Multicore GPU aims at 3nm for data centre designs
Imagination Technologies has developed a new series of multicore graphics processor unit (GPU) aimed at the data centre.
The B Series GPU IP cores based on the PowerVR tile rendering architecture and provide higher performance than the A Series launched in December and can be grouped in clusters of four GPUs. The scalable GPU design delivers up to 6 TFLOPS of compute, with a reduction of up to 30 percent in power over the A-series by using optimised voltage and frequency domains across the core to provide up to 2.5x the fill rate than competing IP cores. These are being used with designs on current 7nm and 5nm process technologies with 3nm chip designs in evaluation.
The BXT uses a controller core in each GPU that can all act as a primary, while the BXE series uses one controller to manage three secondary GPUs. These can be implemented on a single chip or as chiplets that can be combined for different products.
“We are engaged from 7nm down so there is active work on 7, 5 and 3nm,” said Kristoph Beets, Senior Director of Technical Product Management at Imagination. “A lot of customers are excited about the process nodes and chiplets are an interesting approach – the main cost is to create the die and the challenge is how they are coupled together,” he said.
Imagination uses several different controller architectures to link the cores, with 33 different multicore options.
“With the B-series multicore we can put multiple instances of a GPU down and connect them with a light interface for up to four times the performance, so its quite a big step,” said Andrew Girdler, product manager at Imagination. “We have paired that with the power efficiency that gives a 30 percent improvement in same process node over the A series.”
“The de-centralised design with four repeated cores means you don’t have to worry about tightly coupling them,” he said. “The core controller breaks up the work of rendering a frame across multiple cores.”
“Traditionally GPUs are a central block,” said Beets. “What we have done is have GPU cores picking up work autonomously. There is one core in charge, for example using an AXI interrupt line so there is no special signals, no spaghetti logic,” he said.
“Our cores have always had their own firmware processor – you can either have one processor in control or run the cores individually, which might make sense in the data centre market, where the different cores are allocated to other customers. Historically we have used MIPS cores but we are in the process of moving to a RISC-V core on some. All the automotive variants have a full RISC-V processor – the consumer grade are an older instruction set, some are MIPS for highly area focussed designs while the higher performance for the BXT is a variation of META processor but all run the same code,” he said.
Imagination has had a focus on low power graphics suitable for mobile applications, and had deals with ARM and Apple in the past. This low power consumption is increasingly important for data centre chips that have to provide high performance in a rack with a fixed thermal limit.
The growth in streaming games, where the rendering is done in the data centre, is driving the need for higher performance chips. The B-Series also includes IMGIC image compression technology in the market to provide new bandwidth saving options. It offers up to four levels of compression, from pixel-perfect lossless modes to an extreme bandwidth-saving mode, which offers a guaranteed 4:1 or better compression rate. This gives more flexibility for SoC designers to optimise either for performance or to reduce system cost while maintaining a great user experience. IMGIC is compatible with every core across the B-Series range, bringing the benefits of Imagination’s industry-leading image compression to even the smallest cores.
The GPUs can also be used for mobile, consumer, IoT, microcontrollers, DTV and automotive chip designs.
IMG B-Series is available today and already has lead licensees for every product family with consumer chips expected in the next 12 to 18 months. 7nm chips are already in production at TSMC and volume 5nm chips are ramping up in volume. “These cores are already licensed and delivered to lead partners,” said Girdler.
Imagination is looking at larger clusters than four GPUS. “We do have ambitins to scale that up further rin the future,” said Beets. “As you can imaging putting a small workload on a big GPU is an issue.”
Related articles
- IMAGINATION IN CHINA LAB RISC-V DEAL
- UNISOC AND IMAGINATION STRATEGIC PARTNERS ON AI ACCELERATION
- IMAGINATION PUSHES RAY TRACING
Other articles on eeNews Europe
- Complete DDR5/LPDDR5 memory IP for TSMC’s 5nm process
- Ruthenium shows way to 2nm
- TSMC looks to 2nm process technology
- Data centre AI drives $30bn AMD-Xilinx deal