Cerebras Systems

AMD acquires Taalas to hardwire AI into silicon

Advanced Micro Devices (AMD) has agreed to acquire Taalas, a Toronto startup that builds inference accelerators customized—or hard-wired—for a single AI model. Taalas says its specialized chips can produce outputs for specific models thousands of times faster than traditional GPUs in exchange for less flexibility. The company, founded in 2023, has raised $219 million; its current chip runs a small version of Meta's Llama 3.1 and is manufactured on an older TSMC process using on‑chip SRAM. AMD plans to integrate Taalas' technology into its roadmap and systems alongside its CPUs and Instinct GPUs, including rack-scale offerings such as Helios. The acquisition follows a wave of AI-hardware deals, including Nvidia’s purchase of Groq assets for $20 billion seven months earlier, and fits AMD’s broader buying strategy to assemble components for integrated AI systems.

Lovable and Cerebras Announce Inference Partnership

Lovable and Cerebras Systems announced a partnership to run Lovable’s latency-sensitive inference workloads on Cerebras’ high-performance inference infrastructure. Lovable — which says more than 50 million projects have been built on its platform since its November 2024 launch — will use dedicated Cerebras capacity to reduce response times and enable more interactive, multi-step software creation workflows. Cerebras’ Wafer-Scale Engine is highlighted as keeping an entire model’s weights on a single wafer to deliver greater memory bandwidth and faster token generation than GPU-based systems. Both companies said they will jointly explore new product experiences enabled by faster inference; technical results and availability details will be shared later.

AMD Advances: Can It Break the CUDA Moat?

SemiAnalysis analyzes AMD's progress on AI accelerators and software, arguing AMD now has a meaningful chance to compete with Nvidia if it solves two major risks: Helios rack production/backplane SerDes and a chronic shortage of stable internal GPU clusters for CI and developer testing. The piece reviews the MI455X (gfx1250) silicon and Helios rack architecture (2nm, CoWoS-L packaging, 12 HBM4 stacks, Broadcom Tomahawk 6 switches, UALoE scale-up fabric), notes Anthropic’s public 2GW AMD deployment and Microsoft’s reversal to deploy MI455X Helios, and documents substantial ROCm/ROCm.ai software improvements and upstreaming efforts (NIXL->NIXL/NIXL merges, MoRI work). The author warns that AMD's hardware advances must be matched by robust CI, distributed-inference composability (WideEP, MoRI-EP), and reliable rack production to materially erode Nvidia’s lead.