5th Generation APU/NPU, a Massive ISP, and New 5G & WiFi

APU/NPU AI/ML Upgrades

Moving away from CPU and GPU, we find MediaTek’s new “APU”, or AI Processing Unit, or rather NPU as we more generically tend to call it nowadays. MediaTek always had one of the earliest in-house implementations of such an IP, and the Dimensity 9000 now implements the 5th generation of such an IP.

MediaTek promises a +400% performance and power efficiency improvement over the previous generation Dimensity 1200 implementation, which is quite a step-up, however should be contextualised to other more performant competitor platforms.

The company had made one slide available of showcasing MLPerf against Apple’s A15, noting +108% at +75% power efficiency versus the iPhone chip, the company here actually quoted our own published MLPerf figures, however we’ll avoid using the slide as the comparison isn’t great as the iOS variant of the app isn’t fully optimised and doesn’t take advantage of CoreML acceleration.

A more apples-to-apples comparison would be the ETHZ Ai Benchmark across Android devices, here the chip should take advantage of all accelerator blocks including CPU, GPU and NPU, and the Dimensity 9000 is advertised to beat the Google Pixel 6 and the Tensor SoC, which recently had notably outperformed the competition. MediaTek doing even better here seems to be promising for ML performance of the SoC. Naturally, we’ll have to see more detailed benchmarks for a more thorough analysis of the ML performance, but it seems MediaTek has developed a quite solid contender here.

Media Pipelines & Massive ISPs

Moving onto the media side of things, MediaTek also throws in everything but the kitchen sink into the new SoC. We’ll get to the ISP in a bit, but in terms of video encoding and decoding, MediaTek supports all popular current codecs. There is no AV1 encoding yet, but the company says that the chip is the first in the industry to support 8K AV1 decoding – previous generation Dimensity and competitor SoCs only are able to do 4K decoding at the moment.

The display pipeline supports WQHD+ to 144Hz, or FHD+ to 180Hz with full HDR+ Adaptive (10-bit), so the chip should power the highest resolution and refresh rate screens out there.

Getting back to the ISP, MediaTek claims that the camera subsystem is massively revamped in this generation. The Dimensity 9000 features triple ISPs with the ability for concurrent operation, with the chip being able to push through 9 Gigapixels/s. We’re not sure exactly if the figures are comparable across the vendor claims, but Qualcomm’s Snapdragon 888 “only” supports 2.7GPixel/s. Interestingly enough MediaTek says this is only a 2x throughput increase versus the Dimensity 1200.

The Imagiq790 ISP claims to be the world’s first to support 320MP sensors and MediaTek claims they’re working closely with the sensor vendors to enable such functionality. We’ve heard last year about 200MP sensors when Qualcomm announced the Snapdragon 888, and in September Samsung had announced the HP1 at this resolution, though we haven’t seen devices with it yet.

In triple-camera operation, the ISP supports concurrent 32+32+32MP sensor operation. We’ve seen such operation become more popular in sensor-fusion like scenarios for computation photography, or concurrent video recording on multiple sensors.

The new ISP also vastly improves its capabilities in terms of bit-depth, as MediaTek now upgrades it to a full 18-bit pipeline. Although mobile image sensors today natively are barely 10- or 12-bit at most in terms of their ADC bit-depth, multi-exposure image stacking has become the norm, especially with sensors now also getting more advanced HDR techniques such as staggered HDR captures. The new higher bit-depth ISP now is able to better do exposure merging across multiple frames, with the Dimensity 9000 able to do 270 frames per second (at presumably 4K resolution) across three sensors on its three ISPs.

The raw throughput and processing power here would be immense, and a massive leap over any other current SoC in the market.

MediaTek says that they’ve improved the video pipelines as well, being able to tightly interact with the APU in a memory coherent fashion, bypassing the need to copy data over DRAM, increasing performance and decreasing bandwidth requirements and latency. Presumably vendors would be able to take advantage of the architecture to implement video recording with ML-based image processing models on-the-fly, essentially the same kind of technology Google had presented on its recent Pixel 6 phones and the Google Tensor chip.

