Zettascale10: Oracle announces fastest supercomputer cluster

In the race for the world's fastest supercomputers, Oracle is entering a new preliminary benchmark: "Zettascale10" is expected to achieve up to 16 Zettaflops, which translates to 16.000 Exaflops, or 16 quintillion operations per second, a 16 followed by 21 zeros.

However, Oracle is using a trick for this immense number. The value applies to simple 4-bit floating-point values (FP4), which are sufficient for many AI algorithms. It is not comparable to supercomputers on the Top500 list, whose benchmarks are generated with 64-bit operations (FP64). With perfect scaling, Zettascale10 would achieve 32 Exaflops, which would still be the world's fastest supercomputer by a huge margin. The fastest individual supercomputers today achieve one to three FP64 Exaflops.

Cluster of Data Centers

Zettascale10 is not a single data center, but a network of several campuses within a two-kilometer radius in Abilene, Texas. The Stargate supercomputer, which is already partially completed, with a future 450.000 Nvidia Blackwell GPUs, is the flagship in the Zettascale10 network. Oracle operates it together with OpenAI.

The first completed Zettascale10 iteration is expected to use a total of 800.000 Nvidia GPUs and consume several gigawatts of electrical power. The cluster is scheduled to be operational in the second half of 2026.

Crude Extrapolation

For the announced 16 Zettaflops, Oracle simply sums up the FP4 computing power of each Blackwell accelerator (20 Petaflops each). The actual performance depends on factors such as latency between the systems.

For networking within the server racks, Oracle relies on Nvidia's network technology. Ethernet connects the data centers at a higher level. The GPUs access memory via Remote Direct Memory Access (RDMA) over Converged Ethernet (RoCE).

(mma)