PowerVR GT7900: redefining performance efficiency in graphics

This year MWC and GDC 2015 are happening during the same week which is unfortunate for me because I really enjoy going to both events.

Although attended by different demographics, a common theme you will be hearing next week is how the exponential increase in performance is paving the way for console-quality graphics on next-generation embedded devices, including smartphones and tablets.

In this article I would like to give you a teaser of what I’ll be talking about next week to the many press and analysts who will be visiting our booths in Barcelona and San Francisco, respectively.

PowerVR GT7900 is a super-GPU designed for ultimate performance

At last year’s Mobile World Congress, I unveiled PowerVR GX6650 – a high-end Series6XT GPU for flagship devices; the announcement was covered by AnandTech, The Register, VentureBeat, Android Authority and many other publications.

Now I’d like to focus on PowerVR GT7900, the highest performing GPU from our Series7XT family. We have already seen eight cluster configurations used in mobile devices and we expect this trend to continue, especially since the new 14/16 nm process nodes offer significant improvements in performance, power and area.

PowerVR-Series7XT-GPUPowerVR Series7XT GPUs scale from two to sixteen cluster configurations

However there is a new category of devices that can use these embedded GPUs: affordable game consoles.

PowerVR GT7900 is a super-GPU designed for ultimate performance in embedded applications; this new design takes everything from its predecessor and turns the dial to 11:

For example, peak GFLOPS performance is more than tripled thanks to a 2.6x increase in ALU cores and further architectural improvements.

Analyzing the architecture displayed in the diagram below, we find 512 ALU cores (vs. 192 in the GX6650) arranged in 16 unified shading clusters (USCs). For those who want to learn more about the internal architecture of our Rogue USCs and our definition of an ALU core, you can read this blog post from my colleague Rys.

PowerVR GT7900 GPU - PowerVR Series7XT The internal architecture of our high-end PowerVR GT7900 GPU

PowerVR GT7900 delivers the ultimate PC-class gaming experience for embedded devices: it packs the complete feature set for OpenGL ES 3.1 + AEP (Android Extension Pack) compatibility, including full hardware support for tessellation and ASTC LDR and HDR texture compression standards.

We’ve also added virtualization at the GPU level to complement the same technology found in our MIPS Warrior CPUs; this is a unique feature in embedded graphics which enables multiple operating systems to run concurrently in secure containers at the platform level.

Take the heat off

One of the most common problems affecting flagship devices targeting ultimate performance is overheating. This is where an unbalanced design can really cause issues under long-term usage due to an inefficient architecture.

PowerVR GPUs deliver sustained performance thanks to a sophisticated TBDR architecture that has been refined over the years to deliver unmatched efficiency.

The chart below shows compares several mobile devices running a demanding graphics application – the same type of content you would encounter when playing a console-level game; notice how our quad-cluster PowerVR Rogue GPU outclasses competing solutions while also maintaining sustained performance:

PowerVR-Rogue-GPU-vs-competition-long-term-performancePowerVR GPUs deliver sustained performance over time

To help system designers prevent thermal throttling, we’ve also implemented an advanced power saving mechanism called PowerGearing which enables the GPU to shut off parts of the design in power-constrained applications.

PowerVR_GT7900_PowerGearingPowerGearing is an advanced power saving mechanism designed to prevent thermal throttling

Furthermore, PowerVR GT7900 includes a low-power FP16 mode which can provide more than 1 TFLOPS of graphics and compute performance within constrained power budgets.

At a target frequency of 800 MHz for 16nm FinFET+, PowerVR GT7900 delivers 800 GFLOPS in FP32 mode and 1.6 TFLOPS in FP16 mode.