Upgraded 5G Modem

On the modem side, MediaTek has had quite a lot of success with its recent 5G modem implementations. The new Dimensity 9000 modem upgrades things even further, advertising for the first time 3CC carrier aggregation with up to 300MHz of Sub-6 bandwidth, allowing for download speeds of up to 7Gbps.

The modem is fully compatible with 3GPP release 16, with one larger change being UL Tx switching, which allows for better uplink capabilities and utilisation of spectrum in multi-band 5G NR deployments.

MediaTek notes that its modem has been extremely power efficient compared to competitor products, and the Dimensity 9000 will continue on this patch by providing advanced power saving technologies.

Although the modem features all the bells and whistles for Sub-6 5G, it does still lack mmWave. MediaTek notes that this is simply a result of the market need and the company’s current customer focus. Currently, the US remains the only market where mmWave truly remains a critical feature, as most vendors opt to not even equip their global device variants with mmWave modules. The company admits that due to this, we’ll unlikely see US devices powered by the Dimensity 9000, and the company is fine with that compromise, as it tries to cater to and focus on vendors which serve the rest of the world. MediaTek noted that next year, we will see mmWave product announcements for the US market, but these will at first be in the low/mid-range line-up.

Finally, on the Wi-Fi side, MediaTek’s Dimensity 9000 platform also comes with their own in-house solution, now supporting Wi-Fi 6E for 6GHz band support, and 160MHz channel bandwidth, as well as Bluetooth 5.3. The GNSS solution now adds BDS-3 connectivity via B1C signals.

TSMC N4, Maxed-out CPUs, LPDDR5X, Large GPU Conclusion & First Impressions
Comments Locked

150 Comments

View All Comments

  • syxbit - Thursday, November 18, 2021 - link

    >>The GPU side, the chip likely will be the only design for 2022 with a large Mali GPU

    Surely the Tensor 2 will be another large Mali ?
  • arayoflight - Friday, November 19, 2021 - link

    Pretty sure Google will use the AMD IP from S.LSI for Tensor 2.
  • dotjaz - Friday, November 19, 2021 - link

    Surely you know Google will use whatever Samsung is using.
  • UNLK A6 - Thursday, November 18, 2021 - link

    Mediatek slide decks are always a laugh. "World's first", "World's best." But please AnandTech don't publish their bogus slides without getting out your sharpie. Almost every "graph" is visually exaggerated. For example the first graph labeled +35% actually has a bar chart with 54% increase in height. Make fun of their charts and try to keep them honest.
  • Tomatotech - Saturday, November 20, 2021 - link

    You made me go back and actually count the pixels in the first two graphs. The bar charts are proportionally the correct heights for the numbers they represent. You've just wasted 5 minutes of my time and made yourself look a fool. Next time check your facts before making wild claims.

    Thanks for actually making MediaTek look more honest, which probably wasn't your aim.
  • mode_13h - Sunday, November 21, 2021 - link

    > You've just wasted 5 minutes of my time

    Thanks for that. I was inclined to believe the previous poster.
  • UNLK A6 - Sunday, November 21, 2021 - link

    See my comments below for clarity on who to believe.
  • UNLK A6 - Sunday, November 21, 2021 - link

    Well, perhaps you should check again. Not only is it obvious that the first bar is not a 35% lift from the second as claimed on the slide, if you count pixels you'll find it's a 54% lift.

    Perhaps you (like the chart maker) misunderstand that a 35% lift is not the same as a 35% drop. In this case the reciprocal of 1.54 is 0.65. The chart is simply wrong because either the chart maker is clueless or they are being deliberately misleading. A 35% lift is the same as a 26% drop. Clearly not what is depicted.

    I tend to lean on deliberately misleading because MediaTek does this year after year.
  • mode_13h - Monday, November 22, 2021 - link

    Thanks for the follow-up.
  • Wilco1 - Tuesday, November 23, 2021 - link

    You are wrong, the graph is correct - the first bar has 135% of the height of the 2nd bar, so it correctly shows the 35% performance uplift. I suggest you count the pixels rather than making things up.

Log in

Don't have an account? Sign up now