PowerVR Series7XT Plus GPUs give mobile visual computing a boost

Imagination Technologies is announcing two new high-end mobile graphics processing units today in its PowerVR Series7XT portfolio. These designs—the GT7200 Plus and the GT7400 Plus—add integer pipelines and OpenCL 2.0 support to the GT7200 and GT7400 graphics processors. Those features purportedly make the Plus GPUs especially well-suited to tasks like computer vision, computational photography, and deep learning applications on mobile devices.

The GT7200 Plus is a "dual-cluster" design with a total of 64 ALU "cores," while the GT7400 Plus doubles those numbers to four clusters and 128 ALU cores. We've already examined the general principles of the Series7XT GPU, so we'll focus on that new integer pipeline here. The company says that while previous iterations of its Rogue-architecture ALU cores could only perform int32 operations, these Plus designs can now perform int8 and int16 operations, as well. Those more flexible processing paths are apparently useful for visual computing applications like those built using the Khronos Group's OpenVX API.

These Plus GPUs also add hardware features necessary to support the OpenCL 2.0 API. Imagination Technologies says the improved shared virtual memory support in these GPUs means that SoCs with these GPUs on board can share pointers between the GPU and CPU, reducing redundancy in memory and also reducing latency. These GPUs also add support for dynamic parallelism, meaning that the graphics processor can manage its own threads without CPU interaction. Dynamic parallelism purports to simplify the GPU programming model and lessen CPU overhead. It could also reduce SoC power consumption and allow many GPU computing algorithms to execute more efficiently.

As an example of what this hardware can do, Imagination Technologies says that Chinese software and services company Baidu was able to accelerate one of its deep-neural-network applications using PowerVR GPUs in mobile devices. That application can perform image classification in real time using input from the device's camera to distinguish among various objects in a scene (a shirt and purse, for example).

According to Imagination, other applications that could benefit from the Plus-series chips' visual-computing features include face, body, and gesture tracking, smart video surveillance, HDR rendering, driver-assist systems in cars, object and scene reconstruction, augmented reality, visual inspection, and robotics.

Comments closed
    • wingless
    • 7 years ago

    har har har….

    • Klimax
    • 7 years ago

    DX11 is required. It is mainline even with DX12 present. Remember, DX12 is parallel not successor. (Also IMO, it doesn’t have good future compatibility.)

    • Klimax
    • 7 years ago

    Or how about testing same to same and thus FP16 or FP32 everywhere…
    (Would be definitely more interesting)

    • tipoo
    • 7 years ago

    It may be easy to scale up, but I was speaking of the *current* performance which some people ran away with comparing FP16 to FP32. Tests like GFXbench may paint them as similar to 650M/930/Iris Pro levels, but like for like tests show them half to a third as powerful, even the iPad Pro with it’s monster-for-ARM GPU.

    • guardianl
    • 7 years ago

    Recall fail on my part, I thought you worked for Imgtec. Intel, Imgtec, they both start with I…

    • guardianl
    • 7 years ago

    Higher-precision ALU throughput is maybe *the* easiest thing to scale up in a GPU. The FP16 vs FP32 ratio is just a result of the mobile focus with the normal power/perf trade-off.

    If ‘desktop’ GPUs were only used for gaming (no DCC apps etc.) we might very well be using FP16 for pixel shaders in 2015. Given the insane levels of block-level texture compression most games use (partially because of last gen consoles) sometimes it seems like we might as well…

    • Andrew Lauritzen
    • 7 years ago

    No idea about Linux or PVR, sorry.

    • guardianl
    • 7 years ago

    BTW Andrew, I can’t figure out what the status of OpenGL drivers for Linux is for PowerVR. The MIPS Creator-CI20 binary drivers seem to support OpenGL 2.x? Is there some type of support document at PowerVR that specifies what’s available with each GPU series etc? A roadmap? 🙂

    • ronch
    • 7 years ago

    Looking at the diagram above, I can see this architecture is gonna be a FLOP.

    • ronch
    • 7 years ago

    I stopped reading at ‘cores’.

    • tipoo
    • 7 years ago

    Yeah, careful there. Some sites like Ars Technica took one look at GFXbench results and then claimed the iPad Pro had a GPU as fast as the Iris Pro 5200/930M level, and that started making the rounds online. Thing is, that test used largely FP16 on mobile, largely FP32 on PCs. If you looked at other tests you’d find the 650M/Iris Pro level GPUs really scored 3x as fast as the iPad Pro.

    It’s impressive, no doubt – but for ARM. It competes with Core M level GPUs, but not even Iris.

    • auxy
    • 7 years ago

    Sandy Bridge had 6 or 12 EUs, depending on configuration, and Broadwell has 12, 24, or 48.

    It’s really the same as what AMD or NVIDIA do with their GPUs. GCN is a modular design, and AMD just sticks more bits on when they want it to go faster. (*’▽’)

    • Andrew Lauritzen
    • 7 years ago

    This, and workloads really. Remember that stuff written for desktop is definitely not optimized for stuff like TBDR, partial precision math, various mobile texture formats and other stuff that is necessary to make these chips run at their potential. Could change in time, but don’t expect an off the shelf PC game to run well on current mobile chips, even if there were drivers.

    • Pitabred
    • 7 years ago

    Which, given the architecture of DX12/Vulkan, wouldn’t be too horrible. They’re very slim drivers overall. The biggest (only?) stumbling block right now would DX11 and older, because they’ve already got OpenGL drivers on the mobile systems which wouldn’t take a lot of work to port.

    • guardianl
    • 7 years ago

    The issue is software not hardware. They already have decent 10W+ designs that are as fast as entry level Radeon/Geforces (i.e. 930M) – just see the iPad pro reviews. This means they could enter the laptop hardware market basically tomorrow if they had the software…

    Writing a set of competitive DirectX 12/11/10/9 + OpenGL 2x/3x/4x and Vulkan drivers is the real roadblock.

    • tipoo
    • 7 years ago

    I’m actually excited to see the 72EU Skylake GT4e. They’re fairly competent with 40, the architecture seems ok, with 72EUs and the eDRAM bandwidth to match it it could be quite good.

    • UnfriendlyFire
    • 7 years ago

    They could take 3Dfx’s route of sticking 4 GPU chips on one GPU card, and try to scale up.

    Or Intel’s route of throwing moar execution units at the GPU with every revision (Sandy Bridge had like 4-6 execution units, Broadwell has a few DOZENS).

    • Prestige Worldwide
    • 7 years ago

    They would need to do a lot of work to scale up their tech from low power SoC nodes to high power desktop nodes.

    • kuttan
    • 7 years ago

    PowerVR should again introduce GPUs for the desktop PC gamers.

Pin It on Pinterest

Share This

Share this post with your friends!