Intel has ditched the permanent component of the GPU. We know the details of FP64, FP32 and FP16 in Intel Arc Alchemist

Intel’s forums answered the question of what calculations the Intel Arc Alchemist cores will support. It was given by one of the company’s moderators, emphasizing that DG2 GPUs will not have physical accelerators for FP64 calculations, having native support only for FP32 and FP16. FP64 will only be available under very limiting emulation performance.

Intel Arc Alchemist GPUs without hardware FP64 cores

The GeForce RTX 3090, currently the most powerful graphics card in the consumer market, boasts 35.58 teraflops of FP32 performance and a modest 556 gigaflops of FP64. Is it a coincidence? Not at all, because double-precision calculations are mainly useful in the professional sectors, and in the consumer sectors (e.g. when playing), the FP32 and FP16 (single and half-precision respectively) play the greatest role. So it’s hard to blame Intel for abandoning something in the Alchemist architecture that would not be used very often in practice.

FP64 has long been associated mainly with High Performance Computing applications as it provides very accurate computational results. Games do not require such accuracy, being content mainly with FP32 and FP16 cores, so these calculations are given priority.

Also read: Gabe Newell himself contributed several Steam Decks. Everything you need to know about Valve’s Steam Deck

In practice, Intel often abandoned completely useless elements in the graphics processor, which thus freed up space in the physical silicon of the GPU, allowing for example to add more FP32 / FP16 cores, encoders, decoders, or a cache deck.