What is true on Npu?
Nothing on this track
Nothing on this track
| Published | Source | Entity | Captured intelligence | Topic | Actions |
|---|---|---|---|---|---|
| 10h ago | @seeedstudio |  OpenAI | @seeedstudio10h agoOpenAI @seeedstudio on X 🚀 #ProjectSpotlight Using #OpenAI #Codex + #reComputer #RK3576 to build and deploy a #YOLOv8 video detection WebUI, working directly with the board over SSH. What if you could go from an idea to a working Edge AI application with just a… | ||
| 4:42 PM ET | @a16z |  Thinking Machines | @a16z4:42 PM ETThinking Machines @a16z on X In the '90s, general-purpose chips wiped out the specialists. The original Thinking Machines, Cray Computer, and a wave of others shut down, and their best ideas got absorbed into other chips. Zain Asgar of @gimletlabs argues the… | Compute | |
| 1:53 PM ET | @SemiAnalysis_ |  DeepSeek | @SemiAnalysis_1:53 PM ETDeepSeek @SemiAnalysis_ on X Huawei Ascend 950PR Deep dive & µbenchmarks, DSv4.1 Flash Analysis, DeepSeek support for Ascend, AICPU, Tilelang Ascend https://t.co/3j3AvY17wC | Compute | |
| 8:53 PM ET | @kyleichan | DeepSeek | @kyleichan8:53 PM ETDeepSeek @kyleichan on X DeepSeek teaching other Chinese AI labs how to do AI inference ultra-efficiently at scale—and on Huawei 950 AI chips. https://t.co/TYk6J5Rc8r | Compute | |
| 9:51 AM ET | @UnmannedSystem |  Curtiss-Wright | @UnmannedSystem9:51 AM ETCurtiss-Wright @UnmannedSystem on X Curtiss-Wright integrates Link 22 and MAVLink protocols into its TCG HUNTR gateway to connect crewed and autonomous defense assets across land, sea, and air. Read more: https://t.co/pvQlpxhGyD @lantronix https://t.co/eUbsH1Slvd | Physical AI | |
| 6:29 AM ET | @tphuang |  Chery | @tphuang6:29 AM ETChery @tphuang on X Chery also unveiled 紫微芯片/Ziwei chip, its semi/chip division: Already 3 types of product: 750-1500V power module/including flagship 1500V+1450A SiC C100 CPU - 2 of which produces 1000KDMIPS 2 NPU produces 2400TOPS & 600 GB/s memory speed -… | Compute | |
| 6:06 AM ET | @seeedstudio |  Seeed Studio | @seeedstudio6:06 AM ETSeeed Studio @seeedstudio on X Dead Wi-Fi. Frozen kiosk. Angry guest. The fix isn't another chip. It's deciding which answers can't disappear—FAQs, directions, machine help—and running them locally. Define that and deployment gets predictable, support becomes billable:… | Compute | |
| 4:28 AM ET | @seeedstudio | Seeed Studio | @seeedstudio4:28 AM ETSeeed Studio @seeedstudio on X 🚀🔥reComputer #RK3588 Module is Coming Soon!!! ⚡ Octa-core CPU: 4× Cortex-A76 + 4× Cortex-A55 🧠 6 TOPS NPU for on-device AI acceleration 💾 Up to 32GB RAM & 128GB eMMC 🔌 Rich I/O with a 260-pin SO-DIMM interface 🖥️ 8K video decoding at… | Compute | |
| 10:00 AM ET | @seeedstudio |  Raspberry Pi | @seeedstudio10:00 AM ETRaspberry Pi @seeedstudio on X 🚀 #Rockchip RK3576 + #RaspberryPi PoE+ HAT: Power Over Ethernet Meets Edge AI! We adapted the #reComputer #RK3576 Module to the Raspberry Pi #PoE+ HAT, enabling PoE-powered operation through Ethernet while running #YOLO11 from Seeed… | Compute | |
| 2:10 PM ET | @tphuang |  Axera | @tphuang2:10 PM ETAxera @tphuang on X One thing I also enjoy following is edge AI chips, which are used for edge devices. Axera this week came out w/ AX8860 chip w/ 256 TOPS & 128GB LPDDR5X + 273 GB/s memory bandwidth. Uses 6nm process. This type of chip can run Qwen 27B model… | Compute | |
