In a bold move, Oracle has announced it's taking orders for a supercomputer cluster boasting an impressive 2.4 zettaFLOPS performance. However, this eye-catching figure requires closer examination.
The cluster, featuring 131,072 Nvidia Blackwell accelerators, claims to have "three times as many GPUs as the Frontier supercomputer," currently the United States' top-ranked supercomputer. But the devil is in the details – or more precisely, in the bits.
Oracle's headline-grabbing zettaFLOPS figure is based on "AI zettaFLOPS" using 4-bit floating point (FP4) precision. This is a far cry from the standard 64-bit double precision typically used to measure supercomputer performance.
While 4-bit precision can be useful for certain AI tasks, particularly inference on quantized models, it's not commonly used for training most modern AI models. The majority of current AI training occurs at 16-bit floating point or brain float precision.
At 16-bit precision, Oracle's cluster performance drops to about 459 exaFLOPS – still impressive, but not quite reaching the zettascale milestone.
It's worth noting that using lower precision comes at a cost to accuracy. While it can speed up certain operations, it's not suitable for all computing tasks, especially those requiring high precision calculations.
Despite the marketing spin, Oracle's cluster still packs a punch. Even at standard 64-bit double precision (FP64), its peak performance is estimated between 5.2 and 5.9 exaFLOPS. This potentially outperforms the AMD Frontier system by more than three times, though achieving peak performance at this scale is extremely challenging due to interconnect limitations.
Oracle plans to make these Blackwell-based superclusters available in the first half of 2025, joining their existing offerings of H100 and H200 superclusters.
While Oracle's announcement showcases the potential of next-generation AI hardware, it also highlights the importance of understanding the nuances of performance metrics in the rapidly evolving world of supercomputing and artificial intelligence.