# AMD Related Articles

HTX News Center provides the latest articles and in-depth analysis on "AMD", covering market trends, project updates, tech developments, and regulatory policies in the crypto industry.

AMD acquires Taalas: hardware AI manages without scarce HBM memory

AMD has agreed to acquire Toronto-based startup Taalas, which tackles a key bottleneck in AI inference: the constant need to transfer model weights from memory to the processor for each generated token. Taalas's chips eliminate this operation by permanently embedding the model weights into the transistors themselves. This data transfer is what currently limits inference speed and has made high-bandwidth memory (HBM) a scarce commodity. Taalas's first test chip, fabricated on TSMC's 6nm process, reportedly generated tokens for Meta's Llama 3.1 8B model at speeds 48 times faster than comparable Nvidia GPUs. Its architecture features a mask ROM section for fixed weights and SRAM for adaptable components. However, this design comes with a significant trade-off: each chip is permanently dedicated to a single model. Switching models requires a partial redesign and fabrication, a process taking about two months. While the acquisition is seen as part of AMD's rivalry with Nvidia in inference, its broader implication lies in challenging the assumption of a permanent HBM memory shortage. The AI memory market is currently booming, with HBM supply sold out through 2026. Yet, Taalas's technology demonstrates that the memory bottleneck is an engineering challenge, not an absolute physical constraint. This aligns with industry-wide efforts from companies like Nvidia (through model compression) and memory makers like Samsung and SK hynix (developing new packaging and storage technologies) to reduce dependency on scarce HBM. AMD's move suggests that the current high prices for memory, driven by AI demand, may not be sustainable. It highlights a growing engineering push against the premise of perpetual memory scarcity, reminding investors that memory has historically been a cyclical business.

cryptonews.ru08/09 14:56

AMD acquires Taalas: hardware AI manages without scarce HBM memory

cryptonews.ru08/09 14:56

Kimi K3, which used to require 16 B200s, now fits on just 8 AMD cards

This article highlights a key achievement for AMD in the AI hardware race. The company's MI355X GPUs, each with 288 GB of memory, successfully deployed the massive 2.8 trillion parameter Kimi K3 model on a single 8-GPU server. In contrast, the NVIDIA B200 (with 192 GB per card) required a two-server, 16-GPU setup to hold the model, leading to inter-node communication overhead. In performance tests for a 1024-input/400-output token task, the 8-card MI355X system achieved a total throughput of 952 tokens/s and a single-user generation speed of 118 tokens/s. This single-node throughput was approximately 3.8 times higher than the per-node average of the dual-node B200 setup (498 tokens/s total). While NVIDIA's B300 delivered higher absolute performance (1568 tokens/s on 8 cards), a cost-efficiency analysis based on assumed hourly rates showed MI355X offered better value per dollar. Notably, the deployment on AMD's ROCm software platform was relatively straightforward, requiring only minor fixes like patching a missing function for speculative decoding and a simple zero-padding workaround to optimize a prefill kernel for attention heads. This significantly reduced the Time-To-First-Token (TTFT). The article concludes that for extremely large models, memory capacity is becoming a critical differentiator. AMD's strategy of equipping cards with more HBM memory provides a tangible system advantage in single-node deployment efficiency and cost, posing a growing challenge to NVIDIA's CUDA ecosystem dominance.

marsbit08/04 11:22

Kimi K3, which used to require 16 B200s, now fits on just 8 AMD cards

marsbit08/04 11:22

Bernstein Reveals Details of Core Scientific's $14 Billion Deal with AMD

Analysts from Bernstein revealed details of a deal between Core Scientific and AMD with a potential total value of over $14 billion. According to the report, initial contracts for 530 MW of capacity could generate this revenue over 15 years, with AMD acting as a credit guarantor for part of the bitcoin miner's infrastructure. The partnership, announced on July 28, has the potential to allocate up to 2.5 GW of data center capacity for AI. Bernstein broke down the 530 MW into 377 MW of direct triple-net lease for AMD and 152 MW for an unnamed cloud provider backed by AMD's credit. This structure is seen as lowering financing costs and counterparty risk. AMD also received warrants to buy 30 million Core Scientific shares at $23.47 each, which vest upon reaching the 2.5 GW target. Average annual revenue from the deal is estimated at around $0.9 billion, or about $1.8 million per megawatt, which is 5-25% below recent AI hosting deals by other miners. However, the 377 MW triple-net lease for AMD carries a margin close to 100%. Core Scientific expects capital expenditures for the deal to be $11-12 million per MW, totaling about $6 billion. Bernstein views this partnership as a new phase in the transformation of former bitcoin miners into AI infrastructure operators, with AI chipmakers like AMD now acting as direct anchor tenants. Recent similar deals include Hut 8 allocating 704 MW to a tenant believed to be Nvidia, and AMD reserving 200 MW with Riot Platforms. Core Scientific also paid Block $41.9 million to terminate a mining chip supply contract as part of its accelerated diversification into AI.

cryptonews.ru07/30 16:11

Bernstein Reveals Details of Core Scientific's $14 Billion Deal with AMD

cryptonews.ru07/30 16:11

NVIDIA's 20-Year CUDA Moat Collapsed Over a Weekend, Claude Single-Handedly Got AMD's New GPU Running

In a single weekend, Claude, an AI agent from Anthropic, successfully ported and optimized its cutting-edge model to run on a brand-new AMD MI355X server rack without any manual code intervention. This feat demonstrates a potential breakthrough in overcoming NVIDIA's long-established CUDA software ecosystem dominance, built over two decades. Anthropic's team simply instructed Claude to get the AMD machine running. By Monday, it not only worked but was showing a continuously improving performance curve. The achievement impressed AMD CEO Lisa Su and accelerated a major deployment partnership: Anthropic plans to deploy up to 2GW of AMD Instinct GPUs starting in 2027. The key enabler is AMD's new ROCm.AI platform, a toolbox designed specifically for AI agents like Claude. It provides AI-readable documentation, including chip instruction sets (ISA), and tools like the Hyperloom service that allows agents to autonomously profile performance, identify bottlenecks, test configurations, and generate optimized kernels. In a demo, Hyperloom boosted the output speed of a model by 38%. This represents a fundamental shift. While CUDA's strength lies in its vast, human-expert-driven ecosystem of tools and tacit knowledge, AMD's strategy is to make its hardware and software stack directly accessible and optimizable by AI agents. An agent can parallelize tasks—debugging, profiling, coding—that would take human engineers years to master, compressing the traditional software adaptation timeline from years to tasks. The competition is no longer just about peak hardware specs but also about how well AI can read, utilize, and tune a platform.

marsbit07/28 00:09

NVIDIA's 20-Year CUDA Moat Collapsed Over a Weekend, Claude Single-Handedly Got AMD's New GPU Running

marsbit07/28 00:09

CPU Makes a Comeback to the Table, A $170 Billion "Power Seizure" Drama Begins

A new era is dawning for the server CPU (Central Processing Unit), driven by the shift from AI model training to large-scale reasoning and the rise of Agentic AI. This article explores how the CPU is reclaiming a central role in the AI data center. For years, the focus has been on the GPU (Graphics Processing Unit) for AI training. However, as AI moves to the inference and Agent phase—where tasks involve complex, multi-step reasoning, tool calls, and data management—the workload balance is flipping. Studies show CPUs now handle over 70% of the workload in Agentic AI, up from 10-30% in training. This is because Agent tasks generate massive intermediate data (KV Cache) that exceeds GPU memory, forcing it to be offloaded to the CPU's larger, more scalable memory pools. This increased importance is translating into market changes. Major players are taking note: NVIDIA launched its first standalone CPU line, Vera, based on ARM architecture and optimized for Agent performance. AMD doubled its server CPU market forecast to over $1200 billion by 2030. Analyst reports project the total server CPU market could reach $1700 billion by 2030, with AI-driven demand being a primary driver. Furthermore, the classic ratio of CPUs to GPUs in AI servers is rapidly changing, converging from 1:8 toward 1:1 for Agent deployments. This surge in demand has led to a rare industry-wide price increase of 10-15% for server CPUs from Intel and AMD, breaking a decade-long trend of "more performance for the same price." Demand is bifurcating into high-core-count CPUs for in-rack GPU support and moderate-core CPUs for standalone Agent task orchestration. In China, this global trend presents an opportunity for domestic CPU manufacturers like Hygon (海光信息) and Huawei Kunpeng, who are bolstered by both growing AI infrastructure needs and national policies promoting technological self-reliance ("xin chuang"). The maturity of their software ecosystems is also accelerating, evidenced by faster adaptation to new AI models. In conclusion, the narrative is shifting from a GPU-centric view to one where CPU-GPU synergy is critical. The CPU is no longer a peripheral component but a performance-defining bottleneck and a key growth driver in the AI hardware stack, opening a massive new market estimated in the hundreds of billions of dollars.

marsbit06/19 13:41

CPU Makes a Comeback to the Table, A $170 Billion "Power Seizure" Drama Begins

marsbit06/19 13:41

活动图片