Artificial intelligence computing hardware startup Tenstorrent Inc. is pushing hard to position itself as an alternative to Nvidia Corp.’s graphics processing units after raising a massive $1.5 billion in a late-stage funding round today.
The Series E round was jointly led by Trace Capital and Innovation X and saw participation from new and existing backers including Protagonist, Morgan Stanley, Fidelity Management & Research Co., ARK Invest and T. Rowe Price. It brings the startup’s valuation to $6.2 billion.
Toronto-based Tenstorrent is led by its Chief Executive Jim Keller (pictured), who is widely regarded as a legend in the semiconductor industry, having previously helped to design a number of groundbreaking processors for the likes of Intel Corp., Advanced Micro Devices Inc., Apple Inc. and Tesla Inc.
The startup originally set out as a chip designer but quickly pivoted to offer more than just the bare silicon. Today, it sells a complete portfolio of AI compute platforms, ranging from its Galaxy servers, TT-QuietBox workstations and its Blackhole cards for laptops.
Those platforms are powered by its specialized Tensix Neo accelerators and its Ascalon central processing units, which are both based on the open-standard RISC-V instruction set. They’re designed to be extremely scalable and energy-efficient, capable of producing top-tier performance for AI workloads without runaway power consumption, ensuring costs remain low.
Its best-selling products are the Galaxy Tenstorrent servers, which are enterprise-class systems that can work in coordination with others to create massive superclusters for training and running advanced frontier models. According to Tenstorrent, the Galaxy servers are already being used by companies such as Virtu Financial Inc. to power on-premises AI trading agents, while the data center colocation provider Equinix Inc. also makes them available to enterprises.
Developers who need immense compute locally may be more interested in Tenstorrent’s TT-QuietBox (below), which is a liquid-cooled desktop AI workstation that has enough power to test and run cutting-edge models right out of the box, with minimal setup. These systems are powered by Tenstorrent’s Blackhole cards, which are infinitely scalable PCIe boards packed with massive RISC-V cores and GDDR6 memory to enable high-performance AI processing in a very small package. For larger workloads, it’s possible to network multiple Blackhole cards and TT-QuietBoxes to pool their memory and compute.
Hardware is one thing, but where Tenstorrent really sets itself apart is that it also develops all of the software needed to make its platforms usable. It has also developed an open-source software stack that gives developers the ability to bring any AI workload to its hardware.
This stack consists of TT-Forge, a compiler that bridges machine learning frameworks like PyTorch to its RISC-V chips; TT-Metalium, which is a software development kit for hardware optimization and custom kernel development; and TT-LLK, for low-level kernel optimizations. These three components combine to enable developers to compile, fine-tune and then execute AI models on its custom silicon.
Keller said Tenstorrent’s platform approach allows it to scale from a single workstation to massive cloud-based clusters. “We’ve built one architecture and we sell it at every size,” He explained. “The future of compute is open, scalable and owned by the people building on it.”
Constellation Research analyst Holger Mueller told SiliconANGLE that Tenstorrent is one of a growing number of innovative startups trying to build more efficient computing platforms for AI, and the pedigree track record of its CEO Keller means it’s also one of the most promising. “Tenstorrent gets the funding it needs to keep innovating, and this is good news for enterprises,” the analyst said. “Nvidia is not the only option anymore. More competition and deployment options for AI compute can only be beneficial for enterprises, making AI workloads cheaper and more flexible.”
The funds from today’s round mean that Tenstorrent now has the muscle to ramp up production of its Galaxy servers, which will be deployed globally to support developers running massive large language models and video generation workloads. Some of the money will be given to its research team, which will look to advance innovation in the RISC-V ecosystem. Finally, Tenstorrent plans to expand its workforce globally.
Protagonist co-founder and Managing Partner George Bousis said Tenstorrent’s full platform approach sets it apart from the legions of startups scrambling to provide a viable alternative to Nvidia’s GPU ecosystem. “Few companies design the chip and the software that runs on it, but Tenstorrent does both, and its video generation results beat industry benchmarks ahead of our expectations,” he said.