Imagination rounds out D series GPU, looks to E series
Imagination Technologies has launched the final member of its D series graphic processing unit IP, boosting the bandwidth and performance for AI applications as it looks to the next generation, expected to be the E Series,
“This is the last entry in the D family,” said Kristof Beets, veice president of product management at Imagination tells eeNews Europe. “It’s always a little artificial, what’s an iteration or a new architecture. The next step is bigger and more fundamental in performance increase.”
The DXTP core is optimised for data flow for AI applications, adding an extra port and changing the ratio of resources in the core. This adds a new data path and fixed function block.
“There’s enormous amounts of synergy between graphics and compute,” he said.
D-Series articles:
- Ray tracing GPU aims for mainstream mobile gaming
- Safety architecture boosts automotive GPU for chiplets
- GPU IP extends battery life for mobiles
“We are making it easier to implement and this is why we keep saying graphics becomes a form of compute and we are now slotting in graphics as a use case as its really compute first and the graphics is enabled by that functionality. In the past it was the other way round and that is very much starting to reverse. We have seen that from developers in the graphics market as well as they by pass fixed function unts and use the programmable compute so they can use their own tricks to speed up.”
“At the beginning we had a lot of pressure for density so we created units with three compute clusters, the texture unit and ALU with fixed function geometry blocks with data paths out the top and out to the system,” said Beets. “The wider you go, the more dense it can be, but you have a balance with performance and bandwidth. What we found by looking at smartphone applications was that architecture needed to be more geometry heavy as you could be limited by the fixed function blocks.”
“Now we have three ports to give more bandwidth so it’s a different balance point. We have also changed the ratio to 3 x 2 for the six building blocks (shown above) and that gives us more bandwidth, more hidden surface removal with a whole new data path and fixed function unit.”
Some of this comes from recognising the changes in the market from graphics to AI as well as changes in the AI frameworks that are very network intensive, requiring more focus on the data flow to keep the NPUs filled.
“For every compute there is set up work to prepare the registers and that is 16x faster by tuning the fixed function logic from one setup per clock to go massively parallel. Some of the processing jobs were relatively small, and in those cases we were set up limited and that makes a big difference,” he said.
There is also more focus on in memory compute.
“Our GPUS traditionally have a lot of storage so we are now very much applying that same principle so it becomes tile based compute and we keep data on chip. We have half a megabyte (500K) per unit and that is fully usable so one GPU is 3Mbytes of very flexible local register space. This is also a reuse buffer with 32K to reuse.”
The Imagination GPU IP is not just for mobile applications. Self driving startup Zelos is using the DGXT core as a compute engine.
“This is a massive growing area for us. We developed the computational libraries for them focused on neural networks, FFTs and linear algebra for ADAS,” said Beets. “We pick relevant networks but there are odd ones they can look at the optimised libraries and map it to their own flow so we focus on the intrinsic building blocks.”
“That’s very similar to what we see in CUDA. Developers are coming at a higher level and calling CUDA libraries that Nvidia optimised.”
All of this sets the scene for where Imagination is going, says Beets, with a focus on compute and AI, software and tools for the next generation E-series. “The neural processing unit has gone but what we found that most customers have their own so this is about how we fit into that framework and how we make our IP fit with those blocks,” he said.
If you enjoyed this article, you will like the following ones: don't miss them by subscribing to :
