Tachyum said that one of its customers plans to purchase hundreds of thousands of Prodigy processors to create a powerful supercomputer with a performance of 50 Exflops. Typically, such indicators are associated with AI calculations using 8- or 16-bit mathematics with a floating point, and not with 64-bit calculations, which are usually used in high-performance computers. However, tachyum claims that the system will be 25 times more powerful than” the fastest supercomputer built this year. “
Probably, this is a mention of the recently put into operation by the Aurora supercomputer from the National Argonne National Laboratory with a performance of more than two ex -Flops. Tachyum also states that their future system will be able to reach up to eight Zetttflops for processing volumetric language models and have a DDR5 memory of hundreds of Petabyt by 2025.
In order to better understand the scale of Tachyum plans, it is important to pay attention to their Chip Prodigy, which the company has been working on in recent years. Prodigy is presented as a universal processor that can cope with various tasks. To support modern standards, such as X86, ARM and RISC-V, an qemu emulator was adapted to the architecture of Prodigy.
CEO Tachyum, Radoslav Danilak, said that most of the 600 mm ProDigy crystal is dedicated to its 192 64-bit processor nuclei. These nuclei have their own architecture of a set of instructions and will execute four OUT-OF-Order instructions on each clock cycle at frequencies over 5 GHz.
According to Danilak, the 50 exseflops supercomputer will be absolutely gigantic. “This is a huge, huge car,” he said. If you believe the words of Tachyum, to create such a machine will require about half a million chips.
It is worth noting that despite the ambitious plans of Tachyum, the Prodigy chip has not yet been put into production. However, the company remains optimistic and hopes for the successful completion of the project in the coming years.