2026-06-07 Domingo

Centro de Notícias - Página 58

Obtém notícias cripto em tempo real e tendências de mercado com o Centro de Notícias da HTX.

New AMD Paper Overturns Conventional Wisdom: FP4 Training Instability's Cause Is Not Insufficient Randomness

AMD's new research challenges the conventional understanding of FP4 training instability. While reducing precision from FP8 to FP4 promises doubled computational throughput and is supported by new hardware like NVIDIA Blackwell and AMD MI350 series, training large language models natively with FP4 has been notoriously unstable, often attributed to insufficient stochasticity. The paper "Pretraining large language models with MXFP4 on Native FP4 Hardware" demonstrates successful end-to-end FP4 pre-training of Llama 3.1-8B on AMD MI355X GPUs using the MXFP4 format, achieving a 9-10% overall speedup over FP8. Crucially, it identifies the root cause of instability: not randomness, but the accumulation of *structural micro-scaling errors* along the sensitive weight gradient (Wgrad) path. Through controlled experiments, researchers found that quantizing the Wgrad operation to FP4 caused significant convergence degradation. Counterintuitively, common stochasticity-based mitigation techniques like stochastic rounding and randomized Hadamard transforms worsened performance. In contrast, applying a *deterministic* Hadamard transform successfully stabilized training by ensuring consistent error patterns, reducing the extra token cost from 26-27% to just 8-9%. This work has significant implications: 1) It provides a clear diagnostic for low-precision training instability, steering focus towards structural errors. 2) It pushes FP4 from a primarily inference-focused format into the realm of viable training. 3) It leverages the open OCP Microscaling (MX) standard, promoting cross-vendor compatibility. The research marks a critical step towards more economical large model training by further pushing the boundaries of low-precision computation.

marsbit05/27 06:19

New AMD Paper Overturns Conventional Wisdom: FP4 Training Instability's Cause Is Not Insufficient Randomness

marsbit05/27 06:19

Will ONDO's 'Tokenization Narrative' Change After Its CEO's Unexpected Passing?

Ondo Finance founder and CEO Nathan Allman has passed away unexpectedly. Allman, a Brown University graduate with a background in private credit and Goldman Sachs' digital asset team, was a key architect of Ondo's pivot from DeFi structured yield products to becoming a leading Real-World Asset (RWA) protocol. He drove the strategy to tokenize traditional financial assets like US Treasuries (OUSG), yield-generating dollar assets (USDY), and US stocks/ETFs (Ondo Global Markets) for on-chain accessibility. The company announced that President Ian De Bode, a former McKinsey partner with a strong institutional strategy and operations background, will succeed Allman as CEO. While Allman's sudden departure presents a near-term challenge, testing market confidence and Ondo's continuity, the project is seen as more than a founder-driven narrative. It has an established product suite and a management team with deep traditional finance experience. The long-term impact hinges on the new leadership's ability to execute. De Bode's expertise in compliance, distribution, and institutional partnerships aligns with RWA's next phase of scaling infrastructure. The core question is whether Ondo can maintain its product momentum and institutional relationships. Ondo's native ONDO token represents governance and RWA narrative value, not direct revenue from the underlying assets. Its future as a "top tokenization play" will depend on the team's continued delivery of product growth, asset scale, and real-world demand, moving beyond the initial emotional shock.

marsbit05/27 05:33

Will ONDO's 'Tokenization Narrative' Change After Its CEO's Unexpected Passing?

marsbit05/27 05:33

Bankless Co-founder's Confession on Selling Off ETH: Ethereum Did the Right Thing, but 'ETH as Money' Has No Future

Bankless co-founder David Hoffman recently sold his remaining ETH holdings, sparking debate within the Ethereum community. In a detailed explanation, Hoffman clarifies that his decision was not based on bearish sentiment towards Ethereum itself, which he remains highly optimistic about, but rather on the conclusion that the "ETH is Money" narrative has largely run its course. Hoffman argues that for ETH to achieve its envisioned status as global money, Ethereum needed to execute flawlessly across multiple layers—governance, technology, and market dominance—in a highly coordinated manner. He acknowledges Ethereum's significant successes and current justified valuation but suggests the window for a major revaluation based on this monetary narrative is closing. The post examines several challenges: the strong correlation between L1 chain activity/fees and native token value; the perceived failure of the "strong version" of crypto (user-owned, egalitarian systems) versus the rise of a "weak version" (efficient ledger technology for traditional finance); and the possibility that ETH's momentum as money was uniquely tied to the distorted conditions of the 2020-2021 period. Crucially, Hoffman highlights a structural tension: Ethereum is architected as a "giver, not a taker," providing critical infrastructure like secure block space and tokenization at cost. This ethos benefits the broader ecosystem (applications, L2s) but doesn't prioritize extracting maximum value for ETH itself. The "ETH is Money" thesis required Ethereum to win a war of overwhelming market dominance—a war its design philosophy refuses to explicitly fight. Therefore, while he sees continued immense success for the Ethereum network and its ecosystem (following a "fat application" theory where value accrues to apps and L2s), Hoffman finds it increasingly difficult to foresee a structural upward revaluation for the ETH asset based on the monetary narrative. His capital reallocation reflects a belief that this particular investment thesis has played out.

