If you’ve been keeping track of the evolution of mobile graphics in the last decade, you might have noticed several trends that have changed the industry to the core. Probably the most visible and often talked about is the rapid increase in performance and feature set. It is becoming clear that high-end mobile GPUs are delivering more innovative technologies that surpass desktop PCs and consoles when it comes to the overall efficiency (i.e. performance per mW) of the design.

However, there is another important trend revolutionizing the mobile graphics industry. At the start of this decade, a new family of GPUs was introduced every two to three years, followed usually by a minor update in between; these designs were reused in different configurations across a range of devices. Now we have reached the point where the accelerated pace of innovation in mobile graphics means we deliver yearly significant scaling in performance as well as feature set. This accelerated release schedule is also designed to coincide with the progress made by APIs and Moore’s Law

We are constantly pushing the boundaries of what’s capable in mobile.

Introducing the new PowerVR Series7XT GPUs

This brings us to our big announcement today. With our Imagination Summit taking place this week in China, we are unveiling two brand new PowerVR families.

This article focuses on the PowerVR Series7XT family, while a second blog post covers the PowerVR Series7XE line-up. Read on after the break to find out how we have improved performance, reduced power consumption and added market-leading features to create the ultimate GPU roadmap for next-generation flagship devices.

PowerVR Series7XT GPUAll PowerVR Series7XT GPUs are based on the same underlying architecture:

The PowerVR Series7XT family provides an important feature upgrade over Series6XT, bringing new levels of performance to apps, games and GPU compute applications.

PowerVR Series7XT GPUs offer the right mix of performance and features, offering a solution that hits the sweet spot for high-end to mainstream platforms targeting Android 5.0 Lollipop and AEP (Android Extension Pack) support. The main highlights include:

  • Industry-leading performance: PowerVR Series7XT GPUs deliver unrivalled performance per mW. For example, PowerVR GT7900 includes 16 clusters and 512 ALU cores, delivering more than 1 TFLOPS graphics and compute performance in FP16 mode.

PowerVR Series7XT GPUs are up to 60% faster clock for clock, cluster for cluster compared to how Series6XT counterparts perform in graphics benchmarks.

  • Improved overall efficiency: our new co-issue mechanism makes better use of existing resources and delivers significant improvements in PPA (performance, power, area).
  • Virtualization support, a key enabler for improving overall system security and robustness in mobile, automotive and other embedded applications
  • Hardware tessellation: we’ve added tessellation hardware to ensure users have access to an unbeatable PC-class gaming experience on AEP (Android Extension Pack) devices. On top of that, we continue to support geometry and compute shaders as well as ASTC LDR and HDR texture compression formats
  • Optional DirectX 11 support: we deliver the best PPA support for desktop PC-class devices running Microsoft Windows
  • Optional HPC support for High Performance Compute (HPC) and supercomputing applications which require a variable amount of FP64 and OpenCL™ 1.2 Full Profile capabilities.

All these new added hardware capabilities ensure we have the right functionality required to support the latest feature set of the most popular graphics and compute APIs, including OpenGL® ES 3.1 + AEP, OpenGL® 4.4, OpenCL 1.2 FP, RenderScript, and DirectX 11.

PowerVR Series7 - GFXBench 4_0-2GFXBench 4.0 from Kishonti is a next-generation graphics benchmark

Desktop-class performance

The performance increase mentioned above is enabled by our versatile scaling capabilities; the new GPUs scale from two to sixteen shading clusters, giving developers the extra ALU horsepower they need to achieve more realistic effects, more complex pixel shaders, and console-quality lighting in games.

  • PowerVR GT7200: 2 shading clusters, 64 ALU cores
  • PowerVR GT7400: 4 shading clusters, 128 ALU cores
  • PowerVR GT7600: 6 shading clusters, 192 ALU cores
  • PowerVR GT7800: 8 shading clusters, 256 ALU cores
  • PowerVR GT7900: 16 shading clusters, 512 ALU cores

This unprecedented level of performance offers developers the chance to implement console-quality effects like HDR rendering, realistic 4K texturing, physically-based shading in mobile applications. The animation below illustrates the level of realism that PowerVR Rogue GPUs offer for mobile graphics.

OpenGL ES 3.0 demo for PowerVR Series7XTConsole-quality rendering on PowerVR Rogue GPUs

PowerVR Series7XT USC internal architecture

We have made several changes to our USCs (Unified Shading Clusters) which improve performance and deliver better efficiency. The internal architecture of the Series7XT USC is presented in the diagram below:

PowerVR Series7 - Series7XT_USCThe internal architecture of PowerVR Series7XT USCs

While the number of FP32 and FP16 slots remains unchanged from Series6XT, we’ve added an optional FP64 ALU for every pipeline to ensure developers have access to every type of precision and computing resources when needed:

  • FP16 is still widely used in mobile applications, offering sufficient accuracy at lower power.
  • FP32 ALUs can be used when a higher rendering precision is required
  • FP64 can be added to designs which are used in high-end supercomputing-style applications.

