沙巴体育平台 www.yousuperb.com Today MediaTek announces a new series in its product line-up: The new G-series starting off with the G90 and higher binned G90T. The new chips seemingly are a marketing exercise for MediaTek as it tries to battle again Qualcomm’s newest Snapdragon 7xx devices. The new G90 on paper very much looks like an update to the P90 which was announced late last year – updating the CPU and GPU IP whilst also slightly improving the camera capabilities of the chip.

MediaTek Current P- & G-Series
SoC Helio P90 Helio G90
(Helio G90T)
CPU 2x Cortex A75 @ 2.2GHz
6x Cortex A55 @ 2.0GHz
2x Cortex A76 @ 2.0GHz
(2.05GHz)
6x Cortex A55 @ 2.0GHz
GPU PowerVR GM 9446 @ 970MHz Mali G76 MP4 @ 720MHz
(800MHz)
APU / NPU / AI Proc. / Neural IP 2x +140GMACs
(Tensilica DSP)

+ In-house Inference Engine
1127GMACs total
2x APU

+1TOPs total perf
Memory 2x 16bit LPDDR4X @ 1866MHz LPDDR4X @ 2133MHz
ISP/Camera 1x 48MP
or
2x 24+16MP
1x 48MP (64MP)
or
2x 24+16MP
Encode/
Decode
2160p30 H.264 & HEVC 2160p30 H.264 & HEVC
Integrated Modem Category 12/13

DL = 600Mbps
3x20MHz CA, 256-QAM, 4x4 MIMO

UL = 150Mbps
2x20MHz CA,64-QAM
Mfc. Process 12FFC

The new G90’s main feature update is the switch from Cortex A75 cores to new Cortex A76 based IP. The new cores are clocked in lower at 2.0GHz, which is 10% lower than the 2.2GHz of the P90. The higher binned variant, the G90T, ups the frequency slightly higher by 50MHz at up to 2.05GHz. The big cores are accompanied by the same core config as on the P90- 6 additional Cortex A55 cores running at up to 2.0GHz.

Another big change in the IP setup is that MediaTek is dropping Imagination’s 9XM core sin favour of a Mali G76. The new GPU comes in a MP4 configuration (MediaTek also likes to specify three execution engines per core), running at up to 720MHz in the G90 and 800MHz in the G90T. MediaTek promises 26% faster performance than its direct competitors, the competition here being Qualcomm’s Snapdragon 730.

Part of the gaming theme of the SoC, MediaTek promotes its new “HyperEngine” game technology. It looks here that this is a response to Huawei’s GPU Turbo or Qualcomm’s own variant – it’s a plethora of software optimisations that promise to improve the experience of the phone. The most important aspect for MediaTek here seems to be the promise of 60% shorter rendering latency.

In terms of the APU performance we don’t have immediate details of the G90’s, but it looks like things have remained relatively unchanged compared to the P90, with possibly Tensilica DSPs augmenting MediaTek’s own inference engine IP which is here stated to be able to operate at up to 1TOPs.

On the camera department, the regular G90 remains the same as the P90 with up to 48MP single-sensor ability or 24+16MP multi-sensor capture. The G90T clocks the ISP higher and promises compatibility with the newest 64MP sensors that have been announced by vendors such as Samsung.

The new chip continues to be manufactured on the 12nm FFC process and thus MediaTek should likely be able to price the chip extremely competitively against Qualcomm who has more advanced process node chips in this range.

Related Reading:

POST A COMMENT

14 Comments

View All Comments

  • FloconDeNeige - Tuesday, July 30, 2019 - link

    G76MP4? And this is being marketed for gamers? I know Mediatek is a value company, but this seems very low considering it is Mali (not reputed for good performance).
    Difficult for manufacturers to put this chip in phones which would be advertised for their gaming performance without risking disappointing their customers IMO.
    Reply
  • Plumplum - Tuesday, July 30, 2019 - link

    Hi,

    I calculate something...
    As ARM double the number of lanes between G72 and G76, G76 is at least twice better
    If you take GFxBench Manhattan 3.0 test, you will get :
    20 fps on Helio P60 (1920x1080) with Mali G72mp3 800mhz
    On G90 with G76mp4 720mhz, you can expect around :
    20x4x2/3x720/800=48fps

