Tachyum’s Universal Processor Still Promises to Run CPU, GPU, and TPU Functions on the Same Chip
High-performance computing (HPC) maker Tachyum is bringing a server and data center processor to market with specifications to compete with AMD and Intel. The Prodigy features 128 cores, 5.7GHz clock and 16 channels of DDR5 memories. In addition, the company claims that Prodigy is the “world’s first universal processor”.
Capable of executing CPU, GPU and TPU (Tensor Processing Unit) tasks on the same chipset, depending on the need, the Prodigy was baptized as ‘universal processor’. It comes to the server and datacenter market to compete with the latest and greatest hardware solutions from Nvidia and Intel such as Nvidia’s H100 GPU as well as AMD’s Instinct M1210.
Performance and weight PTO
The line will feature 8 models with different configurations, with the top-of-the-line version being the T16128-AIX, with 128 cores, 5.7GHz and 16 channels of DDR5 memories. The company claims that the Prodigy has 4 times the performance of an Intel Xeon 8380 and 3 times the performance of the AMD 7763HPC. According to manufacturer data, the top-of-the-line version T16128-AIX offers up to 90TFLOps of processing power, which can be allocated in racks up to 4 processors.
Lithography is 5nm from TSMC with up to 32TB of DDR5-7200 RAM in 16 channels and 64 PCle 5.0 lanes. Due to the weight settings the Prodigy also has considerable power consumption: TDP of 950W. Despite this, as the module has a ‘universal function’ the high power consumption would still be an advantage over the combination of CPUs and GPUs needed for data centers.
– Continues after advertising –
The manufacturer also reported that the chips will be able to translate code from other architectures, in addition to the native one from Tachyum, such as x86, ARM and RISC-V. As we can see, the intention is really to have a ‘universal processor’.
To achieve translation and reading in different architectures, the chipset supports calculations in FP64, FP32, TF32, BF16, INT8, FP8 and TAI formats, in addition to being compatible with processes to speed up the processing of data from matrices for AI, such as ‘Scatter’ and ‘Gatter’.
Until the moment, there is no information if performance on architectures like x86 will affect the performance of the Prodigy line, but either way it’s an interesting solution. As it is a product intended for servers and datacenters, Tachyum did not reveal the official prices of the Prodigy line. In the schedule presented by the company, the line will be launched from the second half of 2022, with plans to launch the 2nd generation with more 3nm cores and lithography in the second half of 2024.
Thinking about buying a product online? Discover the Adrenaline Save extension for Google Chrome. It is free and offers you price comparisons at major stores and coupons so you can always buy at the best price. Download now.
Source: Tweaktown, Tachyum, Wccftech