You can find the full specs for every member of the Series7XT family here.

Virtualization for GPUs

Although you usually tend to hear people talking about virtualization in the context of CPUs, it is actually a system-level requirement that can only be implemented optimally if all the components in the chip support it.

PowerVR Series7XT and Series7XE GPUs implement a set of extensions to the Rogue architecture that enable efficient hardware support for virtualization. These virtual GPUs can be added to virtual machines running on hypervisors. A priority based mechanism ensures each virtual machine efficiently gets the required amount of performance, ensuring robust performance across all virtual clients.

PowerVR Series7 - virtualization_zonesImplementing secure virtualization at the SoC level

One of the important features of virtualization is isolation — the foundation of an improved security framework for mobile and embedded devices. Since virtual machines are isolated from each other and from the underlying physical architecture by default, they deliver the security principles needed by device manufacturers in mobile, automotive and other embedded markets. For example, software developers can implement better secure mobile payments in games or prevent malware attacks when consumers are interacting with the UI.

What is tessellation?

You’ve recently heard a lot about AEP (Android Extension Pack), a new optional feature set that Google has introduced in Android 5.0 Lollipop. This collection of extensions is a superset of OpenGL ES 3.1 and contains several extensions that provide native support for advanced graphics functionality in mobile devices.

Tessellation is one of the new features of the AEP and offers developers the possibility to create scalable artwork and smoother looking characters or to seamlessly adjust the level of detail in large, open environments. The animation below shows how tessellation can be used for terrain tessellation effects:

07-Tessellation demo for PowerVR Series7 Tessellation effects running on PowerVR Rogue GPUs

PowerVR Series7XT GPUs support dedicated hardware that handles tessellation in fixed function logic which is far more power efficient than competitive solutions running excessive amounts of shader code.

Final words

PowerVR Series7XT GPUs are ideally suited for a wide range of mid-range and high-end devices where ultimate performance and efficiency are essential. Devices where you’ll likely see these new graphics processors include mid-range to flagship smartphones, premium to high-end tablets, gaming consoles, smart cars, notebooks and other personal computing platforms.

The PowerVR Series7XT family comes less than one year after Series6XT was announced at CES 2014, and roughly a year after the first devices using Series6 GPUs were in the hands of consumers.

Imagination continues to set the pace of innovation in mobile, creating a new benchmark for the graphics industry with every PowerVR launch. We are very excited to see Series7XT and Series7XE GPUs appearing in devices later next year.

For more news and announcements related to PowerVR, keep coming back to our blog and follow us on Twitter (@ImaginationPR, @PowerVRInsider, @GPUCompute).

 

Editor’s Note

* PowerVR Series7XT GPUs are based on published Khronos specifications, and are expected to pass the Khronos Conformance Testing Process. Previous generation PowerVR Rogue GPUs have already achieved conformance. Current conformance status can be found at www.khronos.org/conformance.

OpenCL and the OpenCL logo are trademarks of Apple Inc. used by permission by Khronos.

OpenGL is a registered trademark and the OpenGL ES logo is a trademark of Silicon Graphics Inc. used by permission by Khronos.

Comments

  • Viraf Kohli

    Next year I can guess on one fruitful tablet use case. So how does series 7 coexist with your Raytracing IP? Can we expect some announcements in that direction?

    • The GPUs we’ve announced today are based on the traditional TBDR architecture. Some exciting news on ray tracing will be coming soon, as promised.

      Regards,
      Alex.

  • regs

    So GT7900 is the one that is said to be 1,5 TFLOPs? So will be more powerful than 750Ti and R7 250X? Why not then to double it then, to make affordable desktop gaming card?

    • Hi,

      We can actually scale beyond that – including going multicore. The press release says that other configurations are available and customers can add any feature pack they wish, including desktop-class options such as DirectX 11_1.

      Regards,
      Alex.

      • regs

        Does it mean we will see some reference design card?

        • We usually have test chips for Rogue GPUs; we use them for driver development until platforms from customers become available.

          • regs

            Yeah, that’s for intergrated solutions. For discrete solutions it’s more common now to have a reference design, so manufacturers can easily copy it.

          • Yes, but since we don’t make chips, we don’t need to build reference designs.

            It is usually our partners (companies such as Intel, Allwinner, MediaTek) that produce reference designs; for example, Intel was showing a smartphone prototype using Atom at MWC 2014 while Allwinner was demonstrating a development board.

