Intel Arc Alchemist GPUs Won't Support Hardware Accelerated FP64 Compute

[harlan4096]

With how useless FP64 is in the consumer space, this might be a good thing

A moderator on the Intel forums confirmed that Intel's consumer-focused Arc Alchemist GPUs will not feature hardware-accelerated FP64 cores, relegating the GPUs to native FP32 and FP16 support. The only exception to this is emulated FP64 support on Arc Alchemist which will be supported for niche cases.

However, due to the nature of emulation, FP64 calculations on Arc will run much slower than on GPUs with native hardware-accelerated FP64 cores.

Keep in mind that this configuration only applies to Intel's gaming-centric Arc Alchemist GPUs, not its upcoming Ponte Vecchio GPUs for the enterprise space.

FP64 is a computer number format commonly used in high-performance computing (HPC) applications. In addition, the large format has proven advantageous in very complex mathematical applications, such as physics, weather predictions, and simulations of many sorts, thanks to the wide dynamic range of numeric values FP64 provides.

However, FP64 is really only beneficial in the enterprise realm and has rarely ever become useful in the consumer space. Simply put, FP64 provides such a wide amount of numerical values that it's too big of a number to process for smaller calculations, done in workloads such as 3D gaming. In general, smaller calculations can be performed quicker on cores with FP32 and even FP16 capabilities. This is why you see consumer-focused gaming GPUs prioritizing FP32 and FP16 performance over FP64.

For instance, if we take a look at the GeForce RTX 3090 featuring Nvidia's latest Ampere architecture, you'll find that its FP32 performance comes in at an impressive 35.58 Teraflops. But the RTX 3090's FP64 compute capabilities pale in comparison to that performance metric, coming in at just 556 Gigaflops (not even a single Teraflop) -- or just 64th the performance of the GPU's FP32 capabilities.

This demonstrates just how unimpressive FP64 is in the consumer space. In the real world, this kind of FP64 performance is only beneficial for FP64 demos or benchmarks and rarely anything more.

With this in mind, it seems Intel's strategy to completely negate any hardware accelerated FP64 support on Arc Alchemist might be a good thing. The lack of "useless" FP64 cores gives Intel more room on the GPU dies to add important hardware, such as more FP32 or FP16 cores, additional hardware encoders and decoders or bigger caches.
...

Continue Reading