    On Snapdragon 730, you get 38fps on this test...
    As Qualcomm say their 730G is 25% better than 730, you can expect around...48fps

    What is sad is that Mediatek keeps 12nm process...even if a 720mhz GPU may be stable
    Reply
  • FloconDeNeige - Tuesday, July 30, 2019 - link

    That's not how that works.
    Firstly, with the doubling the quad size in the G76 compared to the G72, the theoetical bump in performance is not at least twice better, but exactly twice as good (from the article in Anandtech about it, the scaling is supposed to be very good, so let's say ~2).
    Secondly you cannot compute the performance like that, no GPU scales linearly with execution units and frequency. There are many other considerations, such as memory system, power/thermal envelope, variations from arch to arch, CPU...
    For example, the Exynos 9820 in the S10 get about 102fps in Manhattan 3.0 even though it has triple the quads, and about the same frequency (with your computation, it should have reached around 140fps).
    Reply
  • Plumplum - Tuesday, July 30, 2019 - link

    I think memory won't limit G90 as it is 2x16bits lpddr4x 2133mhz...
    I think even 730G with slower memory won't be limited

    We're already agree to say G76 is twice better than G72...
    We're agree too on execution units...when you compare mp4 to mp12...but here, we compare mp3 to mp4, you won't loose that much.

    About frequency, I'm disagree...you can compare for exemple Intel Z2520, Z2560 and Z2580 equiped with PowerVR544...framerate increase nearly linearly with frequency on OpenGLES...
    You can even cross OS and find similar behavior with Apple A5 with PowerVR543 (nearly the same as 544).

    If I remember monitoring session on Android 6, OpenGLES didn't use CPU that much on my old Helio X20 and mt6752...I can't verify nos as monitoring no longer works on newer Android versions.

    I don't say my numbers are exact, I use the word "around", but I think they aren't far from reallity.

    The interrogation may come from 12nm and thermal...
    I'm confident with G76 running at 720mhz
    A little less at 800Mhz...

    Plus I'm agree with PeachNCream
    Reply
  • PeachNCream - Tuesday, July 30, 2019 - link

    I'm going to go out on a limb and say that there are not many people that have done the research or fully understand mobile SoC performance, much less even care about it. You buy a phone and whatever it happens to run well in the way of games, you end up playing. Not many people own enough handsets to compare performance so I doubt there will be much disappointment out there either. Reply
  • regsEx - Wednesday, July 31, 2019 - link

    Yeah, MediaTek keep stepping on same rakes again and again. Reply
  • ToTTenTranz - Tuesday, July 30, 2019 - link

    MediaTek really can't help but use anemic GPUs. Even for their so-called "Gaming" SoC they're using a ~150 GFLOPs GPU. Is this even going to be substantially better than the 124 GFLOPs PowerVR in P90?
    I can't see that in the table, but I bet they're using 32bit LPDDR4 too.

    Why not just call this P95 or P100?
    Reply
  • Plumplum - Tuesday, July 30, 2019 - link

    First, I don't know where you find your 150GFlops...
    If you get G72's, it's absolutly false as G76 has twice the instructions per execution unit compare to G72...get the numbers from a G72mp4 is false, it's more like a G72mp8

    Second, GFlops means nothing when you compare Mali to Adreno. One exemple :
    https://gfxbench.com/compare.jsp?benchmark=gfx50&a...

    MaliT880mp12 and Adreno 530 gets close framerate on many tests.

    MaliT880mp12 : 265GFlops
    Adreno 530 : 498GFlops

    My advice : forget GFlops
    Reply
  • Wardrive86 - Tuesday, July 30, 2019 - link

    Dont trust the obline gflops you find if we are counting Fmadds the Adreno 530 has 256 ALUs @ 710 mhz = 364 Gflops. And your right Compute isnt everything (if its texture, pixel or geometry bound) but newer workloads tend to be more compute bound so it will mean more as we continue going forward Reply
  • Plumplum - Wednesday, July 31, 2019 - link

    He probably get the online one...I don't trust.
    What newer workloads are you talking about? AI?
    Even these ones...when Qualcomm uses Hexagon and Adreno, others prefer using dedicated NPU
    Reply

Log in

Don't have an account? Sign up now