We have a total of 128 5.7 GHz cores that handle 64-bit instructions, and there are a total of sixteen controllers (channels?) For 7,200 MT / s DDR5 memory (max. 8 TB per socket) and 64 PCIe 5.0 interface lines, or all modern equipment. Tachyum also boasts that Prodigy handles the role of a simple processor, HPC processor and is also suitable for AI or graphics rendering using ray tracing. However, we get a more specific idea thanks to the list of supported data formats: FP64, FP32, TF32, BF16, Int8, FP8 and TAI, while the chip contains 1024-bit vector and 4096-bit matrix units and handles 4 out-of-order instructions per clock. The equipment is completed by more than 128 MB L2 + L3 cache.
We also learn that the consumption of this processor should be only a tenth compared to “traditional hardware” and the total costs only a third, but these are only very general and meaningless numbers when we do not know exactly what they relate to. We won’t learn much from the claim that the Prodigy is four times more powerful than the best Xeon on the market. In addition, the manufacturer boasts that it should have three times higher gross performance in HPC applications.
The Prodigy T16128 Universal Processor is due to hit the market sometime next year, when production begins, so Tachyum may not be ready to release more detailed tests yet and is currently limited to general performance and performance statements.
However, we know that this is a 5nm chip, the manufacturer of which has not been revealed (but it is only offered by TSMC or Samsung), which is located on an FCLGA case measuring 64 x 84 mm. In terms of performance, it will offer “12 AI PetaFLOPS and 90 TeraFLOPS for HPC load” and will be able to run application binaries written for x86, ARM, RISC-V and ISA.
In addition to the T16128, Tachyum is also preparing weaker versions of the T864 and T832 with 64 and 32 cores.