4K resolutions and beyond

PowerVR GT7900 is the second generation of GPUs to feature native 10-bit YUV color support. The diagram below shows how system architects can build a high-performance platform capable of handling 4K resolutions at 60 fps and beyond from capture to display; this includes:

  • Main processor: dual-threaded, quad-core MIPS I6400 CPUs organized in two clusters delivering best-in-class performance for Android 5.0 Lollipop
  • Multimedia: PowerVR GT7900 graphics processor (GPU), PowerVR Series5 video encode and decode processors (VPU), PowerVR V2500 camera processor (ISP)
  • Connectivity: Ensigma Explorer radio processor (RPU) for high-speed, low latency 802.11 ac 2×2 Wi-Fi, Bluetooth 4.0 (and Bluetooth Smart) and digital radio

PowerVR GT7900 - game console chipAn example of a chip designed for affordable game consoles

For consumers, PowerVR GT7900 – and the Series7XT family in general – will usher a new era of console-like gaming experiences. To give you an idea of the level of performance these new high-end GPUs will be able to achieve, here are a few screenshots from Dwarf Hall, our latest demo to be shown at GDC 2015.

PowerVR_DwarfHall_demoDwarf Hall implements console-quality effects and runs on high-end PowerVR Rogue GPUs

Some of the effects we’ve implemented in Dwarf Hall include:

  • 1 million triangles per frame in some scenes
  • Deferred shading with 128-bit physically-based G-buffer
  • Many dynamic lights
  • Multiple specular probes treated as lights
  • Soft particles
  • Lens flare effects
  • Full post-process pipeline with colour correction, saturation, dynamic exposure and HDR tonemapping

I’m really excited to see how our partners will use these new high-performance parts in designs coming later next year.

If you are at GDC 2015 and want to learn more about the future of mobile and embedded graphics, make sure you don’t miss our idc15 developer day. We will also be showing our latest demos running on PowerVR Rogue GPUs.

If you will be at MWC 2015, drop by our booth to see some console-quality content running on PowerVR Rogue GPUs; you can also click on the button below if you’d like to book a meeting.

Book a meeting at MWC 2015
Make sure you also follow us on social media (@ImaginationPR, @PowerVRInsider) for the latest news and announcements from both these events. You can also keep in touch with me on Twitter at @alexvoica.

, , , , ,

  • Pingback: This is what the PS4 and Xbox One should really be afraid of : TechWorthy()

  • Pingback: This is what the PS4 and Xbox One should really be afraid of | android news updates()

  • Sean Lumly

    Outstanding!

    I’m really glad to see that Imagination is continuing to target higher-power mobile devices, and this GPU, at 14/16nm, is simply incredible in the sheer number of computations and memory accesses that are possible. Outside of external bandwidth (and more relevantly memory-units) it is not too far removed from current consoles in the types of content that should be possible, especially with on-chip deferred rendering, and other budget-conscious innovations. And external bandwidth can be greatly increased with ASTC at low bit-rates offering a very nice tradeoff.

    This is especially relevant for mobile VR that will likely see 4K resolutions @ 60fps before long and require the ability to sustain much more computation for good performance.

  • http://blog.imgtec.com/author/alexvoica Alexandru Voica

    Thanks, VR is definitely an application that is suited for high-resolution, high-performance GPUs. Add ray tracing to the mix, and you get even better photorealism.

  • Sean Lumly

    Indeed. Realtime ray-tracing has the potential to literally re-define a tremendous swath of interactive computer graphics starting with transparency, occlusion, and reflections — very expensive and poorly serviced areas of rendering. It would take a lot to be able to render a ray per pixel at 4K (nearly 500M rays per second), but even at a quarter of the resolution (1080p), it would provide far better fidelity than the competing rasterization solutions.

    For VR this is invaluable, especially for reflections that are less convincing with stereo rendering.

    I’m looking forward to higher numbers of load/stores and texture units being packed into these GPUs. There are many effects that work well when sampling over large numbers of pixels (eg. 256) and can be quite cache friendly due to spatial locality. The trouble is getting this data into the GPU core’s registers in the first place! As resolutions go up (something tells me that we’ll stall at 4K for some time), the number of cycles per pixel goes down, which is a memory limiting factor outside of external bandwidth.

