Imagination Promises Big Performance, Efficiency Gains for Graphics and AI with the New DXTP GPU
Imagination's latest GPU IP comes with a claim of 20 percent efficiency gains — and will be the last in the D-series GPU family.
Imagination Technologies has announced its next-generation graphics processor IP, the Imagination DXTP — claiming that not only can it offer hardware-accelerated 3D graphics with better power efficiency than its predecessor but can also accelerate on-device machine learning and artificial intelligence (ML and AI) workloads, with up to two tera-floating point operations per second (TFLOPS) at FP32 precision and eight tera-operations per second (TOPS) at INT8.
"Imagination DXTP is a real example of many small steps achieving big gains and of the resourcefulness of Imagination's engineering teams," claims Imagination's chief product officer James Chapman of the company's latest launch. "The set of performance and efficiency improvements inside DXTP will enable future smartphones to run the next wave of gaming and AI applications with even lower power consumption than before."
The latest — and last, the company says — entry in Imagination's D-series graphics processor family, the DXTP GPU IP targets portable and embedded infrastructure from smartphones to laptops and from desktop computers to in-car infotainment systems. Imagination rates the part at 64 gigapixels per second for graphics throughput, two TFLOPS for floating-point compute at FP32 precision, and eight TOPS for integer compute at INT8 precision when running at the recommended 1GHz clock speed — and says that its in-house hardware-based virtualization system allows it to run graphics and compute operations side-by-side with "minimal overhead."
Compared to Imagination's last-generation DXT IP, the new DXTP delivers a claimed 20 percent improvement in power efficiency for popular graphics workloads, expressed as frames per second per watt (FPS/W), thanks to an unspecified range of microarchitectural improvements. For those focusing on AI workloads, the company has also released optimised OpenCL compute libraries that, it says, can deliver up to 80 percent GPU utilization for common compute workloads.
More information on the Imagination DXTP IP is available on the official product page; at the time of writing, the company had not disclosed any design wins, nor when the first products featuring its new IP are likely to hit store shelves.