From Silicon To Software

 

Customer Spotlight: Tachyum’s Universal Processor for Hyperscale Data Centers

HPC processor

By Editorial Team

Social media posts and video streaming. E-commerce transactions and e-mail communications. Online gaming and business applications. These are just a few examples of everyday activities that rely on data centers. As data center usage continues to skyrocket to support increased internet traffic, their impact on the planet’s energy usage will need to be mitigated before it all hits a wall. What’s more, there’s opportunity to make better use of existing data center resources, diverting otherwise idle processing capacity toward various practical pursuits.

Tachyum, with offices in the U.S. and Slovakia, has a solution for this performance plateau: Prodigy, the world’s first universal processor for servers, artificial intelligence (AI), and high-performance computing (HPC).

Targeted for availability later this year, the Prodigy universal processor is designed to run faster, be up to 10x lower power, and deliver up to 4x lower total cost of ownership compared to other processors for this market. Dr. Radoslav Danilak, the company’s co-founder and CEO, notes that extending current data center growth hinges on reducing power consumption of these compute-intensive systems by at least 10x. Data centers currently consume about 3% of the planet’s energy. At their current growth rate of 15%, this number could be an untenable 50% by 2040.

With the capabilities of its Prodigy processor, Tachyum aspires to bring “superhuman brain-scale AI to the mainstream” over the next few years, according to Danilak.

10x More AI and HPC Resources from Advanced, Homogeneous Chip Architecture

To address the energy and performance demands, the Prodigy processor unifies CPUs, GPUs, and specialized chips (such as hardware accelerators) into one advanced, homogeneous architecture. At smaller geometries, more traditional silicon chips are hampered by the slower wires that stem from the increased resistivity of shrinking cross-sections. The architecture of the Prodigy processor jumps over the performance plateau while eliminating wasteful idle time that other devices often experience. For example, let’s consider a typical data center workload for a social media platform. There will be peak periods of utilization in a 24-hour cycle, along with periods where the average utilization could be well under 50%. The servers still need to continue running during these low periods of usage. What if the excess capacity could be directed toward another purpose, such as AI training, so that the server is more fully utilized?

The Prodigy processor, which is scheduled for tapeout later this year, carries the promise of 10x more AI and HPC resources compared to the current baseline. Designed with encrypted memory to protect sensitive data, the processor can be used for a variety of compute-intensive applications: HPC, various types of AI workloads, neural networks, and traditional data center workloads. Its ability to seamlessly switch among different types of workloads allows the Prodigy processor to improve the economics and energy efficiency of hyperscale data centers.

Tachyum has teamed up with PosAm, Towercom, the Slovak Academy of Sciences, and MIRDI (Ministry of Investment, Regional Development and Informatization) on I4DI (Innovations for Digital Infrastructure). I4DI will use the Prodigy processor in the design of what it considers to be the world’s fastest AI supercomputer for the National Supercomputing Centre (NSCC) in Slovakia. The machine is expected to provide 64 AI exaflops and 500 double-precision petaflops of processing power. The universal processor offers the potential for substantial benefits in a variety of application areas, for example:

  • Faster drug and vaccine discovery
  • More accurate climate change modeling
  • Better renewable energy planning
  • Enhanced biofuel catalyst design
  • Better design of low-emission engines

Renewable energy planning with AI

Renewable energy planning can get a boost from AI. 

IP Increases Performance, Lowers Power, and Manages Temperature

HPC SoCs typically host hundreds of large CPUs operating at different frequencies in multiple voltage domains and speeds. To support the multi-voltage, multi-domain requirements and control leakage and power, designers can take advantage of specially tuned logic libraries to achieve their specific PPA goals. In Tachyum’s case, the company has a fast chip that integrates more than four million instances for a high-performance core. Needing a way to keep its processor cool while providing high performance for demanding data centers, the company selected Synopsys DesignWare® Logic Library IP and DesignWare PVT Monitors and Sensors.

Delivering high speed, high density, and low power, DesignWare Logic Libraries provide a broad set of standard cells to optimize the Tachyum processor’s circuits for performance, power, and area. The combination of high-performance compute logic libraries with large muxes, hand-tuned complex combinational cells, optimized multi-bit flops, and 72X and 168X high-drive buffer cells for clock spine implementation resulted in > 4-GHz FMAX with a balanced VT-profile at lower voltage.

The DesignWare PVT sensors offer an embedded, distributed sensor network to address the dynamically changing conditions of the SoC, including in-die process speed, supply variation, and thermal activity. The unique Distributed Thermal Sensor (DTS) enables faster and more area-efficient thermal sensing in areas with multiple hot spots, which are common in this type of complex AI device. The implementation of the PVT Controller also allows for a much quicker and easier integration of the monitors and sensors, which can help speed up time to market.

“As we develop the Tachyum Prodigy processor to deliver industry-leading performance for data center, AI, and HPC workloads, we also challenge ourselves to support a greener era of computing. Reducing power consumption while providing 10x faster performance than leading processors is critical to our success. DesignWare Foundation IP, including logic libraries and PVT sensors, helps us balance our power and performance requirements while significantly reducing our customers’ data center total cost of ownership,” said Danilak. “Startups generally have limitations, but Synopsys has teamed up with us to find solutions that work all around.”

Danilak added, “Having IP with the capabilities of the DesignWare Logic Libraries, available when we needed them even given the disruptions of the pandemic, helped smooth our transition from 7nm to the high-density, high-performance TSMC N5 FinFET process. In addition, the Synopsys support team was great at educating us, and was essential to helping us meet our deadlines.”

As Tachyum plans its next chip, the company looks forward to potentially furthering its collaboration with Synopsys and evaluating Synopsys solutions for design, verification, and prototyping. “Based on our good experience with the DesignWare IP and team, we will look at the broader DesignWare IP and Synopsys portfolio for our next generation,” said Danilak. “The combination of increasing performance and lowering power and cost puts Synopsys ahead of the alternatives.”

AI for All, AI for Good

Tachyum’s vision is to make AI more accessible and enable entities to direct more AI projects toward the good of humanity. Today’s supercomputers cost billions of dollars to build. A universal processor like Tachyum’s Prodigy, however, can reduce costs and power while delivering enough performance to support shared workloads. This opens the door to scenarios such as private enterprises renting unused data center capacity to government agencies pursuing a variety of worthwhile AI projects.

“A revolution to create a machine more powerful than the human brain only happens once,” said Danilak.

In Case You Missed It

Catch up on these recent AI-related blog posts: