Imagination’s new GPU takes the fight to Apple, Qualcomm, and Arm
A new mobile GPU has been launched by Imagination, and it is set to bring some welcome competition to the mobile GPU arena. Lots of attention is given to Qualcomm’s Adreno GPUs — found in its Snapdragon processors — and to Arm’s Mali GPUs found in Samsung’s Exynos processors and Huawei’s Kirin processors. But, as I have written previously, those two aren’t the only mobile GPU makers out there. There is another.
Imagination’s PowerVR GPUs are currently found in processors from MediaTek, Unisoc, Allwinner, and Rockchip. I say “currently” because alliances are fluid in the world of mobile processors. Up to and including the release of the Apple A9 processor, Imagination was the sole supplier of GPU tech to Apple. That deal went sour, causing Imagination to regroup and find new partners. MediaTek is one of those partners, but the relationship isn’t exclusive. For example, MediaTek’s new Dimensity 1000 5G SoC uses Arm’s Mali GPU.
While Imagination’s management team has been looking for new markets and new partners, it seems that its engineers just got on with the job at hand, namely developing a next-gen GPU. The new IMG A-Series GPU has the potential to become a serious contender — not only in smartphones, but also in other mobile devices like Chromebooks.
When it comes to GPUs, there are three key metrics: performance, power, and compute. In other words, how fast will 3D games run? How much battery will it use? And can it be used for other tasks like machine learning? According to Imagination, the A-Series delivers significant improvements, at the same clock frequencies and using the same manufacturing process. It offers 2.5x the performance and 8x faster machine learning, all while using 60% less power than current PowerVR devices that are currently shipping. Those are some impressive numbers! The only fly in the ointment is that this specific comparison is with devices that are shipping in the real world, not with all existing GPU designs from Imagination. Still, these gains are not to be disregarded.
GPU designers need to offer their customers (i.e. the chip makers) lots of flexibility. That means being able to configure the same basic GPU design in different ways depending on the market. For example, Imagination offers the high end IMG AXT 64-2048. It delivers 2.0 TFLOPS, 64 Gpixels, and 8 TOPS of AI performance. This would be an ideal GPU for Chromebooks or even as an integrated GPU in a server.
The combination of possible configurations continues with the IMG AXT 48-1536 for premium mobile (1.5 TFLOPS, 48 Gpixels and 6 TOPS) through to the IMG AXM 8-256 for mid-range mobile. There are even variations for entry-level TVs and low-end smartphones with the IMG AXE 2-16 (2 PPC, 16 GFLOPS, and 2 Gpixels) and the IMG AXE 1-16 (1 PPC, 16 GFLOPS, and 1 Gpixels). The latter, Imagination claims, is the fastest Vulkan-capable GPU in its class.
As with previous PowerVR GPUs, the new A-Series uses tile-based deferred rendering technology, which means it only draws what is visible on the screen. To help deliver sustained performance, the IMG A-Series incorporates Pro-Active DVFS (Dynamic Voltage and Frequency Scaling) and Deadline Scheduling algorithms, which means that if parts of the GPU aren’t fully utilized or needed for processing, they are immediately slowed down or even put to sleep to ensure optimal power efficiency.
The internal design of the A-Series is quite different from previous PowerVR GPUs. At the heart of a GPU is a fast and efficient multiply-add unit for doing fused multiply and add (FMA) operations. These units perform a multiply and then an add, all in one step. Previous PowerVR GPUs used a 32 wide set of FMA units in a logical component known as the ALU (Arithmetic Logic Unit). The new A-Series widens the ALU to 128 operations. That is 256 lots of 32-bit floating-point multiply-add operations per clock cycle (256 because it multiplies two FP32 numbers).
The result of this new design is that the ALU pipelines are simplified while increasing the thread-level parallelism. The previous design relied more on instruction-level parallelism, which in turn required a more complex ALU and a more complex compiler.
The IMG A-Series uses Imagination’s HyperLane technology. With it, individual hardware control lanes allow memory isolation, meaning different tasks can be submitted to the GPU simultaneously for secure GPU multitasking. The GPU can be divided up according to raw power (i.e. one quarter for one task, three quarters for a different task) or dynamically with set levels of performance per task. For automotive applications, this could mean that the infotainment unit would get the lowest priority compared to the digital dashboard. This not only applies to traditional graphics processing, but also to machine learning tasks running on the GPU.
To handle the concurrency aspects of the GPU, the A-Series includes a built-in microprocessor, based on an unspecified architecture. In the past Imagination would have used a MIPS based microprocessor for such a task. However, after the Apple episode, MIPS was sold off. One thing we do know is that Imagination has been getting very cozy with various RISC-V chipmakers over the last 12 months. I will leave you to draw your own conclusions.
Transferring data inside of a GPU uses energy. The less data you transfer, the more efficient the GPU. Most GPUs use some form of internal compression because, for large streams of data, it is actually more power efficient to compress the data and then send it rather than sending it raw. The IMG A-Series GPUs use PVRIC4.1 lossless or visually lossless image compression for bandwidth reduction all of the time. The new compression technique is lossless in most cases, or visually lossless in exception cases. On the plus side, a 50% compression ratio is guaranteed, regardless of the data.
When can we expect these to come to market?
Announcements are nice, but when we will see a system-on-a-chip with an IMG A-Series GPU? The answer is sooner rather than later. The IMG A-Series has already been licensed in multiple markets and the first SoCs are expected in 2020. Also in 2020 Imagination will release its next GPU, the successor the A-Series, the cleverly named B-Series. It will deliver a 30% performance improvement compared to the Series-A and should come with hardware ray tracing!
A new powerful GPU has the potential to really mix things up in the mobile processor market. What do you think? Can Imagination break Qualcomm’s stronghold? Can it persuade any of Arm’s customers to switch? Let me know in the comments below.