Artículos Relacionados con Inference

El Centro de Noticias de HTX ofrece los artículos más recientes y un análisis profundo sobre "Inference", cubriendo tendencias del mercado, actualizaciones de proyectos, desarrollos tecnológicos y políticas regulatorias en la industria de cripto.

NVIDIA Stuns with First Vera Rubin Test, DeepSeek Throughput Skyrockets 30x

NVIDIA has unveiled the first on-silicon test results for its next-generation flagship cabinet, the Vera Rubin NVL72, using DeepSeek-V4-Pro on real "agent coding" workloads. The results are staggering: compared to the current leading GB300 NVL72, Vera Rubin delivers up to a 30x increase in throughput per megawatt and reduces token generation cost by up to 35x. The benchmark used was the AgentX test from SemiAnalysis, which captures full AI agent workflows with growing context, tool calls, and sub-agent generation, moving beyond traditional LLM benchmarks. This highlights a shift from the LLM era to the Agent era. Key innovations behind Vera Rubin's performance include extreme co-design, separation of services, distributed KV cache, KV-aware routing, MegaMoE architecture, 4-bit NVFP4 quantization, and 6th-gen NVLink for efficient MoE model execution. Simultaneously, NVIDIA announced the full-scale production of two new chips: 1. **Groq 3 LPX:** A low-latency inference accelerator designed for the Vera Rubin platform. When paired with Rubin GPUs for context processing, it achieves record-breaking output speeds—e.g., running Gemma 4 31B at 3,400 tokens/sec—drastically reducing multi-step agent task times. 2. **Vera CPU:** A processor specifically built for agentic AI, featuring 88 custom Olympus cores and 1.2TB/s memory bandwidth to handle the complex orchestration of agent tasks. SpaceXAI is already deploying it, with plans for space-based Vera Rubin systems by 2028. NVIDIA's strategy has evolved from selling GPUs to providing a complete, optimized "AI factory" stack—encompassing GPU, CPU, and specialized accelerators—to power the emerging Agent AI economy.

marsbitHace 9 hora(s)

NVIDIA Stuns with First Vera Rubin Test, DeepSeek Throughput Skyrockets 30x

marsbitHace 9 hora(s)

Arthur Hayes Unveils Flop Network and FLOP. Airdrop Scheduled for Q4

Investor Arthur Hayes has announced the Flop Network project in the agent economy, featuring the native FLOP token. Launch is expected in early 2027, with an airdrop planned for Q4 2026. Flop Network is described as a protocol based on Proof-of-Useful-Inference consensus. The FLOP token is central, serving as "fuel for AI agents" for performing inference and memory storage in a decentralized, censorship-resistant manner. Miners provide computing power and process blocks for rewards, while validators secure the network and store agent memory. Hayes will lead Flop Labs and emphasized there will be no FLOP presale or involvement from major venture capital. The announcement notes the project's architecture is still approximate and under development. The move aligns with growing industry focus on agent economies, with other projects like Coinbase's x402 protocol also emerging to handle transactions between AI agents, whose numbers are expected to rise significantly.

cryptonews.ru08/19 08:08

Arthur Hayes Unveils Flop Network and FLOP. Airdrop Scheduled for Q4

cryptonews.ru08/19 08:08

Arthur Hayes Makes a High-Profile Comeback, Flop Labs Aims to Become the "Fuel" for the Agent Economy

Arthur Hayes, co-founder of BitMEX, has announced his return as head of a new project called Flop Labs, declaring "FLOP is food for your AI agent." This marks his first high-profile return to a leadership role since stepping back from BitMEX. Flop Labs aims to be a native monetary network and verifiable computation settlement layer for the AI Agent economy, distinguishing itself from existing AI Agent projects. Its core concept is "Proof-of-Useful-Inference," which seeks to integrate useful AI inference tasks into the blockchain consensus mechanism, allowing FLOP tokens to serve as the native currency for Agents to purchase computing resources and store memory. The ecosystem involves four key roles: Miners (provide GPU compute), Validators (verify services), Agents (consume services), and KOLs/Partners (community growth). The project promises a 100% fair launch with no presale or VC involvement. A large airdrop is planned for Q4 2026, with the genesis block targeted for Q1 2027. While the announcement has generated community excitement, particularly around the airdrop, the project is in an early conceptual stage. It currently lacks a published whitepaper, detailed tokenomics, technical specifics on its consensus mechanism, or smart contract audits. Arthur Hayes has defended the venture, separating his belief in the potential of the Agentic Economy from concerns about an AI stock bubble. The success of Flop Labs will hinge on its ability to deliver substantive technical progress and ecosystem development in the coming months.

