The CPU specs of 'Tachyum Prodigy', which handles CPU, GPU, and TPU with one chip, are revealed, and the maximum clock frequency is 5.7GHz, 128 cores, and TDP 950W monster specs.
Tachyum has announced the ' Prodigy ' chip, which integrates the functions of CPU, GPU, and TPU into a single architecture. Prodigy's CPU specs are up to 128 cores, operating frequency is 5.7GHz, 16-channel DDR5 memory, TDP is 950W, and technology media Wccftech describes it as 'insane specs'.
Tachyum's Prodigy CPU Specs Unveiled: 5nm Universal Processors With Up To 128 Cores, 5.7 GHz Clocks, 16-Channel DDR5-7200 Memory Support & 950W TDP
https://wccftech.com/tachyums-prodigy-cpu-specs-unveiled-5nm-universal-processors-up-to-128-cores-5-7-ghz-16-channel-ddr5-7200-memory-950w- tdp /
Prodigy is a universal processor that can perform CPU, GPU, and TPU tasks on the same chip. Since all tasks can be performed on the same chip, it has the characteristics of high performance while being able to save costs compared to competing products. In a presentation that announced Prodigy, Tachyum said, 'Prodigy performs four times as much as Intel's Xeon CPU on HPC , three times as much on HPC as NVIDIA's H100 , and six on AI & inference workload. We will achieve double the performance. ' It also emphasizes that it can achieve more than 10 times the performance with the same power compared to the systems of competitors.
The characteristic specifications of Prodigy are as follows.
-Equipped with up to 128 high-performance unified 64-bit cores that operate at up to 5.7 GHz
16-channel DDR5 memory controller
・ PCIe 5.0 has 64 lanes
· 4-socket and 2-socket multiprocessor support
・ A rack solution for data centers that supports both air-cooled and liquid-cooled systems.
-The integer arithmetic performance of SPEC CPU 2017 is about 4 times that of Intel Xeon Platinum 8380 and about 3 times that of AMD EPYC 7763 HPC.
-Double precision floating point performance is about 3 times that of NVIDIA H100
・ The performance of AIFP8 is about 6 times that of NVIDIA H100.
The specifications of each model of Prodigy are as follows.
|Number of cores||Clock frequency||memory||PCIe||TDP||Market segment|
|T16128-AIX||128 128||5.7GHz||16x DDR5-7200||Gen5 x64||950W||High-end AI / HPC|
|T16128AIM||128 128||4.5GHz||16x DDR5-7200||Gen5 x64||700W||Middle range AI / HPC|
|T16128-AIE||128 128||4.0GHz||16x DDR5-7200||Gen5 x64||600W||Entry AI / HPC|
|T16128-HT||128 128||4.5GHz||16x DDR5-6400||Gen5 x64||300W||throughput|
|T864-HS||64||5.7GHz||8x DDR5-6400||Gen5 x32||300W||speed|
|T864-HT||64||4.5GHz||8x DDR5-6400||Gen5 x32||300W||efficiency|
|T832-HS||32||5.7GHz||8x DDR5-6400||Gen5 x32||300W||entry|
|T832-LP||32||3.2GHz||8x DDR5-4800||Gen5 x32||180W||Low power|
Since Prodigy supports two or four multi-sockets, it is possible to build a system with up to 512 cores and TDP 3800W consisting of four T16128-AIX. All Prodigy is manufactured on TSMC's 5nm process node, and the instruction set architecture supports native and x86 / ARM architecture / RISC-V .
Some of the unique features related to HPC and AI include:
2x1024 bit vector unit per core
4096-bit matrix processor per core
-FP64, FP32, TF32, BF16, Int8, FP8, TAI data type
・ Sparse data type optimizes efficiency
・ Quantization support using low-precision numerical type
-Scatter / gather for efficient storage and loading of matrices
Tachyum has air-cooled racks and liquid-cooled racks dedicated to Prodigy, and the T16128-AIX double precision floating point number (FP64) performance is about 90T (tera) FLOPS when using NVIDIA's H100 DGX POD. On the other hand, it is possible to improve the performance up to 6.2P (peta) FLOPS for a dedicated air-cooled rack and up to 12.9PFLOPS for a liquid-cooled rack.
Prodigy is scheduled to be released in the second half of 2022, and the next-generation model 'Prodigy 2' is scheduled to appear in the second half of 2024. For Prodigy 2, Tachyum states, 'With a manufacturing process of 3 nm, we have more core memory memory bandwidth and support PCIe 6.0 and
in Hardware, Posted by logu_ii