    IMGtec does well with 2 issues per shared unit, though. A monster like this could do 16 fetches in a cycle which is pretty darn good! I think we’d start seeing some really interesting things enabled if that number were doubled.

  • Pingback: myandroid.com | This is what the PS4 and Xbox One should really be afraid of()

  • dumb

    Can you give spec on the ISP like Qualcomm did.

    Are you guys vaguely claiming 4K 60 fps encoding.

    There should be law to force you to state the TDP when mentioning
    GFLOPS.

    Not even a cursory mention of OpenCL. What a tangle web.

  • http://blog.imgtec.com/author/alexvoica Alexandru Voica

    Hi,

    Can you please let me know what specs you are looking for? I will gladly answer any questions; here is the original press release about the ISP:
    http://www.imgtec.com/news/detail.asp?ID=837

    We are not making vague claims, we have a very scalable video architecture that can definitely handle the complete toolkit for HEVC processing at 4K and beyond. Here is what we’ve announced so far:
    http://blog.imgtec.com/powervr-video/powervr-d5500-decoding-hevc-in-10-bit-colours-at-4k-resolutions

    Since we license GPU IP, we can’t really talk about TDP; TDP is relevant for chips – however, the chart above showing performance efficiency over time for a heavy graphics workload should give you an idea of how our GPUs tackle high-performance graphics. If you are asking about the total TDP of a mobile device, you would be looking at 1.5-2W for a smartphone and 4-5W for a tablet.

    The original PowerVR Series7XT announcement talks about OpenCL. We have
    added HPC support for compute servers that require FP64 and OpenCL 1.2 Full Profile capabilities. More details below:
    http://blog.imgtec.com/powervr/powervr-series7xt-gpus-push-graphics-and-compute-performance

    Hopefully this answers your questions.

    Regards,
    Alex.

  • GPU

    Really good. Nvidia X1 and GT7900 the best GPU for mobile devices!!!

  • LDM

    Hi Alex, excellent article!

    It seems this beast GPU will have a certain market targets. Would you think would be possible seeing it in the tablets too? Or it is too much power constraint to get inside a tablet?
    You guys seem to target Android environment, any chance to see this beast running direct d12?
    Cheers
    L

  • http://blog.imgtec.com/author/alexvoica Alexandru Voica

    I would say you’d need to wait for 10nm (or lower) to consider using this type of GPU in a mobile device.

    We offer DirectX support too so a customer could definitely license PowerVR to build a Windows device.

  • MaqueGenio

    Or maybe you can use in a tablet (or even a handheld console) but with a GPU freq. not quite as aggressive as in the example above (800mhz)… I mean this beast has to fit in mobile devices, otherwise it would be a waste and pointless to put effort into a GPU design that will end up serving such a miniscule market that is android based micro consoles.

  • http://blog.imgtec.com/author/alexvoica Alexandru Voica

    A lower frequency should help, but it is about size too.

  • Pingback: This is what the PS4 and Xbox One should really be afraid of | Technology Trends()

  • Pingback: Imagination PowerVR GT7900 GPU is Made for Game Consoles and Notebooks()

  • Pingback: PowerVR GT7900 una nueva GPU de Gama alta | AndroidPC.es()

  • Pingback: Imagination avtäcker PowerVR GT7900 för Android-spelkonsoler » KATHING()

  • Pingback: This is what the PS4 and Xbox One should really be afraid of | Gadget Guru Reviews()

  • Pingback: PowerVR goes 4K with GT7900 for game consoles | CSR PRODUCTIONS Entertainment Group, Inc.()

  • Pingback: Imagination Technologies revela PowerVR GPUs 7XT Series, monstrinho capaz de 1,6 TFLOPS no modo FP16 - eXtremeMods - Tecnologia()

  • Pingback: Чего боятся PS4 и Xbox One? Или консоли на Android наносят ответный удар! - GadgetsVik - новости, обзоры и советы()

  • Pingback: PowerVR GT7900: la nuova GPU mobile batte, e non di poco, una GeForce GT 730M | NUTesla | The Informant()

  • Sean Lumly

    Does Imgtec have boards with these GPUs available for developers (purchase or otherwise)?

