MediaTek Announces Flagship Dimensity 9000 5G SoC on TSMC 4 nm

[TheLostSwede]

A few years ago, MediaTek decided to leave the high-end mobile device SoC space as it decided to focus on the mid-range and entry level market, but as of today, that is no longer the case. The company has announced its new flagship Dimensity 9000 SoC and it packs all the latest cutting edge mobile SoC features you'd expect to see.

The Dimensity 9000 is based entirely on the new Arm v9 architecture and the chip itself is built in TSMC's 4 nm node, which gives MediaTek a lead on all its competitors, but Apple. As for the Arm cores in question, the main core is a Cortex-X2 at 3.05 GHz, which is accompanied by three Cortex-A710 cores at 2.85 GHz and four power efficient Cortex-A510 cores at 1.8 GHz. All of these new Arm cores were only announced at the end of May this year, so MediaTek has clearly been burning the midnight oil to get the Dimensity 9000 out before its competitors' new chips based on the Arm v9 architecture.




The GPU in the Dimensity 9000 is also from Arm, in the shape of the Mali-G710 in an MP10 configuration, which means it has 10 GPU cores. Arm claims a performance improvement of around 20 percent compared to the older Mali-G78 GPU core, so it's unlikely that the Dimensity 9000 will offer the fastest mobile GPU in the market for long, but MediaTek claims its 35 percent faster than other Android flagship phones currently in the market, without going into any details.



When it comes to display support, MediaTek seems to have been a little bit conservative, as the highest official resolution supported seems to be WQHD+ or 2960x1440 pixels, albeit at 144 Hz, so it's possible that higher resolutions are supported at lower refresh rates. On the other hand, FHD+ or 2220x1080 would enable 180 Hz support, so we might see some gaming phones based on the Dimensity 9000.

The Dimensity 9000 is also the first mobile phone SoC to support LPDDR5X memory and it does so up to 3,750 MHz or 60 GB/s memory throughput. MediaTek has also outfitted the Dimensity 9000 with 6 MB of system level cache, which is twice that of the Snapdragon 888 series, but 2 MB less than both Samsung's Exynos 2100 and Google's Tensor SoCs. However, the shared L3 CPU cache is twice the previously mentioned mobile SoCs at 8 MB.

Other interesting features of the Dimensity 9000 includes the new Imagiq790 18-bit camera HDR-ISP which is capable of supporting 320 Megapixel sensors, or three 32 Megapixel sensors simultaneously for computational photography. There's also what MediaTek calls a 5th generation, six core APU or Ai processing unit, also known as an NPU to further help with computational photography and it's said to be 16 percent faster than the Google's Tensor chip in some Ai workloads.



When it comes to video playback, the Dimensity 9000 is pretty much as cutting edge as it gets, with 8K30p AV1 decoding support, which MediaTek claims as as a first, as well as 8K60p decoding and 8K30p encoding support for H.265/HEVC, H.264 and VP9. However, it's unclear if the Dimensity 9000 can output these types of resolutions to an external display, making support for such resolutions somewhat pointless.

5G support is pretty much given on a high-end mobile SoC these days and although MediaTek didn't implement support for mmWave in the Dimensity 9000, it does support 3GPP release 16, also known as 5G R16 which enables download speeds of up to 7 Gbps on supported networks. There's also support for what is known as UL Tx Switching, which means upload speeds will be faster and MediaTek has tested upload speeds of up to 3.2 Gbps on a Nokia based 5G base station.

Other wireless features include WiFi 6E, Bluetooth 5.3 support - which apparently is another world's first in a phone SoC - and Bluetooth LE Audio support for Dual-Link True Wireless stereo audio.

Overall this looks like a very competent mobile SoC and so far it seems like several major mobile phone makers are onboard, so expect devices from Motorola, OnePlus, OPPO, Samsung and Xiaomi, with the first one arriving as early as Q1 next year. Until then we won't really know how the Dimensity 9000 stands up to the competition and although it looks impressive compared to what's currently in the market, we still don't know the full details of what Qualcomm has coming next, nor Samsung or Google for that matter.

View at TechPowerUp Main Site

 

[zlobby]

'Demensity' typo? Seems quite in order to me!

 

[Valantar]

Wow, that is ballsy. Not bad. Also, a process advantage on everyone but Apple? Isn't Apple firmly on (a semi-custom) 5nm?

Either way, looking forward to seeing how these perform. Hope the implementations of the cores are good, as the later generations of SoCs from Samsung and the like have shown just how important this is.

 

[TheLostSwede]

'Demensity' typo? Seems quite in order to me!

Ooops, fixed.

Wow, that is ballsy. Not bad. Also, a process advantage on everyone but Apple? Isn't Apple firmly on (a semi-custom) 5nm?

Either way, looking forward to seeing how these perform. Hope the implementations of the cores are good, as the later generations of SoCs from Samsung and the like have shown just how important this is.