  • Renato Roškar

    So, the real big quastion here is, are you going to partnership with more smartphone manufacturers than you are now? If your 7XE Series are realy going to be that much power capable than it would be the best choice for other brands like LG, HTC, Sony etc. to go for those chipsets and not for Qualcomms… and when are we going to see the GPUs rolling out next year? Q3, Q4? I just want to stay on “hard ground” to be real, Q1 and Q2 next year would be to early and for Q3 and Q4 would I say “maybe”?! I am sure you have some realy nice answers. Have a nice day 🙂

    • Hi Renato,

      We are targeting a wide range of markets with the new Series7 GPUs. This includes smartphones too but also tablets, smart TVs, automotive, digital TV/set-top box and even desktop-class or HPC applications.

      However, we don’t design chips like Qualcomm or NVIDIA. Instead, we license our designs to silicon vendors like Allwinner, Intel, Texas Instruments, MediaTek and others and they produce the SoC. Given that it usually takes between 12-18 months between a customer licensing a PowerVR GPU and taping out, I would assume you would start to see products using Series7 GPUs later next year.

      Regards,
      Alex.

      • Renato Roškar

        Thank you for the given information.

      • kmare

        It really sounds great. Are there any plans on open sourcing or releasing open source drivers? Maybe helping out with the specs?

        Thank you for any information.

        • The open source discussion is a complicated one and I think I’ve replied to it a few times.

          We are not a silicon vendor and we do not control the manufacturing chain directly therefore pushing out drivers onto platforms that are not ours poses a lot of legal implications.

          Furthermore, even if we would make the source code of our Android driver public there’s no easy way for you to use that to replace the
          driver on your device.

          This is because companies have to customize the memory allocator and display controller integration parts on Android so you’d need to get your SoC vendor to give you those parts
          too, then you’d need privileged access to push it to the device. Other companies choose to implement their own driver.

          Regards,
          Alex.

          • Mitch

            I don’t think that many people even care about open sourcing the code for your drivers – I personally understand the pains of having to open source a code base that has license and IP restricted material in it, and it’s entirely understandable.

            All people want is for you guys to publicly document your GPU registers, ALU opcodes, and fixed-function command stream opcodes (if such a concept exists in your architecture – but you get the idea), without the need for signing NDAs to get access to your DDK (which then actually prevents people from doing anything useful with said doco).

            This is no different to how every single CPU architecture out there works with their TRMs. Could you imagine a world where ARM/MIPS/Intel MCU and MPUs, and SoC’s containing them had no TRM at all, or worse, were like imgtec and stopped people from using their documentation for practical purposes?

            We can and will happily do the rest for you, since you guys have shown for almost a decade now you have no interest in doing it yourselves.

      • LDM

        Hi Alex, you forgot to mention you guys license also for the big A ;)..

  • LDM

    Hi Alex, in Anand website it’s mentioned a powervr series 6 with 8 clusters. Is it a new model? I remember the x6650 is a six cluster GPU, isn’t ?
    L

    • PowerVR Series6 and Series6XT families could scale up to 8 clusters while the new Series7XT family goes up to 16 USCs.

      For Series6XT we’ve announced GX6650, one of the many available configurations. This indeed is a six cluster part.

      Regards,
      Alex.

      • LDM

        Thanks. Does it mean sound will be announced an official GX66 with 8 clusters? Indeed the six cluster in the A8X is already a beast; the 8 cores will be incredible!! And you guys have announced a series 7 with 16USCs!!

        • The publicly announced configs for Series6XT are GX6240, GX6250, GX6450 and GX6650 but the family can scale to eight clusters.

          • LDM

            Alex, according to Anand, it turned out Apple uses a powervr series 6 with 8 cluster configuration in their proprietary GPU A8X!

          • Roninja

            Seems like the family did just that…

  • #?

    I had been waiting on this announcement before 6XT was revealed. I just knew you guys where much further ahead than you were letting on. Nothing gets me excited like the stuff you guys are doing. 7XT and MIPS in iPhone, iPad and Macs next year.. I’ve known it was coming since 2012, even skipped several product generations patiently waiting. Very impressive.

  • dr.one

    1. No mention of H265 or its performance.
    2. Demphasis on OpenCL especially 2.0 on mobile,
    is it because Metal is all that is needed.
    Why wouldn’t you suport OpenCL 2.0 for HPC scientific applications.

    • Hi,

      I’m not exactly sure what you are referring to. H.265 is a video codec, what we are announcing today is a lineup of GPUs – they are two separate, albeit somewhat connected things. Video codecs (encoding and decoding) are handled by our PowerVR Series5 VPU (video processing unit) family; you can read more here:

      https://www.imgtec.com/blog/powervr-video/powervr-d5500-decoding-hevc-in-10-bit-colours-at-4k-resolutions

      It is true that you can use the GPU to offload some of the computational elements of the video encoding/decoding pipeline (instead of using the CPU) but having dedicated hardware that is designed specifically for this task is usually the optimum solution.

