Tenstorrent has reduced the tensor core count on its Blackhole p150 AI accelerator cards from 140 to 120 cores through a firmware update, claiming only a 1-2% performance impact despite the significant hardware reduction.
Starting with firmware version 19.5.0, Tenstorrent has reduced the tensor core count on its Blackhole p150 AI accelerator cards from 140 to 120 cores, affecting both new cards and existing units already in customers' hands. The change was communicated to customers via email and is documented on the firmware update's GitHub page, though the company has not provided detailed reasoning for the reduction.
{{IMAGE:3}}
The Numbers Behind the Change
The core reduction represents a 14% decrease in hardware resources. When comparing the old and new Blackhole p150 product pages on Tenstorrent's website, the raw compute capacity drops from 774 to 664 TFLOPS. This is a substantial change, particularly for multi-card configurations like the QuietBox system.
Despite this hardware reduction, Tenstorrent claims that typical workloads will see only a "non-material" 1-2% performance difference when using TT-Metalium, the company's low-level developer kit that operates close to the hardware. The company also notes that users may observe changes in grid size in metal, which could require updates to applications that depend on grid layout.
Customer Impact and Reactions
Existing users face a difficult choice: update the firmware and lose cores, or forgo new features of this nascent, in-development product. The change has sparked discussion in Tenstorrent's Discord server, with users questioning whether the SDK fully utilizes the underlying hardware and expressing concerns about how changes to core grid layout and timing could negatively affect existing, tested code.
Predictably, the 14% core reduction has generated jokes and skepticism within the community. Some users have speculated about potential yield issues with the Blackhole chips, though this seems unlikely given that they're manufactured on a 6-nm process—considered mature technology by today's standards.
Product Context and Market Position
Tenstorrent's Blackhole products represent an ambitious entry into the AI accelerator market. The p150 model, even with the reduced core count, can deliver 664 TFLOPS and comes with 32 GB of VRAM for $1,399—approximately one-third the price of a single RTX 5090 while theoretically delivering several times the performance.
The cards can be wired together to pool VRAM, and the chips combine CPU and tensor cores on the same die, bypassing limitations of conventional designs. This architecture represents a significant departure from traditional GPU-based AI acceleration approaches.
Company Background and Future Plans
Tenstorrent, founded by renowned chip architect Jim Keller, has distinguished itself from many AI startups by actually delivering products rather than vaporware. The company is already in conversations with multiple 2-nm chip manufacturers for its next generation of products.
However, Tenstorrent has yet to deliver its upcoming two-chip, 64 GB Blackhole p300 model. The current turbulence with the p150 firmware update highlights the challenges of creating an entirely new AI accelerator architecture, complete with software stack and tooling, particularly when using an unusual chip layout.
Industry Implications
The Blackhole p150 situation illustrates the growing pains of emerging AI hardware companies attempting to compete with established players. While Tenstorrent's approach of targeting small-scale AI accelerators for small businesses, laboratories, and individual developers may carve out a viable niche, the company must balance innovation with reliability and clear communication.
For the broader AI hardware market, this incident underscores the importance of transparency when making significant changes to shipped products. As more companies enter the AI accelerator space with novel architectures, customers will likely demand clearer explanations for hardware modifications that affect performance characteristics.
{{IMAGE:2}}
The Blackhole p150's price-to-performance ratio remains compelling even after the core reduction, but Tenstorrent will need to rebuild trust with its early adopter community. The company's ability to execute on its ambitious roadmap while maintaining product consistency will determine whether it can establish itself as a serious competitor in the specialized AI hardware segment.
Comments
Please log in or register to join the discussion