
Tachyum runs x86-64 binaries on FPGA model
Chip designer Tachyum has hit a key milestone by running non-native x86_64 application under Linux running on a model of its Prodigy chip on an FPGA emulation system.
The ability to run native x86 binaries as well as a range of AI frameworks is key for Tachyum for its the Prodigy Universal Processor when it becomes generally available next year.
x86 accounts for the widest installation base for datacentre workloads today. Tachyum, based in Slovakia and the US, is aiming the Prodigy chip as a way to run control, data and AI apps on the same chip, sidestepping the dominance of Intel and AMD for control and Nvidia for AI accelerators. The aim is for Prodigy data centre servers to dynamically switch between computational domains such as AI/ML, HPC, and cloud in a single homogeneous architecture.
This eliminates the need for dedicated AI hardware and can dramatically increase server utilization, reducing both OPEX and CAPEX.
Tachyum’s architecture does not include any x86 specific hardware, as this would limit Prodigy’s performance. Instead, a standard dynamic binary translator provides the ability to run unmodified Linux x86 binaries out of the box. Prodigy allows users to mix x86 applications with native Prodigy applications by seamlessly running Prodigy native Apache web servers combined with x86 Linux binary databases.#
Tachyum customers can initially use a dynamic binary translator while they port their legacy x86 applications to Prodigy native ISA, usually within 12-18 months.
- Tachyum designs 192-core AI/ML processor chip
- Tachyum plans 3nm Universal processor
- Tachyum teams up with Cologne Chip on FPGA, AI
Until customers migrate to fully native software environments, Prodigy will deliver twice the performance expected instead of the anticipated 3x performance versus Intel high-end CPUs on SPECint2017 rate. However, for most customers, only 20% of runtime will be legacy x86 applications, with the remaining 80% of runtime dedicated to native applications such as TensorFlow, PyTorch, databases which Tachyum has already ported to native Prodigy ISA. As a result the combination of the x86 binaries and native Prodigy code would give a 2.8x performance boost.
“Demonstrating the ability to run x86-64 binary applications on the Prodigy processor emulation is a key milestone for Tachyum and further validates our architecture before tape out,” said Dr. Radoslav Danilak, founder and CEO of Tachyum. “Proving all these existing binary applications on our FPGA emulator is critical in order to get Prodigy ready for volume production next year, allowing us to begin fulfilling the billions of dollars of orders we already have in our sales pipeline.”
- Tachyum sues Cadence for $211m over failed IP
- Tachyum settles lawsuit with Cadence
- Tachyum shows design of 20 ExaFlop supercomputer
The company has also expanded its software ecosystem with support for LLVM for AI and Linux Rust.
The LLVM Project is a collection of modular and reusable compiler and toolchain technologies. Rust is a multi-paradigm, general-purpose programming language with an emphasis on performance, type safety and concurrency, and has become the second language officially accepted for Linux kernel development.
LLVM plays a large role in every major AI framework, including PyTorch and Tensorflow, with AI compilers based on LLVM for native instruction generation. Standalone AI compilers such as Apache TVM are also based on the LLVM compiler infrastructure.
An LLVM backend supports the Tachyum ISA, particularly vector and tensor instructions in both full and low precisions, allowing it to take full advantage of Prodigy’s hardware features that are specifically designed to boost AI workload performance. This AI compiler backend provides the ultimate performance for inference applications.
Rust for Prodigy
Rust provides both fast and memory efficient code with thread safety. Rust is being incorporated into Prodigy’s Linux kernel, relying on LLVM. Support for GCC Rust will come later when the language becomes readily available.
“As impressive as the hardware architecture of the Tachyum Prodigy Universal Processor is, it is the software that we support that will allow customers and partners to unlock the full potential of the chip,” said Danilak.
“With initial availability of LLVM planned for the first quarter of 2024, I am proud of our software engineering team’s efforts to have it available ahead of schedule. Our team is now hard at work building Tachyum software distribution images, and a quarter after the alpha release, a beta version will be available for our customers to test.”
