←
日常焦虑帝5,277 posts
日常焦虑帝 avatar

日常焦虑帝

@gpuhell

攻城狮/业余投机/右侧交易/游戏开发/Haskell/Rust/C++/Unity3D/C#/对java有偏见/RL/智商欠费/浅尝辄止故平庸/反乌托邦/竹林中/地狱变/手撕菠萝蜜/胸口碎榴莲/单机推特中/乐视一生黑(乐视已阵亡)/华为一生黑(迟早会阵亡 )/中华跪族/器材党/预防式B台支

◷ Joined June 2012▧ 513 Photos and videos
日常焦虑帝 avatar
日常焦虑帝@gpuhell·18 Oct 2017

我觉得国内应该全面禁止AR。对于一群做着中国梦的人来说,“增强现实”是致命的 (手动滑稽

日常焦虑帝 avatar
日常焦虑帝@gpuhell·3h

J.P. Morgan: Z.ai management expects to launch a new flagship model with approximately 3 trillion (3T) parameters in Q4 2026. Management also disclosed a new 700B-parameter model, primarily developed as a testbed for the larger training run, which reportedly already outperforms GLM-5.3. Its release remains unconfirmed, but if launched, it would precede the 3T flagship.

日常焦虑帝 avatar
日常焦虑帝@gpuhell·11h

According to multiple sources, Xiaomi’s large language model (LLM) unit, MiMo, is at a critical juncture over a potential spin-off. Several VC firms and CVC investors have initiated preliminary discussions, but Xiaomi has yet to formulate a concrete spin-off and fundraising plan. Sources indicate that Xiaomi founder Lei Jun and MiMo head Luo Fuli disagree over the proposed spin-off, with limited progress in discussions. Luo has reportedly approached a small number of investors and financial advisors (FAs) and may consider launching her own startup. mp.weixin.qq.com/s/7LwGupEtm…

日常焦虑帝 avatar
日常焦虑帝@gpuhell·Nov 17

Will M3 Pro save MiniMax’s stock price? If it can’t achieve SOTA, I’m afraid it will end up like BABA after Qwen4’s release.

Post attachment

CITI: MiniMax's management shared that the M3 Pro is soon approaching release at a ~3T parameter.

日常焦虑帝 avatar
日常焦虑帝@gpuhell·Nov 17

CITI: MiniMax's management shared that the M3 Pro is soon approaching release at a ~3T parameter.

日常焦虑帝 avatar
日常焦虑帝@gpuhell·Nov 16

According to Citi, the key component shortages that slowed NAURA's revenue growth in Q3 are expected to be largely resolved in Q4.

Post attachment
aog T avatar
⟳ 日常焦虑帝 retweeted
aog T@Taog_1575·Nov 14

Summary of the Kirin 9050 and SMIC Roadmap The table in Figure 1 summarizes SMIC’s node progress and products from 14 nm to the current N+3. In the metal layers, “unknown” indicates that the pitch is unknown but these metal layers are present. Key observations are listed from point 1 through point 8: 1) Liang Meng-song once commented that the area of logic chips shrank by 63% from 14nm to N+1. And this density change does indeed align with the G96H480 to G63H264. 2) N+2 is an optical shrink version of N+1, with a 4.76% reduction in both the X (CPP) and Y (cell height) directions. To maintain the same number of metal tracks and a 1.5 gear ratio (CPP/M1), M0–M2 were scaled down accordingly. 3) As mentioned above, N+2 is an optical shrink version of N+1. It can be seen that in subsequent N+2 products, the pitch of M3–M8 remains consistent, just as it was in N+1. The design rule unchanged in these layer. 4) The HD SRAM in N+2 and N+3 remains in the 1 bit/1 row format. Its cell height is consistent with that of the Logic HD cell, so adjacent peripherals using Logic HD cells can be tiled on grid. Figure 2 lists the SRAM HD/HP cell information for N+2 and N+3. 5) In the Kirin 9050, the RDL layer (which subsequently connects to bonding vias and then the following bonding pads) uses Cu RDL instead of Al pads to achieve a high-density bonding pitch. Figure 3 shows the appearance of an Al pad. Because the Al pad process creates concave where vias are present, this causes under-etching when the subsequent bonding via trenches are etched open to land on the RDL metal. Consequently, bonding vias must avoid areas on the RDL layer where vias are located, which imposes limitations on the bonding pitch scaling. In fact, Intel adopted Cu RDL as early as Intel 4, and TSMC followed it with its N2 process. 6) The Kirin 9050 adopt logicfolding and uses Cu-Cu hybrid bonds at the interface between the top and bottom dies, with a bond pitch of 1.5 µm. However, the bonding in the Kirin 9050 is of the “via-to-pad” type, as shown in Figure 4. Since the top die uses vias to bond to bottom die Cu Pad, it has one fewer metal layer than the bottom die (the bottom die has 13 layers—M0 through M12—plus 2 layers for the RDL and bond pads, totaling 15 layers; the top die lacks the bond pads, resulting in 14 layers). The “via-to-pad” bonding method offers better bonding fault tolerance (alignment, etc), but its electrical resistance is likely higher compared to “pad-to-pad” bonding due to the smaller bonding area. It should be noted that Huawei’s LogicFolding technology uses wafer-to-wafer bonding (Figure 5), and Huawei will continue to rely on wafer-to-wafer bonding as it moves toward smaller bond pitches in the future. As mentioned earlier, it is currently confirmed that Western manufacturers in the mobile will not adopt “Logic on Logic” designs (while “DRAM on Logic” may be possible, but I do not know), and fine pitch hybrid bond only feasible with wafer to wafer bonding now. AI chip/ servers currently have no sub-micron bonding pitch demand due to cost / yield concern, so they will continue to use die-to-wafer bonding with pitch to reach 4.5~3um by 2030. However, “DRAM on Logic” is poised for explosive growth and it will definitely utilize high-density hybrid bonding; the current trend is toward “sub-micron” wafer-to-wafer bonding pitch and I can tell you at least two foundry/company is progressing well on it with DRAM on Logic purpose XD (As mentioned earlier, the packaging manufacturer for the Xuaomi O100’s 1.4 µm hybrid bond pitch is not the same to Huawei’s logicfolding. 🤫) 7) Huawei’s density formula differs from the commonly used Angstronomic formula. While both densities are proportional to CPP × cell height, Huawei uses a constant of 2 (Figure 6), whereas Angstronomic uses a constant of approximately 1.47. Therefore, for the G63H252, Huawei lists the N+2 density as 126 mtr—which can be obtained by 2/(63 × 252)—while the N+3 density is 2/(57 × 227) = 155. However, according to the commonly used Angstronomic formula, the N+2 density is 92.8, and the N+3 density is 113.9. Due to logic folding, Huawei states that the Kirin 9050 has a density of 238; when converted back to the Angstronomic density, this equates to 175.3 mtr. Figure 7 shows the chart about Logic HD cell density (cell level), calculated using the Angstronomic formula. Huawei’s Kirin 9050 for 2026 has a density slightly lower than that of SF3/18A; by 2030, it will reach the N3E density level; and the sharp increase in 2031 is generally believed to result from Logicfolding being stacked to the third layer, with a density slightly lower than that of Intel’s 14A2 in 2030—which indeed corresponds to the 1.4 nm density level. 8) Although not labeled in Figure 1, the Kirin 9050’s N+2 top die makes extensive use of SRAM, and while the N+3 bottom die also includes SRAM, it uses HP SRAM. ————— Finally, as mentioned above, Logicfolding is expected to surge to the 1.4nm density level in 2031 (though performance, clock speed, will remain at the 2nm level). This is generally believed to result from stacking to a third layer, but process improvements possibility cannot be ruled out—especially since SMIC is set to launch N+5 in 2031, with a density slightly higher than that of N4. That said, I originally thought N+5 density would be achievable as early as 2027; now it seems the situation is far behind my original expectation. I also used to believe Huawei would never use “Logic on Logic” in smartphone chips (due to the heat dissipation trade-offs, not only heat density) unless they were desperate for process improvements. Now it seems SMIC is indeed moving too slowly, forcing Huawei to turn to Logicfolding, which has a very low yield. (You can see even Kirin9050 N+3 die is much smaller than Kirin 9030 (137mm2 vs 113.82), the supplying chip is very limited and much smaller than the later one.)

