Amir Salek

Prior to joining Google, Salek spent approximately eight years at NVIDIA as a Senior Director of Engineering. During his tenure at NVIDIA, he founded and led the company's System-on-a-Chip (SoC) Design organization and contributed to the development of NVIDIA's GPU and Tegra product lines.^[3]

In 2013, Google recruited Salek to establish and lead custom silicon development capabilities for the company's datacenters.^[5] As founder and head of Custom Silicon for Google Technical Infrastructure and Google Cloud, Salek directed the development and deployment of multiple generations of specialized processors. According to the Financial Times Tech Tonic podcast, Salek was "the brains behind Google's TPU chip," playing a key role in developing the TPU architecture specifically designed for AI tasks.^[2] Under his leadership, the organization developed Google's first production custom silicon chip and created the Tensor Processing Unit family, which includes:

• TPU (Google's first production custom silicon chip)
• TPUv2 (the industry's first production deep-learning training chip)
• TPUv3 (an enhanced training processor with doubled matrix multiply units and improved memory bandwidth)
• TPUv4 (delivering more than 2x performance improvement over TPUv3)
• Edge TPU (a compact deep learning accelerator for edge devices)
• VCU (Video Coding Unit)
• IPU (Image Processing Unit)
• OpenTitan (an open-source security chip)

Throughout his tenure at Google from 2013 to 2022, Salek oversaw the strategic direction of custom silicon development for machine learning, networking, graphics, video, and mobile processors.^[4] In recognition of his role in advancing open-source hardware standards, Salek spoke at the CHIPS Alliance inaugural workshop in 2019 on the "Vision for Open Source Hardware Developments."^[6] His organization's work established custom silicon design as a core competency at Google, and the TPU architecture became instrumental in the company's infrastructure and influenced the broader industry's adoption of specialized AI accelerators.

The TPUv1 development was documented in the seminal 2017 paper "In-Datacenter Performance Analysis of a Tensor Processing Unit," presented at the 44th International Symposium on Computer Architecture (ISCA 2017), which demonstrated significant performance and energy efficiency advantages of purpose-built machine learning accelerators over general-purpose processors.^[7] The subsequent development of TPUv2 and TPUv3 was detailed in the 2021 IEEE Micro paper "The Design Process for Google's Training Chips: TPUv2 and TPUv3," authored by researchers including Salek, which described the co-design philosophy and architectural decisions that enabled rapid development and deployment of these training accelerators at scale.^[8]

In March 2022, Salek joined Cerberus Capital Management as a Senior Managing Director and Partner of Tracker Ventures, Cerberus' stage-agnostic venture capital platform focused on investments in deep technology companies.^[9] In this role, he leads strategic investments in sectors including semiconductors, edge computing, artificial intelligence and machine learning, 5G infrastructure, autonomous systems, aerospace, next-generation compute platforms, cybersecurity, and Web 3.0 technologies.^[9]

Salek's work in establishing Google's custom silicon capabilities demonstrated the viability of rapid application-specific integrated circuit (ASIC) development and deployment at hyperscale, establishing a model that has since been adopted across the technology industry. The TPU family has become foundational to Google's infrastructure and has influenced the broader industry's shift toward purpose-built processors for machine learning and artificial intelligence workloads.^[2]