Tachyum^® today announced details regarding how performance improvements are achieved for its 2nm Prodigy^® Universal Processor. It will enable AI models with parameters many orders of magnitude more than those of any existing solution at a fraction of the cost.

Prodigy Ultimate provides up to 21.3x higher AI rack performance than NVIDIA Rubin Ultra NVL576. Tachyum DDR5 DIMM (TDIMM) increases bandwidth 5.5 times from 51 GB/s to 281 GB/s. Prodigy Ultimate’s 24 channels provide 8 GB/s, an 8.2x increase over the previous design, and 11x increase over 12-channel CPUs. The TDIMM supports 256 GB in standard, 512 GB in tall, and 1 TB in extra tall height. TSV increases capacity up to 8 times to 2 x 96 TB per Prodigy Ultimate, or up to 3 PB per node.

The TAI data types reduce bandwidth up to 4x, making TAI inference performance like with 27 TB/s of bandwidth. Tachyum 16-socket node provides 16 GB cache to reduce memory bandwidth, which is 130x more than 126MB of NVIDIA B300, and older NVIDIA V100 has HBM memory size comparable to Prodigy’s node caches.

U.S. AI companies have spent billions on hardware. China focuses on efficiency. The Chinese DeepSeek inference is activating 20% parameters, enabling 5x larger models for the same bandwidth and compute. The Chinese Moonshot AI Kimi K2 activates 32 billion parameters out of 1 trillion, enabling 32x larger models for the same processing.

Human brains contain 150 trillion synapses. AI ChatGPT 4 has 1.8 trillion parameters, and BaGuaLu has 174 trillion parameters. The ultimate AI, trained on all the knowledge of humanity, is exceeding 100,000,000 trillion parameters. The Nvidia solution would cost $8 trillion and require 276 gigawatts of power. Tachyum’s solution will cost $78 billion and require 1 gigawatt of power, making it accessible to multiple companies and nations.

OpenAI CEO Sam Altman in September stated plans to build 250 gigawatts of AI data centers by 2033. India consumes 250 GW of electricity emitting twice as much carbon dioxide as ExxonMobil. 60 million GB300 with 288 GB memory for FP4 weights will need 17.3 exabytes of memory to support 34,560,000 trillion parameters. With $50,000 per B300 on the solution level, it would cost $3 trillion.

Using $200 per 48GB DDR5, the 17.3 exabytes (EB) would cost $72 billion. The Tachyum TAI data type reduces that to 8.6 EB for $36 billion and retraining to TAI reduces that to 4.3 EB for $18 billion. The $50,000 24-channel Prodigy Ultimate attaches 6TB of low-cost 256 GB DIMMs. The 4.3 EB of memory requires 720,000 Prodigy Ultimate costing $36 billion, or 360,000 with tall TDIMMs costing $18 billion, or 180,000 with extra tall TDIMMs costing $9 billion. That is 100x less than $3 trillion needed for NVIDIA B300.

At 3 kW per Prodigy Ultimate with DRAM, a 180,000 Prodigy Ultimate would consume 540 MW, or 460x less than the NVIDIA B300 solution. Inference performance of 60,000,000 B300 with 18 petaflops is 1,080 zettaflops. The 180,000 Prodigy Ultimate, each at 400 FP4 petaflops delivers 72 zettaflops, and models retrained for TAI inference deliver 216 zettaflops, which appears to be optimal performance for inference. If needed, the 720,000 Prodigy Ultimates deliver 864 TAI zettaflops for $54 billion and 2.2 GW of energy.

“With Prodigy Ultimate, the unsustainable $3 trillion datacenter consuming 250 gigawatts of power can be shrunk to a $27 billion consuming 540 megawatts in 2028 instead of 2033,” said Dr. Radoslav Danilak, founder and CEO of Tachyum. “With Prodigy Ultimate, humanity can move to an era of AI trained on all written knowledge produced by mankind.”

The Prodigy Universal Processor delivers orders of magnitude higher AI performance, 3x the performance of the best x86 processors, and 6x HPC performance of the fastest GPGPU. Eliminating the need for expensive dedicated AI hardware and dramatically increasing server utilization, Prodigy reduces data center CAPEX and OPEX significantly while delivering unprecedented performance, power and economics.

Those interested in reading the technical details for Tachyum’s latest Prodigy Universal Processor architecture can download the Tachyum Prodigy Feature Upgrade document at https://www.tachyum.com/documentation/tachyum-prodigy-feature-upgrade.pdf

Follow Tachyum

https://x.com/tachyum

https://www.linkedin.com/company/tachyum

https://www.facebook.com/Tachyum/

About Tachyum

Tachyum is transforming the economics of AI, HPC, public and private cloud with the world’s first Universal Processor Prodigy unifying the functionality of a CPU, an HPC GPGPU, and AI accelerators to deliver industry-leading performance, cost and power efficiency. Tachyum has offices in the United States, Slovakia, Taiwan and the Czech Republic. For more information, visit https://www.tachyum.com/.

View source version on businesswire.com: https://www.businesswire.com/news/home/20251117724844/en/

Tachyum today announced details regarding how performance improvements are achieved for its 2nm Prodigy Universal Processor