      We actually showed our support for OpenCL 2.0 when it was initially introduced:

      https://www.khronos.org/news/press/khronos-finalizes-opencl-2.0-specification-for-heterogeneous-computing

      Regards,
      Alex.

      • dr.one

        Intel is said to use (rumor) OpenCL for Broadwell to support H265 while the dedicated chip block arrives in Skylake.

        Your link is 2013 post only mentions decoding.
        You know that everyone is waiting for encoders
        to be ready now otherwise it is another 2 years of wait.

        You may be mentioned OpenCL 2.0 with Khronos PR
        but now you only mention 1.2 while AMD and Intel
        have beta OpenCL 2.0 drivers available.

        May be it is just a driver issue but the message it sends seem ominous.

        • Anybody

          He’s provided links to dedicated encoding and decoding IP from Imagination and that’s what the wide majority would use in a SoC in order to save even more power. Passing encoding/decoding entirely through the GPU is a bad idea for power consumption.

  • Roninja

    Whilst Series 7 will work with any cpu very keen to understand the tighter fabric with MIPS Warrior cores.

  • Cayl Sharp

    Non-sense I don’t blieve this:
    The GT7900 looks monsteruous, it’s the first time IT uses a 16 core configuration. what a treat woud it be if Apple make use of it nxt year for the A9X! I guess theoritically it should push well over 1 TFLOPS at a mere 300 Mhz or so, could we get some more numbers on that please.
    When I think I just started doubting IT for an instance… you realize that’s XONE PS4 territory we’re talking about here in less than one year fromtheir release?
    Thank you IT for what you have done for mobile all these years>>>

    • #?

      GT7900 probably would’t be a viable option for iPad form factor. XB1 and PS4 were announced with already dated hardware, they were not as revolutionary as their predecessors.

      It’s meaningless to compare current gen consoles for reasons mentioned – they were made to capitalize on an ever-growing market (gaming) not to revolutionize it. Comparison to GTX 750i was appropriate in my opinion.

      These guys at IMG are pioneers, I’m not surprised by their year over year progress. The likes of MSI, Nvidia and AMD have lazily remained stagnant for years due to lack of competition in markets they are active in.

      Maxwell was purposely delayed for 2 years, 40/32/28nm and Kepler milked to death. Those guys were arrogant, they didn’t even realize their markets had transitioned. Gaming went mobile years ago, people got bored of waiting and sitting around.

      There is bountiful convenience in performance on the go. The platform is still mocked by those still ignorant to the facts despite the innovative leaps being made over other platforms. Gaming is thriving on mobile and has been quickly dying everywhere else since the smartphone revolution came in 2007.

    • Anybody

      Actually a 16 cluster 7900 at 300MHz would end up at 307 GFLOPs FP32 or 614 GFLOPs FP16. In order to reach 1 TFLOP FP32 you’d need a tad less than 1GHz frequency and for 1 TFLOP FP16 a bit under 500MHz.

  • Alter

    I bet 500 GFlops handheld console is around the corner right now.

    • kratos

      There is no handheld console which has 500 Gflops. Ps vita has only 53 gflops.

  • mfaisalkemal

    What about power consumption series 7xt compare to series 6xt clock for clock, cluster for cluster at same node? And what about power consumption series 6xt compare to series 6 clock for clock, cluster for cluster at same node?

  • Nvidia always

    Blah blah, nvidia joined the mobile seen just a couple years ago and they are already paring with img, img has been in the mobile scene since 2006! Pioneers indeed, but nvidia is hardcore gaming!! They are the kings GPUs, ever heard of gtx titan z? Hehe, we’ll that gpu is the leading of all gpus, now as we speak they are making a mobile gpu that will handle 3+Tflops with maxwell architecture, so I don’t want to hear img beating nvidia for $hit nice try that they r trying though.

  • Biggmuscle

    Can this GPU Hardware Decoding for YUV 10bit H264/H265 4k 60fps??

    Please reply!!

  • Daniel Moravek

    will PowerVR one day return to desktop? 🙂

  • Mehmed

    along with gt7xt with 2, 4, 6, 8 and 16 clusters we also need 12 and 24 clusters.
    6XT with 4 Clusters for Apple A8 and 8 Clusters for A8X
    Now with 7XT 6 Clusters for A9 and possibly 12 Clusters for A9X
    The next 10nm Apple SoC will use 8/12 or 12/24 Clusters for sure.

    24 Clusters would be equal to the Xbox One GPU.