Odaily星球日报08/19 05:51

Groq Raises $350 Million at $3.5 Billion Valuation After Nvidia Deal

AI chip startup Groq has raised $350 million in a funding round that values the company at $3.5 billion, nearly half its peak valuation of $6.9 billion from September last year. The company was originally founded in 2016 to build hardware to compete with Nvidia but faced a significant setback after Nvidia licensed its technology and poached its founder, Jonathan Ross, along with other senior staff. This leadership loss forced Groq into a period of restructuring. The new funding round was led by Dallas-based investment firm Disruptive, with Nvidia also participating as an investor. Groq has pivoted its business model from selling chips to operating data centers, aiming to meet the demand for AI inference computing power. The company plans to use part of the new capital to expand its data center infrastructure, targeting over 200 megawatts of total capacity by next year.

cryptonews.ru08/17 18:30

Groq Raises $350 Million at $3.5 Billion Valuation After Nvidia Deal

cryptonews.ru08/17 18:30

Mining as Reasoning: How Does Nockchain Achieve Dual Rewards?

"Mining as Inference: How Does Nockchain's Dual Reward System Work? After looking at macroeconomic conditions, this piece delves into Nockchain's impending 'Logos' upgrade at block height 126,000. This Layer 1 blockchain, which launched fairly in May 2025, originally uses a Zero-Knowledge Proof-of-Work (ZKPoW) consensus where miners generate useful STARK proofs for the NockVM. The Logos upgrade introduces a second mining puzzle: integer matrix multiplication, the core computation for AI inference. The network will split block production, with 30% allocated to this AI puzzle and 70% to ZKPoW. This allows miners to potentially earn both block rewards and revenue from paid AI inference clients for the same computational work. The founders posit this creates a security budget that scales with AI demand. The upgrade, proposed and internally audited by National Compute (a company founded by Nockchain's creators to be its first major client), has been tested and delayed multiple times. While the technical implementation is a significant feat, the article notes the key challenge: National Compute must successfully onboard enough paying inference clients to generate real cash flow and validate the 'security + revenue' model. Currently, the $NOCK token trades around $0.009 with a ~$22 million market cap, awaiting the materialization of this promised market."

marsbit08/17 10:43

Mining as Reasoning: How Does Nockchain Achieve Dual Rewards?

marsbit08/17 10:43

Overnight, GPT-5.6 Sol Was Accelerated 14x by OpenAI

OpenAI, in collaboration with chipmaker Cerebras, has unveiled a limited preview of an "Ultrafast Mode" for its flagship GPT-5.6 Sol model. This new service tier reportedly achieves output speeds of up to 750 tokens per second—a 14x increase over the standard mode's baseline of ~53 tokens/s—without any loss in quality. Key to this acceleration is Cerebras's wafer-scale architecture (WSE-3), which houses model parameters entirely in on-chip SRAM to eliminate the memory bandwidth bottlenecks typical of traditional GPU clusters. In benchmark testing on the challenging "Humanity's Last Exam" (HLE), GPT-5.6 Sol in Ultrafast Mode answered all 2500 questions in 11 hours and 11 minutes, compared to over 78 hours for a competitor model, while maintaining similar accuracy. The speed boost also translated to a 5.6x faster end-to-end performance on the GDP-Val benchmark for economically valuable knowledge work. OpenAI highlights several potential applications for such rapid inference, including real-time event response and reliability analysis, dynamic financial research and security, complex customer support, interactive shopping assistance, and accelerated research and experimentation workflows that enable multiple iterative cycles within a single workday. This advancement may allow users to deploy the highest-tier models for tasks previously requiring slower secondary models, significantly compressing multi-step agent workflows from hours to minutes.

marsbit08/14 00:02

Overnight, GPT-5.6 Sol Was Accelerated 14x by OpenAI

marsbit08/14 00:02

Cerebras Plummets! Hardware Revenue Unexpectedly Declines, Challenging Nvidia Hits Demand Fluctuations

