AI Updates Aggregator

🌍The Next Web (TNW)•May 13, 2026Freshcollected in 63m

Fractile raises $220m for in-memory-compute inference chip production

🌍Read original on The Next Web (TNW)

#ai-hardware #inference #semiconductors #fundingfractile-inference-chip

💡New hardware architecture aiming to solve the memory bottleneck for LLM inference with major industry backing.

⚡ 30-Second TL;DR

What Changed

Secured $220 million in funding led by Accel.

Why It Matters

This funding signals a shift toward specialized hardware architectures that bypass the memory wall, potentially offering significant latency improvements for large-scale LLM inference.

What To Do Next

Monitor Fractile's public benchmarks against H100s to evaluate if their in-memory architecture fits your specific inference workload requirements.

Who should care:Developers & AI Engineers

🧠 Deep Insight

Web-grounded analysis with 12 cited sources.

🔑 Enhanced Key Takeaways

•Fractile was founded in 2022 by Walter Goodwin, an Oxford PhD in robotics, who developed the concept while researching large language models (LLMs) for general-purpose robots.
•The company projects its chips to deliver AI inference performance that is 100 times faster, 10 times cheaper, and 20 times more energy-efficient than current Nvidia GPUs, specifically for LLMs like Llama2-70B.
•Fractile's in-memory compute architecture utilizes SRAM (Static Random-Access Memory) to integrate memory and compute directly on the same die, thereby mitigating the data transfer bottleneck prevalent in conventional GPU-DRAM systems.
•The startup emerged from stealth in July 2024 with an initial $15 million seed funding round and has committed £100 million to expand its UK operations, including establishing a new hardware engineering facility in Bristol.
•The recent $220 million funding round, co-led by Factorial Funds, Accel, and Peter Thiel's Founders Fund, reportedly values Fractile at over $1 billion.

🛠️ Technical Deep Dive

<ul><li>Fractile's core technology is an in-memory compute architecture designed for AI inference, particularly for large language models (LLMs).</li><li>The architecture integrates compute and memory directly on the same silicon die.</li><li>It employs SRAM (Static Random-Access Memory) to co-locate memory and processing units, aiming to eliminate the 'memory wall' or 'data-shuttling bottleneck' associated with moving data between traditional GPUs and off-chip DRAM.</li><li>This approach is projected to achieve a 100-fold increase in effective bandwidth and significantly higher energy efficiency.</li><li>Fractile claims its accelerators can run LLMs like Llama2-70B 100 times faster, at one-tenth the system cost, and 20 times more energy-efficiently than Nvidia H100 GPUs for decode tokens per second.</li><li>The company is developing custom multiply-accumulate (MAC) circuits that also store state.</li><li>There is speculation that Fractile may optimize MAC arrays for general matrix-vector (GEMV) operations rather than general matrix-matrix (GEMM) operations to enhance efficiency.</li><li>The Fractile team includes experienced engineers from companies such as Graphcore, Nvidia, and Imagination Technologies.</li><li>Fractile is developing its own software stack in conjunction with its hardware.</li><li>Commercial readiness for its chips is anticipated around 2027.</li></ul>

🔮 Future ImplicationsAI analysis grounded in cited sources

Fractile's technology could significantly reduce the operational costs and energy consumption of large-scale AI inference.

By integrating compute and memory on a single die using SRAM, Fractile aims to eliminate the memory bottleneck, leading to substantial improvements in speed, cost, and energy efficiency compared to traditional GPU architectures.

The emergence of companies like Fractile will intensify competition in the AI inference chip market, potentially diversifying the supply chain beyond dominant players like Nvidia.

Fractile's innovative approach and significant funding, coupled with Anthropic's reported interest in diversifying its chip suppliers, indicate a growing market for specialized inference hardware that challenges existing solutions.

Fractile's success could accelerate the development and deployment of more complex 'reasoning models' in AI.

Pat Gelsinger noted that reasoning models are memory-bound and require generating thousands of output tokens, a limitation Fractile's in-memory compute aims to overcome, enabling faster and more efficient execution of such advanced AI.

⏳ Timeline

2022

Fractile founded by Walter Goodwin.

2024-07

Fractile emerged from stealth and announced $15 million in seed funding.

2024-10

Received a $6.52 million grant from the UK government's ARIA program.

2025-01

Pat Gelsinger announced his investment in Fractile.

2026-02

Fractile announced plans to invest £100 million to expand UK operations, including a new hardware engineering facility in Bristol.

2026-05

Secured $220 million in funding led by Accel, Factorial Funds, and Founders Fund, valuing the company at over $1 billion.

📎 Sources (12)

Factual claims are grounded in the sources below. Forward-looking analysis is AI-generated interpretation.

🌍Read original article on The Next Web (TNW)

Weekly AI Recap

Read this week's curated digest of top AI events →

👉Related Updates

Same topic

Explore #ai-hardware

Same product

More on fractile-inference-chip

Same source

Latest from The Next Web (TNW)

UK Startup Fractile Raises $220M for AI Chip Production

Bloomberg Technology•May 13

AI-curated news aggregator. All content rights belong to original publishers.
Original source: The Next Web (TNW) ↗