In The News

As Tachyum continues towards tape-out for Prodigy, the company has chosen to establish manufacturing, testing, assembly, ODM, logistics and support facilities in Taiwan to ensure worldwide shipments of the completed product without the need for intermediaries.

The biggest news is that Tachyum’s Prodigy processor will adopt a multi-chiplet design and each compute chiplet within that system-in-package (SiP) will feature 256 universal cores.

With a total of more than $300 million invested over three funding rounds, Tachyum is poised to revolutionize AI data centers by solving the most challenging barriers to AI and expects that Series C will propel the company to an IPO, potentially as early as 2027.

Tachyum’s AI team has demonstrated that a foundation model fine-tuned on a task-specific dataset in FP4 data type, which represents value using 4-bit floating-point format rather than standard FP32 or BF16, achieves parity with traditional FP32 training baselines.

Tachyum’s engineers ported Kprobes (Kernel Probes), which plays an important role in the eBPF JIT technology and serves as trigger for eBPF subroutines.

Tachyum is at the forefront, pushing the boundaries with its Prodigy Universal Processor which promises unparalleled versatility across various applications. A recent development has taken center stage with the successful port of the eBPF Just-In-Time (JIT) compiler to their software emulation platform.

The white paper explains that MoE can achieve performance comparable to dense models while using approximately four times less computing power and memory bandwidth, though memory capacity requirements increase by about four times.

Tachyum today announced the release of a new white paper detailing how it efficiently scales Large Language Model (LLM) training and inference through the Mixture of Experts (MoE) approach. The company’s method is further improved by a DeepSeekMoE architecture with 4-bit FP4 data types for activations quantization and 2-bit Tachyum AI (TAI2) sparse weights quantization.

Validating SMM shows that Tachyum’s Prodigy is capable of running in this distinct and isolated processor environment, independently of the operating system and transparently to executive and software applications.

Tachyum’s membership connects the company with key players in the industry driving the future of AI infrastructure. Tachyum looks to extend scale up of its systems in future designs through the incorporation of UALink switches.

Tachyum’s DRAM Failover technology offers superior memory error correction, providing a higher level of protection than traditional Error Correction Code (ECC).

A single Prodigy processor with 640 or 1280 DRAM chips attached would mean 64,000,000 DRAM chips, a significant scale. With DRAM Failover correct, a failing DRAM die per DIMM would not affect the operation of the system and will not cause failure with Prodigy, unlike GPU accelerators.

The distribution supports up to 1,024 cores in 4-socket configurations, ensuring Prodigy is ideal for a variety of enterprise applications such as supercomputing, artificial intelligence, cloud computing, data analytics, and web hosting.

The company’s software distribution package combines key advancements to deliver optimal performance for Prodigy. Notable updates to RC1 include the inclusion of GCC 14.2, Linux 6.12, LLVM 19, FreeBSD 14.1 and QEMU 9.11. The package also provides users with an optimization guide and performance-tuning tools to ensure maximum efficiency.

Adding the quantum-safe PQC algorithms, approved by the US National Institute of Standards and Technology as a global standard last August, ensures data centre deployments using the company’s universal processor are quantum-resistant and future-proofed for data security.

Tachyum, a leading player in technology innovation, has made significant strides in enhancing data security against quantum computer threats. The company recently announced its successful integration of the National Institute of Standards and Technology’s (NIST) quantum-safe algorithms into its Prodigy software distributions. This integration is pivotal for organizations aiming to fortify their data infrastructures against emerging quantum threats.

Major companies in Artificial Intelligence (AI) Hardware Market are: NVIDIA, Intel, Google, AMD, Qualcomm, IBM, Apple, Samsung, Huawei, Baidu, Tesla, Microsoft, Arm, Xilinx, Cisco, MediaTek, Graphcore, Cerebras, Alibaba, Broadcom, Hailo, Cambricon, Wave Computing, Mythic, Kneron, Sambanova, Tachyum, Tenstorrent, Groq.

Tachyum has released a 1,600-page guide for optimizing the performance of its Prodigy Universal Processor FPGA hardware. Even though the company has yet to tape out its Prodigy processors after years of delays, it has released a performance optimization manual for the chips, which have a unique instruction set architecture and optimization strategies, well before actual products start sampling or hit the market.

Tachyum’s Prodigy Performance Optimization Manual provides detailed information on how to fully benefit from the performance features that are built into Prodigy. It includes the required design guidelines for the development of high-performance software for a broad range of applications, including cloud, AI, and HPC.