Register and share your invite link to earn from video plays and referrals.

Search results for 鑫隆
鑫隆 community
One keyword maps to one global community path.
Create community
People
Not Found
Tweets including 鑫隆
CXMT closing price after the first trading week China’s memory chip giant CXMT (长鑫科技) has officially made its Shanghai debut. Predict its closing price at the end of its first trading week on July 31, 2026. Now Live➡️
Show more
SITUATION ANALYSIS: CXMT lore explained We like to joke on MTS that our favorite company is China’s largest memory manufacturer, ChangXin Memory Technologies (长鑫存储). Today, CXMT surpassed Tencent to become the most valuable company in China, with a market cap of $555 billion. So how did we get here? CXMT was founded in 2016, less than a decade ago, in the Tier 3 city of Hefei, Anhui Province, by an engineer named Zhu Yiming. Zhu was born in Jiangsu Province in 1972. He studied physics at Tsinghua University, one of China’s top colleges, and became deeply interested in semiconductor design and manufacturing. At the time, the industry was deeply centered in the US, and so Zhu moved overseas in 1994 to pursue a physics PhD at Stony Brook University on Long Island. Not wanting to miss out on the 90s tech boom in Silicon Valley, he left Stony Brook early (after getting a master’s in electronics engineering) to work in the memory industry. In 2005, frustrated at his company undervaluing his ideas on chip design, he left to found GigaDevice, which built NOR flash memory and microcontrollers. Zhu saw the future of the memory industry moving to Asia, so he moved himself and his company to Beijing. GigaDevice became one of China’s leading semiconductor companies, and went public in 2016. Today, it’s worth over $40 billion. The same year GigaDevice went public, Zhu partnered with the municipal government of Hefei on a new project: building China’s first domestic manufacturer of DRAM, dynamic random-access memory. At the time, the global DRAM market was concentrated almost entirely into three companies: South Korea-based Samsung and SK Hynix, and US-based Micron. GigaDevice had no cell design or manufacturing process for DRAM, so they had to look elsewhere: to Qimonda, a defunct German DRAM manufacturer that had gone bankrupt in 2009. CXMT acquired many of Qimonda’s patents and recruited many former engineers. With that done, they shifted focus to mass production. Fab 1, backed by the Hefei government, was completed in early 2018 and began production in 2019. By 2020, it had done something no Chinese company had ever done before: sustained mass production of DDR4 and LPDDR4 DRAM chips, which were used mostly for budget Chinese smartphones and PCs. Over the following years, CXMT advanced its process technology further, producing DDR5 and LPDDR5, used for more high-end devices. In 2024, it began producing small amounts of older HBM2, the technology used in AI chips, and in 2025, it released DDR5 server memory. This would turn out to be incredibly useful — around that time, rising AI demand caused the memory market to skyrocket. In July 2026, CXMT went public, and its stock nearly quintupled on the first day of trading. Today, it produces nearly 10% of the world’s DRAM supply, and its memory ends up in mainstream Chinese smartphones, PCs, and data centers. It’s had an incredible arc. If anyone knows where to get CXMT merch, please let us know! Written by @theojaffee. Read more at our link in bio.
Show more
DeepSeek V4.1 Gives Prefill and Decode Different Compute Paths Prompt tokens mostly traverse 20 layers; generated tokens traverse all 40. This Causal Encoder-Decoder design nearly halves long-input prefill while preserving autoregressive generation. Zhihu contributor 潜龙勿用, Changxin Ke(柯昌鑫), a graduate researcher at ICT, CAS, explains how its architecture and post-training were designed together. 1️⃣ A causal encoder, not T5 The 40-layer backbone is split into a 20-layer causal encoder and a 20-layer decoder. Both remain causal. The encoder processes the prompt and supplies the decoder’s global KV. Most prompt tokens avoid the decoder stack, while generated tokens still run through all 40 layers. That means 8B active parameters per prefill token versus 16B per decode token. 2️⃣ Most layers share global memory CSA2 uses three modes: Full creates global KV and an index; Reindex shares the KV but selects new positions; Reuse shares both. Only four layers create independent global KV, four reindex it, and 30 reuse both KV and the latest index. Each layer retains its own query and local SWA state. The first Full layer builds up to 16,384 candidates. Later layers search this pool for their Top-512. From 4K to 1M context, decode FLOPs per token rise only about 25%. 3️⃣ Serving approximations enter training Exact reconstruction of decoder SWA states would replay 2,560 prompt tokens. V4.1 replays only the final 128 encoder outputs and trains the model to tolerate the approximation. With FP4 global KV, cache falls to 890 bytes per token, persistent cache to roughly one eighth of V4-Flash, and prefill compute close to half. Bounded replay and constrained retrieval are not last-minute serving tricks. The model experiences them during post-training. 4️⃣ Post-training is an evolving Agent system Each RL task combines a problem, environment, and verifier. New trajectories can reveal shortcuts, broken environments, or verifier errors and send the task back for repair. V4.1 trains across multiple harnesses, merges checkpoints between RL runs, and finishes with on-policy distillation from more than 40 teachers. Raising reasoning effort from 25 to 100 increases output length about 2.5×, while average Pass@1 across eight benchmarks rises from 67.1% to 76.3%. ✅ The real design choice V4.1 aligns model structure, cache policy, retrieval limits, training environments, and inference around long-running Agents. The model learns under the same constraints the deployed system will actually impose. 🔗 Full analysis: #DeepSeek# #DeepSeekV41# #LLMArchitecture# #AIAgents# #LongContext# #AIInfra#
Show more