From my understanding, Apple is always on the most cutting edge node from TSMC, even though it might not officially have a name.
This 4 nm node is technically an improved 5 nm node anyhow, so...

It'll be interesting to see how the new Arm cores perform, since all the cores a new, rather than just minor improvements or updates.

 

[Valantar]

From my understanding, Apple is always on the most cutting edge node from TSMC, even though it might not officially have a name.
This 4 nm node is technically an improved 5 nm node anyhow, so...

Yeah, that's probably true. AFAIK the node Apple is currently on doesn't actually have an official name.

It'll be interesting to see how the new Arm cores perform, since all the cores a new, rather than just minor improvements or updates.

Yep, this will be very interesting. IMO the A510 is possibly the most interesting, just because the A53 is so woefully out of date by today's standards. The X2 will also be really interesting of course, though I have less hopes there, I'm expecting more of an "X1 plus a bit" rather than an M1 Firestorm competitor. I'd be happy to be wrong though!

 

[TheLostSwede]

Yep, this will be very interesting. IMO the A510 is possibly the most interesting, just because the A53 is so woefully out of date by today's standards. The X2 will also be really interesting of course, though I have less hopes there, I'm expecting more of an "X1 plus a bit" rather than an M1 Firestorm competitor. I'd be happy to be wrong though!

Well, there was the A55 as well... But it was such a seemingly minor improvement that it didn't matter much.
The X2 doesn't look like a huge improvement in general compute performance, so it and the A710 are more generational improvements than anything, whereas the A510 is supposed to have a 35 percent performance improvement. Not sure how much any of the ML stuff will matter to what most people use their devices for.

 

[Valantar]

Well, there was the A55 as well... But it was such a seemingly minor improvement that it didn't matter much.
The X2 doesn't look like a huge improvement in general compute performance, so it and the A710 are more generational improvements than anything, whereas the A510 is supposed to have a 35 percent performance improvement. Not sure how much any of the ML stuff will matter to what most people use their devices for.

Lol, I forgot the A55. Has anyone really used that except Qualcomm in their little Kryo cores? I might just have a mental filter that reads A55 and A53 as the same A 35% uplift doesn't sound all that impressive, sadly. And I entirely agree that ML performance needs actual practical and useful applications before I'll care about it. Still, this all depends a lot on the node, caches, interconnects, power/clock gating, and other aspects of the specific implementations that appear. Looking forward to seeing this in some actual devices. Hopefully QC doesn't bribe everyone away from making use of these.

 

[TheLostSwede]

Lol, I forgot the A55. Has anyone really used that except Qualcomm in their little Kryo cores? I might just have a mental filter that reads A55 and A53 as the same A 35% uplift doesn't sound all that impressive, sadly. And I entirely agree that ML performance needs actual practical and useful applications before I'll care about it. Still, this all depends a lot on the node, caches, interconnects, power/clock gating, and other aspects of the specific implementations that appear. Looking forward to seeing this in some actual devices. Hopefully QC doesn't bribe everyone away from making use of these.

Yeah, there are several Cortex-A55 SoCs in the market now, as both Amlogic and Rockchip has launched chips based on it.
I think older MediaTek phone chips uses them too, as well as Samsung and of course the Google Tensor chip, so it's fairly common in new designs.
However, clock for clock it doesn't seem to be a lot faster, but it is a four year old design, so it shouldn't come as a surprise that it's in a lot of products, it's just that no-one noticed the difference.

35 percent isn't enough on the power efficient cores for you? Daim!

As we've seen, especially from Apple, it is very much about the overall chip design, since although most Arm based devices out there share a common architecture, they clearly don't perform the same, even at similar clock speeds. Obviously a lot of that comes down to software optimisation as well and this is where MediaTek has a bit to prove.

 

[Space Lynx]

Wasn't 2001 A Space Odyssey computer named something or other the 9000? Or am I thinking of something else? Either way I am getting Skynet vibes from that name for some reason... but I don't know why!

 

[noel_fs]

MediaTek claims its 35 percent faster than other Android flagship phones currently in the market, without going into any details.

nice joke, mali doesnt compete with adreno

Lol, I forgot the A55. Has anyone really used that except Qualcomm in their little Kryo cores? I might just have a mental filter that reads A55 and A53 as the same A 35% uplift doesn't sound all that impressive, sadly. And I entirely agree that ML performance needs actual practical and useful applications before I'll care about it. Still, this all depends a lot on the node, caches, interconnects, power/clock gating, and other aspects of the specific implementations that appear. Looking forward to seeing this in some actual devices. Hopefully QC doesn't bribe everyone away from making use of these.

aside iphones i dont think there is a phone that doesnt use A55 that isnt extremely budget.



sidenote, im not touching a mediatek, not a fan of getting stuck with ancient android versions

 

[Valantar]

aside iphones i dont think there is a phone that doesnt use A55 that isnt extremely budget.

Literally every single high end ARM SoC on the market has some form of "little" low power core. And their performance matters.

