According to The Information, Tensorrent, a chip company led by Jim Keller, was valued at $3 billion in its latest funding round.

About Tenstorrent

Founded in 2016, Tenstorrent is dedicated to developing scalable AI accelerators for NVIDIA GPUs, supporting cloud and edge computing. The company is currently developing a RISC-V CPU and licensing its design to other entities. Notably, the company heavily utilizes open-source technologies in its development process, thus avoiding the use of NVIDIA's currently expensive High Bandwidth Memory (HBM).

Tenstorrent's first commercially available chips were manufactured in partnership with GlobalFoundries. Tenstorrent's Tensix processor consists of processor cores called Tensix Cores. Each Tensix Core includes an array math unit for tensor operations, a SIMD unit for vector operations, an on-chip network (NoC) for transferring data between cores and between chips, five small RISC-V processors to assist in controlling the NoC, and up to 1.5MB of SRAM.

Tenstorrent's RISC-V alternative, Grayskull, aims to significantly simplify programming and improve scalability. Grayskull is the company's first Tensix processor, featuring up to 120 Tensix cores, each equipped with 1MB of SRAM, supporting 8GB of LPDDR4 memory on a 256-bit bus, and simultaneously supporting common AI precision formats (FP8, FP16, BF16) and memory-optimized precision formats (BFP2, BFP4, BFP8).

Wormhole and Grayskull

Wormhole is a smaller and improved version of Grayskull's chip. While the number of Tensix cores is slightly reduced (to 80), the SRAM capacity of the Tensix cores has increased to 1.5MB, and support for more precision formats (FP32 output, INT8, INT32 output, and TF32) has been added. To date, most of the company's orders have come from IP transactions that have improved the overall performance and efficiency of existing formats, thus compensating for the reduced core count. Furthermore, local memory capacity has been increased to 12GB of high-speed GDDR6, and Wormhole is scalable to multi-chip deployments.

The Tenstorrent processor uses a mesh-based architecture composed of Tensix cores, designed to efficiently handle tensor computations of various scales. Each processor is equipped with integrated network communication hardware, enabling direct inter-processor communication over a network without relying on DRAM.

In a media interview this September, Tenstorrent stated that it has begun productizing its RISC-V CPU and AI cores as licensable IP. Aniket Saha, Vice President of Product Strategy at Tenstorrent, revealed that the company already has IP licensing customers including LG and Hyundai.

Tenstorrent will also license its intellectual property to companies planning to develop chip products that compete with its workstation and data center hardware products. "Customers typically use slightly different designs," they stated. "Even within the same market, they can develop complementary products."

Tenstorrent emphasizes that the differentiating advantage of its IP products lies primarily in its open-source approach and the ability to customize its cores. Typically, IP companies do not allow modifications beyond configuration options, as this would make IP maintenance extremely difficult.

The first-generation Tenstorrent RISC-V CPU Ascalon IP will be productized with performance implementations of 10 to 20 SPECint2006/GHz. Today's announcement features the two most powerful and third most powerful products. Saha stated that the most powerful Ascalon-X will compete with Arm's Neoverse V2 and V3 cores. The third most powerful Ascalon-S will roughly compete with the Arm Cortex-A78.

Validated and verified subsystems, along with configurable clusters of 2 to 8 Ascalon kernels, will also be available. The Ascalon interface is designated as a direct replacement for Arm (the kernel has both consistent and inconsistent interfaces).

Tenstorrent is pushing forward with its core release cycle of one generation every 18 months, with the next-generation Babylon expected to improve performance, frequency, and PPA.

The automotive version of this CPU, named Alexandria, adds a feature set for automotive functional safety. This also opens up markets in industries such as industrial and medical.

Tenstorrent stated, "Arm has historically not ventured into the industrial and medical markets because order volumes are typically low. We are looking for ways to integrate some of those orders." A development board called Atlantis will feature an 8-core Ascalon-X CPU with a TDP of 50W. This board will be used by OEMs, Tier 1 automotive suppliers, and independent software vendors (ISVs) for porting software requiring high single-threaded performance.

Tenstorrent indicated that the Atlantis SoC may eventually be commercialized for use in AI PCs, but this is still under discussion.

Meanwhile, Tenstorrent's next-generation Tensix AI core, Tensix-Neo, will be released.

The next generation of Tensix-Blackhole

Compared to the current generation of Tensix Blackhole cores, Neo employs a cluster architecture, sharing memory and NoC among four cores, rather than each core sharing memory and NoC, to improve area efficiency and utilization.

"This is the first time we've known in advance that we're going to sell it as intellectual property, so we built it in a more modular and flexible way," Tenstorrent said.

This flexibility is also reflected in specifying L1 cache size and data type support.

Similar to the Ascalon CPU core, customers will handle the relatively simple configurations, while Tenstorrent will handle customizations beyond these parameters. Customers who wish to perform such customizations themselves can purchase an innovation license.

"(For innovation license customers), we provide the RTL code, verification infrastructure, and our built test platform, and then they are responsible for the verification," Tenstorrent said. "But this is not easy, so we carefully select projects to collaborate on."

Source: Content compiled from Semiconductor Industry Observer

View more at EASELINK