. The MacchiatoBin, GT8K and CX LX2 are clearly products to consider. Today I’ve just asked them about the CX – if they have it, I’ll order one Tuesday.

      2. I am not saying that RK3399 is perfect neither, it’s just that it’s better than the pi4 on mostly any aspect. On cpu workloads they are similar in raw performance (some perform better, some worst), but rk3399 have several times more I/O, better gpu, better memory bandwidth, AES, vp8/vp9. And we are talking about an old chip. And it’s more open source friendly than rpi4. You can run it completely blobless if you want.

        Also, 5v only is the problem, not only the connector. The rockpi4, for example, can work from 5v to 20v (IIRC).

        If you want to build a nas, on rk3399 you have 4 pci 2.0 lanes and 2 usb 3.0 interfaces and AES. If you want to run a desktop you use those 4 pci lanes for an nvme. If you run a desktop you need that higher memory bandwidth bc inside X11 even my radxa zero kick rpi4 gpu performance (s905y2) due the short memory bandwidth that rpi4 have. Wayland is great but not mature enough.

        If you want just to test code, then it’s okay.
        And of course it’s much better than a crappy rpi3.

        1. > rk3399 have several times more I/O, better gpu, better memory bandwidth, AES, vp8/vp9

          For sure for I/O, extensions, and datasheet availability it’s day and night!

          Regarding a NAS, I wouldn’t do one with an RPi, but RK3399 definitely allows one to make one. But the RPi4 is an overall good machine for software developers.

      3. Odroid N2+ is a monster but still have severe problems.. memory bus, like rpi4 is 32 bit… inside x11 it performs poorly. It have even worse I/O than rpi4.

        If you just need a “dumb” monster, it’s great, but it’s still dumb.

        That’s why I like the radxa zero, amlogics, are only cool stuff for simple sbcs.. you can’t make anything complex with amlogic (except A311D2) unless you lie like on the odroid N2 and the C4, exposing 4 usb 3.0 when you only have bandwidth for 1.

        1. comparing raw numbers for memory bandwidth that’s about 2.5-3GB/s for S922’s (for a N2.x at blocks of about 10-500MB) and 4.1-4.3GB/s* for rpi4 (8GB ?), but mostly these SBC’s are cpu power limited and memory bandwith limits are not recognizable with consumer usage profiles (www, office tasks, media streaming, video i/o, ‘easy’ object recognition/inference interaction on data input, database analysis?, gaming?, server tasks or network routing? etc.) and that’s around 2002 amd64 (first announcements of amd64/x86-64) desktop level standards. But lacking specs for bcm2711 would make it a bigger guess what’s going on between VC and media output for newer ‘4kp60@10b is ~23Gb/s (~2GB/s)’ ‘gpu/vc – memory bus – display buffer’ datastreams and preferring specs available makes choice easier for Rockchip, Amlogic, ~Sunxi.Allwinner etc.?

          * https://forums.raspberrypi.com/viewtopic.php?t=281183&sid=453c169dc78bfaff0c99dc278e3bd1b6&start=50

          While above Mediatek Dimensity is on top 50-60GB/s (LPDDR5(X)) on memory bus.

          1. I think the numbers for S922 may just be for the Cortex-A53 cores.
            I get much higher numbers with the Cortex-A73 cores:

            Source: ODROID-N2+ review (Note, it’s clocked higher than most other S922X devices, but it still gives an idea)

          2. While thinking about memset values @7.5GB/s = 75Gbit/s = (2.64GT/s) = about 28.4bit/cpu_cycle (11% below theoretical 2x16bit memory interface). On AXI3 bus (about 400MHz, 90nm?) 4.3GB/s would require >=128bit/cycle bus width between cpu and memory controller (what is at least half of L1_i to L2 cache 256bit interface width)?
            (LPDDR5X related: https://developer.arm.com/ip-products/system-ip/corelink-interconnect/corelink-cache-coherent-interconnect-family/corelink-ci-700, 2013: introduction of AMBA5)

          3. AXI3 bus (about 400MHz, 90nm?)

            information fits for around 2010 SoC; while there’s lot of discussion about cpu/gpu core’s IP node sizes, interconnectors between IP modules evolved from 90nm, 65nm, 45nm, … to 2021’s newest 2nm (IBM), what adjusts core nodes and interconnector nodes on these newest developments (insulated copper, graphene not yet, <2030?). Interestingly there’s few public information about bus speeds for Advanced eXtensible Interface’s, AXI (version 3) or MMU units clock frequencies for (new) SoC’s. AXI frequencies might equal around half of default cpu core top frequencies (on clusters one’s directly interfacing the interconnect bus, on AXI3? AXI4 allows for individual clock frequencies between primary/secondary pairings). Numbers for MMU clock frequencies found were 1/4 of AXI bus frequencies (but sounds more like being individually adjusted for each SoC and IP combinations?).
            (just for update and correction)