35 percent isn't enough on the power efficient cores for you? Daim!

Not on a core that lags the competition by as much as this does and is that old, no. A 35% increase in that long is Intel in the mid-2010s levels of annual performance gain, if not worse. Not impressed, no.

 

[noel_fs]

Literally every single high end ARM SoC on the market has some form of "little" low power core. And their performance matters.

?????

yeah, and ever single little cluster uses A55

the performance is pretty good, im actually interested to see hows the new a510

 

[Valantar]

?????

yeah, and ever single little cluster uses A55

the performance is pretty good, im actually interested to see hows the new a510

Did you miss the entire discussion above? Whether it's A53 or A55 doesn't matter much, as they perform pretty much the same. Both are old, slow, and underwhelming even in efficiency by today's standards.

 

[zlobby]

Wasn't 2001 A Space Odyssey computer named something or other the 9000? Or am I thinking of something else? Either way I am getting Skynet vibes from that name for some reason... but I don't know why!

HAL, anyone?

 

[watzupken]

I used to feel that Mediatek's SOCs are pretty good value for its performance. However, I never quite like using their SOC because performance aside, some apps don't work well with their SOC. For instance, there are GPS apps that don't work with Mediatek SOCs. I am not sure about now, but I do hope that the compatibility issues are largely ironed out.

 

[roberto888]

sidenote, im not touching a mediatek, not a fan of getting stuck with ancient android versions

Hi, from my Realme 6 with Helio G90T and Android 12 custom ROM

 

[Pan]

Promising specs and performance claims. Always good with more competition in the high performance mobile SoC's market and not just in the lower price segment.

 

[R0H1T]

Lol, I forgot the A55. Has anyone really used that except Qualcomm in their little Kryo cores? I might just have a mental filter that reads A55 and A53 as the same A 35% uplift doesn't sound all that impressive, sadly. And I entirely agree that ML performance needs actual practical and useful applications before I'll care about it. Still, this all depends a lot on the node, caches, interconnects, power/clock gating, and other aspects of the specific implementations that appear. Looking forward to seeing this in some actual devices. Hopefully QC doesn't bribe everyone away from making use of these.

Yes that's just you, Mediatek/Unisoc/Allwinner/Samsung/Amlogic all have used A55 cores in multiple SoC's by now!

 

[Valantar]

Yes that's just you, Mediatek/Unisoc/Allwinner/Samsung/Amlogic all have used A55 cores in multiple SoC's by now!

Well, lesson learned I guess. Sadly it doesn't change the fact that these cores are rather terrible in terms of both efficiency and performance - look at the summary graph in AnandTech's Google Tensor review, for example. While the Tensor's implementation seems quite poor, it performs in line with other A55s, and all of them score less than half of the efficiency cores in the Apple A14, let alone the A15. A 35% uplift from that would make them score half of the Icestorm A15 cores instead, I guess? That the A55s also consume more power than the Apple cores under the same workload - up to 5x as much, despite doing less total work (scoring lower)! - is especially damning. Even if this should turn out to be a near ideal set of workloads for Apple (which is rather unlikely), that difference is quite something.

 

[Nike_486DX]

Snap 888 would probably still be better cuz of its adreno gpu and better app optimization

 

[Valantar]

Snap 888 would probably still be better cuz of its adreno gpu and better app optimization

App optimization is a bit of an unknown that might push things in this direction, but given how overall underwhelming the 888 is (at least in terms of power draw and thus also sustained performance and throttling characteristics) MTK do have a pretty good opportunity here. Still, this will mostly be competing with the next generation of Snapdragons, so that'll be a more interesting comparison for sure.

 

[looniam]

HAL, anyone?

funny enough . .apple used hal to party like 1999!

 

[Deleted member 24505]

Snap 888 would probably still be better cuz of its adreno gpu and better app optimization

Imo any snap is better than MediaTek, which is mainly used in cheap phones and tablets. I would never buy any phone with a MediaTek in it. This new Dimensity 9000 might actually be something, as seemingly high end manufacturers are having a sniff.

 

[R0H1T]

Imo any snap is better than MediaTek, which is mainly used in cheap phones and tablets.

No they're not, this is like the thinking from 2 decades back that any Intel is better than AMD

Mediatek usually delivers equally good perf/$ in most cases & sometimes much better in the mid to higher end range. Their dimensity 1200 is on par or better than SD870, depending on the benchmark, & is cheaper. They're also powering flagship tablets using Kompanio 1300 SoC.

 

[Deleted member 24505]

No they're not, this is like the thinking from 2 decades back that any Intel is better than AMD

Mediatek usually delivers equally good perf/$ in most cases & sometimes much better in the mid to higher end range. Their dimensity 1200 is on par or better than SD870, depending on the benchmark, & is cheaper. They're also powering flagship tablets using Kompanio 1300 SoC.

Well i am speaking from what i have seen, maybe you're in a different dimension