Editorial illustration for AMD Buys Taalas, a Startup That Bakes AI Models Into Silicon
AMD Buys Taalas: AI Models Baked Into Silicon
AMD announced it's buying Taalas, a Toronto startup that takes a strange approach to running AI models: instead of loading a model onto a chip, it burns the model's architecture and trained parameters directly into the silicon itself. Taalas only came out of stealth in February 2024, but its pitch got attention fast. Because the model is physically part of the chip, inference runs at speeds standard GPUs can't touch, though each chip only works for the one model baked into it.
The company's demo chip ran Llama 3.1-8B at more than 16,000 tokens per second per user, a figure that dwarfs typical hardware benchmarks. Google is said to be pursuing a similar design for Gemini, suggesting this isn't a one-off idea but a real direction chipmakers are betting on for inference-heavy workloads.
For AMD, the deal adds a specialized tool to sit alongside its Instinct GPU lineup, aimed at customers who want raw speed for a fixed model rather than flexibility across many. The purchase still needs to clear standard regulatory review before it closes.
AMD is buying Canadian AI startup Taalas, which builds specialized inference chips. Founded in Toronto in 2023, Taalas came out of stealth in February with an unusual approach: the company embeds a model's architecture and trained parameters directly into the chip. That makes inference extremely fast but locks each chip to a single model.
Why this matters
Baking a model's weights directly into silicon is a bet that speed beats flexibility, at least for the workloads that matter most. A chip hitting 16,000 tokens per second per user on Llama 3.1-8B is a real number worth watching, but it comes with a catch: that silicon is useless the moment you want to swap in a newer model. For AMD, buying Taalas looks like a hedge against Nvidia's general-purpose GPU dominance in inference, not a replacement for it.
For founders and infra teams, the question is cost. Fixed-function chips make sense if you're running one model at massive scale for years; they make far less sense if you're iterating on architectures every few months, which is most of the industry right now. Toronto's Taalas built something clever, but "locks each chip to a single model" is the kind of tradeoff that sounds fine in a demo and gets expensive in production.
Worth tracking whether AMD turns this into a real product line or just mines it for patents and talent.
Related Reading
Business & Startups Nordic pilot adds Gemini for Education, NotebookLM to boost AI literacy
Business & Startups Kling launches Video O1, all-in-one model with MVL bridge using transformer
Business & Startups DeepSeek Seeks More Capital Weeks After USD 7B Funding Round
From Archives Cloudflare Launches Kitesurf, a Browser Designed for AI Agents
From Archives ByteDance Develops China's Largest AI Model, Currently in Pretraining
Common Questions Answered
What is Taalas's unique approach to running AI models compared to traditional GPUs?
Taalas embeds a model's architecture and trained parameters directly into the silicon chip itself, rather than loading the model onto a chip during runtime. This approach enables inference to run at significantly faster speeds than standard GPUs can achieve, with the company demonstrating 16,000 tokens per second per user on Llama 3.1-8B.
What is the main trade-off of having AI models baked directly into silicon?
While baking models directly into silicon provides extreme speed advantages, each chip becomes locked to a single model and becomes obsolete when you need to switch to a newer or different model. This represents a bet that speed is more valuable than flexibility for specific high-priority workloads.
Why did AMD acquire Taalas and what does this acquisition represent?
AMD's acquisition of Taalas appears to be a strategic hedge against Nvidia's dominance in general-purpose GPU inference rather than a replacement for it. By acquiring Taalas, AMD gains specialized inference chip technology that could compete in specific high-speed inference scenarios where model flexibility is less critical.
When did Taalas emerge from stealth and when was it founded?
Taalas was founded in Toronto in 2023 and came out of stealth in February 2024, quickly gaining attention for its innovative approach to embedding AI models directly into silicon. The startup's pitch resonated with industry players fast enough to attract AMD's acquisition interest within months of its public launch.
Further Reading
- AMD buys Taalas, startup that hardwires AI models into its silicon - CNBC
- AMD Acquires Taalas to Advance Compute Solutions for Rapidly Growing AI Inference Market - AMD Investor Relations
- AMD acquires AI chip startup Taalas to boost inference performance by etching models into silicon - The Register
- AMD Buys AI Chip Startup Taalas—Here's Why It Matters - Yahoo Finance
- AMD Buys Startup Taalas To Bake AI Models Straight Into Silicon - HotHardware