  • http://blog.imgtec.com/author/alexvoica Alexandru Voica

    We have done test chips for Rogue GPUs but we don’t have any boards to sell directly at the moment. There is only a dual-core MIPS CPU + PowerVR SGX board available running Debian 7/Android 4.4
    http://blog.imgtec.com/powervr/mips-creator-ci20-development-board-now-available

  • Sean Lumly

    Thanks!

  • Virat Kohli

    Very impressed with this and from what you say Alex this is not an on paper academic exercise and we can expect you partners to bring GT7900 devices to market in 2016 which is very exciting. I’m therefore encouraged to see such a broad range with series 7 and hopefully you have all bases covered this time as series 6 has been disappointing in terms of end license product dominance – fruits aside!

  • http://blog.imgtec.com/author/alexvoica Alexandru Voica

    Hi,

    Previous Rogue GPUs are included in more than ten SoCs, including chips from Allwinner, Actions, Intel, LG, MediaTek, Rockchip and others.
    Here is an example from a few hours ago:
    http://www.pcmag.com/article2/0,2817,2477401,00.asp

    Regards,
    Alex.

  • LDM

    Hi Alex,

    Correct me if I am wrong. This is the first time IMG announces a product outside your usual market targets: mobile – smartphones and tablets.
    Small consoles and notebooks is something new where you guys are starting to focus for the future.
    Is it something that IMG thinks is a new development area – Android consoles?

    From my personal opinion at moment there are not real games in android where you can push really the envelope. All the tablets and smartphones can easily run any (almost) Android games.

    So how you can jiustify a beast like that then?

    For example, one of your competitor, Nvidia has created a game tablet, but it is struggling to get a proper success in the market. One of the reason is that there many low end tablets that can easily run the same games of Nvidia tablets.

    Do you think the situation will change and hence the necessity of this beast?

    Please don’t get me wrong, I am really excited of this new Powervr series 7. But I reckon there is the need that Android environment starts to push much more the games!

    Cheers

    L

  • http://blog.imgtec.com/author/alexvoica Alexandru Voica

    Hi,

    Our strategy is to continue offering scalable GPU solutions for mobile and embedded devices; that includes markets like game consoles, notebooks or compute servers where a PowerVR GT7900 fits really well.

    We are making sure that if a certain customer would want to break into such a market, we have the right GPU IP available.

    Regards,
    Alex.

  • Steve Smith

    I would love to see PowerVR make a comeback on the desktop PC market. My first 3D accelerator was an Apocalypse 3Dx. Great card for the time.

  • LDM

    Hi Alex,
    Today it has been announced the Nvidia Shield console.

    Is that a similar market you guys want to hit with the Powervr GT7900?

  • http://blog.imgtec.com/author/alexvoica Alexandru Voica

    Yes, that is one of the three market segments this GPU targets, together with notebooks and low-power compute servers.

  • Pingback: Imagination nos ofrece gráficos de consola con su GPU PowerVR GT7900 para Android | Gadgetlogia()

  • Virat Kohli

    Does performance scale linearly with increased frequency up to a point of inflection. Is this 800mhz or higher?

  • http://blog.imgtec.com/author/alexvoica Alexandru Voica

    Yes, in the case of GPUs, performance usually scales linearly with frequency.

  • ray kim

    Have a question about two things.
    First, Is F16SOP real works for real world?

    Some game engine shows performance improvement using F16 intentionally but general application in the market dosen’t have optimized shader considering F16 ALUs. Also there are compatibility issue legacy pattern of shader (almost case are wrong in the shader of app because developer frequently mistakely use precision). So effect of F16 ALUs seems limited. If any result of real application testing, would be proof.