Odaily星球日报05/27 05:26

Bankless Co-founder's Confession on Selling Off ETH: Ethereum Did the Right Thing, but 'ETH as Money' Has No Future

Odaily星球日报05/27 05:26

Bankless Co-founder: Why I Sold All My ETH

Author David Hoffman, founder of Bankless, explains his decision to sell all his ETH, despite being a prominent figure in the Ethereum ecosystem. He clarifies that his move is not a bearish take on Ethereum itself, which he remains highly optimistic about as a network. His core argument is that the "ETH is money" thesis, which he helped popularize, has largely played out. Hoffman argues that ETH has achieved the market valuation it deserves based on Ethereum's current success and competitive position. He details several reasons for this view. First, the path for ETH to become global money required nearly flawless execution and sustained dominance across Ethereum's entire technical and social stack—a coordination challenge he now believes had a narrower window for success than anticipated. Second, market data shows a strong correlation between L1 chain activity/fees and the price of its native asset; Ethereum's fee dominance has been challenged by competitors like Solana. Third, the "strong version" of crypto (decentralized, native crypto economies) that ETH's monetary thesis relied upon has struggled to maintain a positive mainstream narrative and stable adoption beyond a brief period. Finally, Ethereum's architecture as a "giver"—providing secure block space and tokenization capabilities at cost to L2s and applications—means it doesn't capture premium value directly. Its rollup-centric roadmap further directs most profits to L2s and applications ("fat app theory"). In conclusion, Hoffman believes the opportunity for ETH to be revalued significantly upward as money has diminished. He sold not because ETH will fail, but because its monetary thesis has matured, and he seeks to allocate capital to other opportunities he finds more compelling.

链捕手05/27 02:11

Bankless Co-founder: Why I Sold All My ETH

链捕手05/27 02:11

From Issuer to Infrastructure Owner: Circle's Arc Strategy and the Fatal Gap in the GENIUS Act

Circle raised $222 million for its proprietary Layer-1 blockchain, Arc, positioning itself not just as a stablecoin issuer but as the owner of the settlement infrastructure USDC relies on. This move, backed by investors like BlackRock and Apollo, highlights a significant structural conflict unaddressed by the GENIUS Act of 2025. While the act focuses on stablecoin reserves and issuer oversight, it remains silent on the market structure implications of an issuer controlling the underlying network—a scenario akin to a currency issuer also owning the payment rails. Traditionally, financial regulations separate issuers from settlement infrastructure to ensure neutrality. With Arc, Circle gains control over transaction ordering, fees, and network rules, potentially favoring USDC over competitors. The article argues that this creates a permanent structural temptation, even if no abuse occurs. The solution lies in applying established market infrastructure principles: mandating neutral transaction ordering, transparent fee schedules, and governance separated from Circle’s commercial interests. The current pre-mainnet phase offers a critical window for regulators to establish these rules before Arc becomes entrenched. Once operational, enforcing changes would be costly and disruptive. The core question remains: should a regulated stablecoin issuer be allowed to own the settlement network its competitors must use? The GENIUS Act doesn’t answer this, but Circle’s Arc strategy makes it urgent.

marsbit05/27 02:05

From Issuer to Infrastructure Owner: Circle's Arc Strategy and the Fatal Gap in the GENIUS Act

marsbit05/27 02:05

What Are the Key Variables Determining the AI Bull Market?

Title: What Determines the AI Bull Market? Key Variables Revealed Despite rising oil prices above $100/barrel, persistent inflation, and fragile Fed rate cut expectations—a traditionally hostile environment for high-valuation tech stocks—the AI sector continues to drive the market to new highs. According to analysts, the current AI boom is in a phase of "rational fervor": while bubbles exist, they are not yet out of control. The crucial shift is the emergence of Agentic AI, which is evolving from an assisting tool (Copilot) to an autonomous execution tool (Autopilot), creating a clearer commercial path from investment to revenue. This shift accelerates Token consumption and inference computing demand while boosting revenue forecasts for leading firms. The market is now rewarding capital expenditure as it transforms from a burden into a competitive moat, supporting hardware chains like GPUs, optical modules, and storage. However, valuations have already priced in growth expectations for 2027-2028. The forward P/E ratio for the "Magnificent Seven" tech giants is about 35x, compared to 25x for the rest of the S&P 500. This premium implies AI adoption must occur 5 to 8 times faster than past technological revolutions—a scenario with little room for error. The sustainability of the AI bull market hinges on three key variables: 1. **Short-term liquidity shocks**: Risks include sustained high oil prices, resurgent inflation, rising interest rates, and potential unwinding of the yen carry trade. The critical question is whether the upward revision speed of Annual Recurring Revenue (ARR) can outpace the rise in interest rates. 2. **Mid-term industry realization**: Can the actual pace of AI adoption and commercialization match the current lofty valuations? Historically, general-purpose technology revolutions follow a non-linear path with periods of acceleration and deceleration. 3. **Long-term structural constraints**: These include energy and power grid limitations, employment displacement and consumer purchasing power, social acceptance and potential backlash, and potential hardware technology breakthroughs that could disrupt current supply chains. While the long-term prospects for AI remain optimistic with potential for significant productivity gains, the stock market's pricing depends not just on the vision but on the actual speed of realization amid these growing constraints. The direction is clear, but the pace of execution will determine whether the bubble remains controlled or spirals out of control.

marsbit05/27 02:05

What Are the Key Variables Determining the AI Bull Market?

marsbit05/27 02:05

活动图片