Cerebras Systems reported mixed Q2 2026 results. While total GAAP revenue grew 74% YoY to $180.1M, its core hardware sales unexpectedly fell 23% to $54.1M. In contrast, cloud and services revenue surged 281% to $126M, becoming the largest revenue segment and highlighting a business model shift. Despite the hardware slowdown, the company raised its full-year core revenue guidance to $880-$890M and provided a Q3 outlook above market expectations. CEO Andrew Feldman noted hardware growth will be "lumpy." Investors reacted negatively, with shares dropping ~17% after-hours. Concerns centered on the hardware decline, as Cerebras positions its wafer-scale chips as a challenger to Nvidia. The market is reassessing growth sustainability despite strong cloud performance and a $25.4B backlog. Cerebras is betting on AI inference demand, leveraging its architecture which avoids HBM and advanced packaging for potential cost advantages. With substantial cash reserves, it plans to scale manufacturing capacity tenfold in 2026 to meet future delivery commitments.

marsbit08/13 14:21

Cerebras Plummets! Hardware Revenue Unexpectedly Declines, Challenging Nvidia Hits Demand Fluctuations

marsbit08/13 14:21

AMD Buys Taalas: Hardware AI Manages Without the Scarce HBM Memory

AMD has agreed to acquire Toronto-based startup Taalas, which addresses a major bottleneck in AI inference: the need to constantly transfer a neural network's model weights from memory to the processor for each token generated. Taalas's chips eliminate this operation by embedding the model weights directly into the transistors themselves. This data transfer is what currently limits inference speed and has made high-bandwidth memory (HBM) a critically scarce resource. Taalas, founded in 2023, has developed application-specific integrated circuits (ASICs). Its first test chip, fabricated on TSMC's 6nm process, reportedly ran Meta's Llama 3.1 8B model at speeds 48 times faster than Nvidia GPUs. The architecture features a mask ROM section for permanently stored weights and SRAM for adaptable components. AMD plans to integrate these chips into its Helios racks alongside its Instinct accelerators. However, this approach comes with a significant trade-off: each chip is permanently hardwired for a single model. Switching models requires a partial chip redesign, a process taking about two months even with Taalas's accelerated method. This limits its applicability to stable, widely-used models. The acquisition highlights a broader challenge in the semiconductor industry: the current memory shortage. HBM is sold out through 2026, and DRAM prices have surged. Yet, Taalas's technology demonstrates that this memory bottleneck is an engineering challenge, not an absolute physical limit. The industry is actively working on solutions, from Nvidia's model compression to Samsung's zHBM and new high-speed flash memory standards, all aimed at reducing reliance on scarce HBM. From an investment perspective, the deal challenges the assumption that AI-driven memory demand will keep prices permanently high. It serves as a reminder that memory has historically been a cyclical business, and current high prices are funding the very innovations designed to reduce future demand.

cryptonews.ru08/09 20:02

AMD Buys Taalas: Hardware AI Manages Without the Scarce HBM Memory

cryptonews.ru08/09 20:02

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

The CUDA Moat That Huang Built Over 20 Years Was Chiseled Open by AI in Just 10 Hours

NVIDIA's CUDA, the software empire built over nearly 20 years to dominate AI computing, faces a new challenge. A startup called Infinity used its AI programming agent, "Ignition," to create CUDA-like software for AI chip company d-Matrix in just 10 hours. This highlights how AI agents can rapidly generate and optimize low-level GPU kernels, a core but labor-intensive part of chip software development. While impressive, experts caution this does not equal replicating CUDA's full ecosystem of libraries, debugging tools, and developer community. The real battlefield may be the inference market, where cost and efficiency matter more than the extreme performance needed for training. Here, companies like d-Matrix and Rebellions believe CUDA's lock-in is weaker, and specialized, cost-effective chips have an opportunity. However, NVIDIA's advantages remain formidable. Industry veterans point out that generating code is easier than rigorously validating it for production use—an area where CUDA's toolchain excels. Furthermore, NVIDIA itself is employing AI to accelerate its own development. The long-term competition may shift from who has the most code to who builds the superior ecosystem for verification and optimization. While CUDA's dominance in training is secure for now, the inference side could see the first cracks in its armor.

marsbit08/04 13:51