Post attachmentPost attachmentPost attachmentPost attachment
日常焦虑帝 avatar
日常焦虑帝@gpuhell·Nov 13

SMIC's Q3 capex fell, citing a shortage of semiconductor equipment supply as the main reason.

日常焦虑帝 avatar
日常焦虑帝@gpuhell·Nov 11

> The contract with SpaceX will not be extended beyond the middle of next year. emm 🤔

Jukan@jukan05Nov 11

Interesting remarks Anthropic made in a meeting with the sell side: - Anthropic has the stability and the better talent that open source lacks. - The AI market is winner take all. As long as we execute properly, we will take all of it. - The contract with SpaceX will not be extended beyond the middle of next year. - Developing Opus 5.5 cost more than $10 billion.

MiniMax (official) avatar
⟳ 日常焦虑帝 retweeted
MiniMax (official)@MiniMax_AI·Nov 10

4K image generation in 10 seconds! ⚡ Meet MiniMax H3-Image. 16 million pixel resolution. Stunning realism. Precise prompt following & editing. Up to 16 reference images. All in 10 seconds. And it's open-weight. 🚀 Available now via our API, MiniMax Code, and Hailuo AI — with 50% off 16MP+ generations during launch week! Can't wait to see what you create. 🔥 Hugging Face: huggingface.co/MiniMaxAI/Min… API: platform.minimax.io/docs/gui… MiniMax Code: code.minimax.io/about

0:50
日常焦虑帝 avatar
日常焦虑帝@gpuhell·Nov 1

According to CXMT’s earnings call, its 4F² technology has completed engineering sample validation, and the company is now advancing customer sampling and qualification. The first 4F²-based DDR5 RDIMM products are expected to be sampled by the end of 2026, initially targeting high-density, high-value applications such as servers. Volume production will depend on customer qualification progress and market demand.

日常焦虑帝 avatar
日常焦虑帝@gpuhell·Oct 29

Gemini's strategic pivot is paying off; going forward, it remains to be seen whether the market will buy in and what impact this will have on Anthropic.

Post attachment

Gemini 4 is already cooking—pretraining’s underway. The biggest twist this time? They’re really sweating the running costs now. Gemini used to play the bang-for-your-buck card, which kinda left it in the dust in the race for state-of-the-art.

Jasper Dekoninck avatar
⟳ 日常焦虑帝 retweeted
Jasper Dekoninck@j_dekoninck·Oct 26

Kimi K3.1 Pro enters MathArena as the best open model (once they release the weights)! Its estimated overall performance is now similar to GPT-5.5 (April 24), putting the open frontier around 6 months behind on MathArena

Post attachment
日常焦虑帝 avatar
日常焦虑帝@gpuhell·Oct 26

Make a wish: on the eve of Anthropic’s IPO, DeepSeek and GLM release new models capable of matching Opus 5.5 to take a shot at this rogue company.