As we get close to the launch of Intel’s first major discrete graphics family, the latest patent has revealed that the blue team is also working on its own multi-chiplet GPU design for the next-gen Arc gaming lineup.
Intel’s Next-Gen Arc Gaming GPUs To Leverage From MCM Design, Hints New Patent
Intel is no stranger to MCM technology, in fact, their Ponte Vecchio GPU is a full-on Multi-chiplet design though it is being designed and built for the data center segment. However, a new patent titled ‘Position-Based Rendering Apparatus and Method for Multi-Die/GPU Graphics Processing’, which has been discovered by Underfox, talks about how several GPUs (MCM) can work together for image rendering.
Patent: Position-based rendering apparatus and method for multi-die/GPU graphics processing – Intel
Intel MCM GPUs is coming…
More details: https://t.co/GIkfwrXGzV pic.twitter.com/sXGt9nbJ1S
— Underfox (@Underfox3) February 3, 2022
In the patent, Intel proposes a solution to GPU-based graphics rendering by integrating multiple silicon dies into a single cohesive unit (MCM) capable of running a single 3D context in order to address manufacturability, salinity, and power delivery problems. It is stated that doing this will require solutions for multiple classes of scalability and interconnect challenges in order to deliver the best performance on a single 3D application running multiple GPU dies. Some of the algos that are currently in use to solve these problems include AFR (Alternate Frame Rendering) and SFR (split frame rendering).
The patent specifically talks about integrating tile-based checkerboard rendering, which is a very common rendering technique used on current GPUs, along with distributed GPU vertex position calculation to enable more efficient scaling on multi-die GPUs. The draw calls are partitioned in the form of puzzles and sent to the individual GPUs on the MCM chip so that they can execute POSH (Position-Only Shaders) to determine the full-frame visibility data for the draws for all dies.
Visibility data then indicate if the given primitive is present on each of the pre-defined set of screen space tiles (checkerboard). When the relevant visibility data is received, each GPU die will use it to limit the geometry work only to relevant primitives and perform pixel processing work for all the titles of the checkerboard that it owns. These GPUs work really fast in working on the single-frame.
Overall, with such a design, the next generation of Intel Arc gaming MCM GPUs can deliver scaled-up performance with more dies being added to the chip without facing the performance limitations of existing multi-die solutions such as AMD XDMA/Crossfire and NVIDIA SLI/NVLINK. Intel states that these solutions are limited in terms of performance and while one should expect a perfect scaling of 200% performance with a 2-way and 400% with a 4-way configuration, the actual GPU performance is significantly lower, and in most cases, non-existent hence why most GPU makers aren’t investing in these technologies any more.
This would allow the GPU system to scale more efficiently. Now one would say that the 4-GPU/tile solution that is used as the example is official but that’s the design that has been proposed and the final number of GPUs per tile could be vastly different.
As mentioned earlier, all GPU makers are investing in multi-GPU technologies with AMD already out with the world’s first MCM solution in the form of their Instinct MI200 series, NVIDIA releasing Hopper this year with MCM. There are also the gaming consumer parts such as the AMD RDNA 3 lineup which is expected to utilize the same tech when it launches later this year. Intel themselves are working on several tile-based multi-chip CPU & GPU solutions such as their Meteor Lake and Granite Rapids designs and within 1 or 2 gens, we could see a proper MCM solution within the Arc gaming graphics lineup.
Intel ARC Gaming GPU Lineup
GPU Family | Intel Xe-HPG | Intel Xe2-HPG | Intel Xe3-HPG | Intel Xe Next | Intel Xe Next Next |
---|---|---|---|---|---|
GPU Products | ARC Alchemist GPUs | ARC Battlemage GPUs | ARC Celestial GPUs | ARC Druid GPUs | ARC E*** GPUs |
GPU Segment | Mainstream / High-End Gaming (Discrete) | Mainstream / High-End Gaming (Discrete) | Mainstream / High-End Gaming (Discrete) | Mainstream / High-End Gaming (Discrete) | Mainstream / High-End Gaming (Discrete) |
GPU Gen | Gen 12 | Gen 13? | Gen 14? | Gen 15? | Gen 16? |
Process Node | TSMC 6nm | TBA | TBA | TBA | TBA |
Specs / Design | 512 EUs / 1 Tile / 1 GPU | TBA | TBA | TBA | TBA |
Memory Subsystem | GDDR6 | TBA | TBA | TBA | TBA |
Launch | 2022 | 2023? | 2024? | 2025? | 2026? |