The chip giant has confirmed that the Intel Xe (Gen12) architecture is so scalable that they have been able to use it to shape a GPU with nothing more and nothing less than 16,384 shaders, an astronomical figure that leaves us with a truly surprising gross power in FP32, 42 TFLOPs.
I know what you’re thinking, and how did Intel manage to shape such a massive graphics core? Well, it is relatively simple, stacking blocks of execution units. We already anticipated it in this news, where we saw that the series of Intel high performance graphics solutions (HPC) could be divided into three main levels:
- Base level with 512 execution units, which leaves us a total of 4,096 shaders and an estimated power of 12.2 TFLOPs, with a TDP of 150 watts.
- Medium level with 1,024 execution units, a figure that could give us a gross power of 20.48 TFLOPs and a TDP of 300 watts.
- High level with 2,048 execution units, which yield that total of 16,384 shaders with a gross power of 42 TFLOPs and an estimated TDP of 500 watts.
I want to make it clear that the architecture used by these GPUs is of the MCM type, that is, a multichip module. The base level integrates a single block with 512 execution units, the medium level adds two blocks of 512 execution units each, and finally we have the high-end level, which mounts four blocks 512 execution units each.
As our more advanced readers may have guessed, this architecture is based on the same foundation as MCM designs that we have seen in the AMD Zen, Zen + and Zen 2 processors, but applied to the graphics sector. This architecture simplifies the adaptation of designs to the wafer, improves feasibility and success rate, and enables more powerful and complex solutions to be created at a lower cost and investment, at least in theory.
Intel conducted an interesting performance test to demonstrate the potential of its new architecture. In it we saw that a single block (512 execution units) can transcode 10 separate streams of 4K content at 60 FPS HEVC. Yes, it’s awesome, but the best thing is, according to the chip giant, block scaling of execution units is linear, which means that theoretically the most powerful model with four execution unit blocks should be able to seamlessly transcode 40 different sequences.
And what does Intel have in mind for the general consumer market?
Everything we have seen so far will be aimed at the professional sector, but lay a foundation that will be used in the graphics cards that Intel plans to bring to the mainstream consumer market, and therefore it is important to analyze them.
According to a recent leak, Intel is working on an HPG (high performance gaming sector) graphics solution that has been built on three main pillars: efficiency, scalability and performance. It hasn’t confirmed its specs, but we know that it will use GDDR6 memory and that it will support hardware-accelerated ray tracing.
Looking at the gross performance offered by the most powerful models aimed at the professional sector, I think Intel will most likely start from a configuration of a block of execution units, that is, of 4,096 shaders, and from there, it shapes different models with lower configurations to build a catalog of low-end, mid-range and high-end solutions.
If all goes according to plan, Intel’s first high-performance graphics card for the gaming sector will arrive sometime in 2021, a date when it will have to compete with NVIDIA’s Ampere-based RTX 30s and AMD’s RDNA 2-based Radeon RX 6000s.