          4. sorry for misunderstanding, again incorrect: interconnector is a term used for bridge connections between transistors on IP implementations (new 2nm example) and also for bus connections between logics ‘clusters’ assembled from IP modules for ARM-related SoCs (or between chiplets on x64 structures). These seem to be still bigger scale node elements on manufacturing process (stability, bridging gap width, materials, resistance) and still looking for how far below these are from 45nm era on todays packaging densities (~ >=30nm, AFAIK)? Anyone knowing some overview information, what structure sizes (nm scale) are included into a SoC design for different parts? Thanks.
            (just for sorting things into reality:
            256(bits/cycle on L1)*2GHz=~60GB/s for L1_d to LSU(?) bandwidth
            256(bits/cycle on L1)*2GHz/4(cycles for L1_d access, cache hit)=~14.9GB/s for L2 to L1 bandwidth
            fits for ~A72/A73 cores, pmbw for results, and Thanks@all )

          5. (G710 seems having sufficient bandwidth support with 32bit data interface, AFAIK
            4K ~8MP/1Hz
            8K ~30-50MP/1Hz)

    1. No way are they going to put out something that would have less single-threaded performance than Pi4. Unless you meant that it will not use Cortex-X1, which is a huge/super core.

      Most likely scenario is 4x Cortex-A75 on 28nm.

    2. I don’t think they’ll release the Pi 5 for a couple of years either, simply because the current Pi 4 model serves the purpose, and there’s no competition that vastly outperforms them yet. They’d also probably want to keep the $35 price tag for their new board, at least the entry-level version. That’s unless they go “PRO” like Arduino with more expensive boards for industrial/embedded applications.

      I’d be expecting new Raspberry Pi MCUs and boards first, probably with a small AI accelerator and/or built-in wireless (WiFi/BLE…).

      1. Indeed. I think rpi5 will be largely delayed. It makes sense, it will be hard to release an upgrade of the current model on 2022/2023.

        Rk3588 sbcs will be far lower price than expected, but still not as cheap as rpi4 2gb. And no one will release an rk3588 with 2gb of ram. Well, maybe someone does…
        Rk3588 will be a severe contender for rpi4, but mainline linux with panfrost, vpu,etc, will take years. We don’t have vpu inside the desktop on rk3399 out of the box today.

        Rpi4 is the first decent rpi board. But there is still a lot to polish from the software side, specially vpu on aarch64. And hopefully we will get better wayland support on more environments, rpi4 and mostly any arm linix board struggles on X11. On rpi4 is the worst case, glmark2-es2 performs more than 3.5 times better on wayland than on Xorg.

        And for the N2, no one shows results of glmark2 on x11 bc of the same issue.. the difference is drastic.

        On rk3399 it’s less affected, but I also noted a huge diff between DDR3 rk3399s against LPDDR4 ones, the first ones run at 1600 mhz, the second ones at 1866 mhz, and the gpu performance increase at same level (1:1)

      2. > I don’t think they’ll release the Pi 5 for a couple of years

        Me wouldn’t be surprised if an RPi 5 has 64-bit RISC-V guest processors (requires Broadcom going that route of course). This would solve their biggest problem (backwards compatibility with ARMv6/ARMv7) while getting nice press coverage about ‘innovation’ even when performing lower.

        If they figure out how to use the XHCI controller inside the BCM2711 maybe we see another RPi 4 with lower idle consumption and USB3 and PCIe usable in parallel? Who knows, who cares?

        1. That’s a wise guess, I tend to agree with you. It will please their fanboys to surf on the “even more open” hype, would save them from offering more performance, and would justify the initial performance losses by “it’s a new product, give it time to mature”. It could be much cheaper and make lots of adepts since day one.

  6. Candid questions. I guess this SoC aims at phones and tablets, but:

    1. Do we know the max amount of memory it could support?
    2. Do we know how many HSIO lanes it could provide?

    This in the context of SFF PC.

    1. 1.) It doesn’t look like they released that info anywhere. The minimum you could expect is 16 GB since that’s what many flagships have, but 18 GB is a possibility since there are a couple phones with that. 20-32 GB, who knows.

      MediaTek seems to keep their product lines separate. Like Dimensity in the phones but they are called Kompanio in laptops. They could make tweaks to other versions of the SoC.

      Nobody should expect a MediaTek mini PC.

  7. If MediaTek would add a few PCIe lines, then wow.
    (PCIe would also allow board manufacturers to add SATA and GbE/10GbE – even if there isn’t dedicated internal controllers for those peripherals).

Leave a Reply

Your email address will not be published.

Advertisement
Advertisement