RISC-V's AI champion just scored $693M cash infusion

RISC-V chip designer Tenstorrent has won $693 million of investment – an endorsement of its plans to use the permissively licensed instruction set architecture for workloads like AI.

The startup announced the funding on Monday, and named Samsung, AWF Partners, and even Jeff Bezos's investment vehicle as investors.

The Santa Clara-based AI infrastructure biz, led by semiconductor guru Jim Keller, has been developing AI accelerators based on the RISC-V CPU instruction set architecture (ISA) going back to 2016.

The open nature of the base RISC-V ISA means that anyone can use it royalty free to develop open source or proprietary close source products or intellectual property, which they are then free to share or license.

Since its introduction in 2010, RISC-V has been adopted in microcontrollers and SoCs, as silicon slingers see it as a fine alternative to licensing an off-the-shelf ISA or CPU core from the likes of Arm or developing something like it from scratch. The ISA has become a favorite among developers of microcontrollers and modest SoCs for embedded applications, like hard disk controllers.

Tenstorrent has championed RISC-V for AI accelerators.

The AI upstart's Blackhole accelerators, discussed at the Hot Chips conference in August, feature a total of 768 RISC-V cores including 16 big CPU cores capable of running Linux and 752 "baby" cores responsible for memory management, networking, system management, PCIe – and the bulk of the chip's compute capability.

Tenstorrent claims the part is capable of delivering peak AI compute performance of 745 teraFLOPS at FP8 (half that at FP16) making it a hair faster than Nvidia's A100 and L40S accelerators – albeit with less memory. Each Blackhole part features 32GB of GDDR6 memory and an Ethernet interconnect capable of 1TB/sec of bandwidth.

However, the chip isn't intended to be a standalone part, but forms the basis for Tenstorrent's upcoming Blackhole Galaxy appliance. That machine will feature 32 Blackhole accelerators connected in a 4x8 mesh for 23.8 petaFLOPS of FP8 performance, 1TB of memory, and 16TB/sec of raw bandwidth.

Tenstorrent's compute platforms have proven compelling enough to convince a wide variety of venture funds to throw their weight behind the business. Of course Samsung's involvement in the venture is a tad self-serving as the South Korean megacorp's foundry unit is one of Tenstorrent's fab partners.

In addition to supporting the construction of new systems and the introduction of an AI development cloud, Tenstorrent indicated the funding will also facilitate work on its open source software stack, enable it to hire more developers, and expand its footprint.

• China sends cloud powered by homebrew Loongson CPUs into space
• AI chip startup Tenstorrent to train Japan's engineers in $50M government deal
• TSMC reportedly cuts off RISC-V chip designer linked to Huawei accelerators
• RISC-V reaches milestone with RVA23 profile ratification

Tenstorrent is not the only org using RISC-V in the crowded AI arena. Earlier this year, we looked at SiFive's Intelligence XM series of clusters – similar to Tenstorrent’s Blackhole, which actually makes use of about 16 of SiFive's RISC-V cores.

The difference is that while SiFive does offer development systems, it is more interested in licensing core IP and SoC blueprints than building and shipping actual hardware itself. Tenstorrent does license its tech to others who want to customize it, but it's clearly keen on building its own end-user hardware products and developing software to harness it.

Similarly, Google has also made use of RISC-V cores from SiFive – specifically the Intelligence X280 to keep the matrix multiplication units (MXUs) in its tensor processing units fed with numbers to crunch. ®