    Second, What changes of architecture regard F16SOP pipeline
    modification have more efficiency since
    Serise6XT? Serise6 not shows significant merit for using F16 ALUs. But
    how latest architecture offer more efficient way in F16?

  • http://blog.imgtec.com/author/alexvoica Alexandru Voica

    Hi,

    A lot of graphics applications use FP16; for example, it is used in UIs and casual games which are not geometry heavy. Another example where FP16 is used is pixel processing which typically inputs and outputs ARGB8888 (32 bit) colour.

    In Series7XT we have made a lot of efficiency improvements vs. Series6XT, leading to gains of over 60% in real world performance. You can find the full list here:
    http://blog.imgtec.com/powervr/powervr-series7xt-gpus-push-graphics-and-compute-performance

    Regards,
    Alex.

  • EM87

    This would be pretty fucking great for the new Nintendo NX console, with the handheld sticking with something similar to the iPhone 6 GPU (both would use ARMv8 CPUs). Any chances of this happening? Read somewhere that Nintendo talked with IT and AMD for their next hardwares, and this sounds like a great solution for what they’re going after. How much does this GPU cost? How is power consumption?

  • EM87

    Is there a specific reason why you removed my comment? Didn’t know it was against the rules to make a comment and ask some informations about the piece of tech featured in the article, such as cost and power consumption.

  • http://blog.imgtec.com/author/alexvoica Alexandru Voica

    Hi,

    I’ve sent you an email earlier today about your comment and why the system deleted it but you might have missed it.

    Essentially, the comment management system filters any profanity; if you can please repost the comment without that specific word you’ve used, you should be fine and I’ll be able to reply to your questions.

    Thanks,
    Alex.

  • EM87

    Ok, thanks for the answer. I totally missed the email :p

    Keep up the good work.

  • EM87

    This would be pretty great for the new Nintendo NX console, with the handheld sticking with something similar to the iPhone 6 GPU (both would use ARMv8 CPUs). Any chances of this happening? Read somewhere that Nintendo talked with IT and AMD for their next hardwares, and this sounds like a great solution for what they’re going after. How much does this GPU cost? How is power consumption?

  • http://blog.imgtec.com/author/alexvoica Alexandru Voica

    I unfortunately can’t comment on any rumors or various potential discussions we might or might not have with other companies. In terms of cost, we don’t make or sell graphics cards directly; instead, we license our designs to semiconductor designers. You can find an explanation of our business model here:
    http://www.alexvoica.com/the-ip-licensing-business-model-a-love-story/

    In terms of licensing costs, the price depends on a number of factors, including performance, size and target application. For power consumption, there are a number of reasons to consider – including the choice of a process node.

    I would say that for a sub-10nm process node, this would be suitable for use inside a gaming tablet (<10W). Anything above that, it would have to target a slim laptop or a game console.

  • EM87

    Ok, thanks for your answer.

  • Mehmed

    I would like to see what will be possible with 10nm and PowerVR 8/9 Series.
    There should be at least 1024 ALU Cores, right?
    It would be at least 2 Tflops i think. Faster than the PS4! with 128 less ALU Cores.
    14/16nm for 2016, that means 2018 we should see 10nm parts from Samsung/TSMC
    greatness awaits. And now coupled with Vulkan API…

  • Pingback: Imagination Technologies dalle GPU per iPhone alle super-GPU per console - macitynet.it()

  • Mehmed

    I think this year we’ll get a GT 7950 instead of a GT 8900.
    Am i right? (as a design for next year of course)

    And after that the 8900 with 1024 Shader in 10nm finfet 😀 or maybe even 1280.

  • velurpaiyan

    Hi, Could you please point me to any article which discussed the implementation of a PowrVR 6 or 7 series GPU core. Thanks.