| 9:28 AM ET | @MilkRoadAI |  Apple | @MilkRoadAI9:28 AM ETApple @MilkRoadAI on X Apple's first foldable iPhone is officially here, and it is the MOST expensive iPhone ever made. Preorders for the iPhone Duo opened this morning at $1,999, with first deliveries and in store availability starting October 23. Within… | Compute | |
| 11:19 PM ET | @ricci_nov |  NVIDIA | @ricci_nov11:19 PM ETNVIDIA @ricci_nov on X Edge VLM placement is decided by energy per token. Running Qwen2-VL-2B, a Raspberry Pi 5 with a Hailo-10H NPU cuts energy per token about two-thirds versus CPU, and avoids thermal throttling. Jetson Orin Nano wins short replies; the curves… | Compute | |
| 11:01 AM ET | @SemiAnalysis_ |  Huawei | @SemiAnalysis_11:01 AM ETHuawei @SemiAnalysis_ on X China’s frontier models already run on China’s own chips. “Remember, the GLM-5 paper is almost a year old at this point. They say: ‘From the first day, GLM-5 is full-stack adapted to Chinese GPU ecosystems. We have successfully completed… | Compute | |
| 4:12 AM ET | @seeedstudio | Seeed Studio | @seeedstudio4:12 AM ETSeeed Studio @seeedstudio on X 🚀 Deploy YOLO11 on RK3576 Module + Raspberry Pi CM4 IO Board!! We tested the #reComputer #Rockchip #RK3576 Module on a #RaspberryPi #CM4 IO Board and successfully deployed #YOLO11 from #reComputerAILab for real-time object detection! ✅… | Compute | |
| 4:10 PM ET | @WevolverApp |  Mouser Electronics | @WevolverApp4:10 PM ETMouser Electronics @WevolverApp on X Embedded vision systems are handling more sensors, interfaces, and AI workloads while operating under tight power and latency constraints. Low-power FPGAs can act as aggregation and preprocessing hubs, bridging interfaces such as MIPI… | Sensing | |
| 12:06 PM ET | @WevolverApp |  MIPS | @WevolverApp12:06 PM ETMIPS @WevolverApp on X Specialized workloads are reshaping compute platform design, particularly for physical AI systems that process sensor data locally. These systems must meet strict requirements for latency, power, safety, security, and deterministic… | Physical AI | |
| 3:19 AM ET | @seeedstudio | Seeed Studio | @seeedstudio3:19 AM ETSeeed Studio @seeedstudio on X Users don't wait for the end of an answer. Build voice that handles it: local speech, interruptible by design, model placed where the job needs it. Build on a tested foundation → https://t.co/tHsiLSoyk9 #VoiceAI #EdgeAI | Sensing | |
| 9:35 PM ET | @seeedstudio | Seeed Studio | @seeedstudio9:35 PM ETSeeed Studio @seeedstudio on X From human demos to autonomous action 🤖 Using @huggingface @LeRobotHF , @knaveen trained an #ACT policy for Seeed’s #SOARM101, collecting 100K data steps and training for 11 hours. The #robot arm can now pick up a cube and place it into a… | Hands | |
| 9:09 PM ET | @pequityresearch | NVIDIA | @pequityresearch9:09 PM ETNVIDIA @pequityresearch on X Expert Takeaways: Unitree Humanoids > Robot chips split into low-compute parts (e.g., Rockchip RK3588, 6 TOPS) for low latency/fast communication and high-compute parts for high-bandwidth throughput. > Rockchip controls ~95% of the… | Humanoids | |
| 6:11 AM ET | @logi_today | NVIDIA | @logi_today6:11 AM ETNVIDIA @logi_today on X アスメック、搬送ロボット向け高性能AI端末 https://t.co/ikUNUph8W4 アスメック、ロボット向けエッジAIコンピューター「AIE-KT78/68」の取り扱いを開始。AMRやロボットアームに搭載し、クラウドを介さず現場で認識・判断・制御を実行する。 | Physical AI | |
| 6:00 AM ET | @seeedstudio | Seeed Studio | @seeedstudio6:00 AM ETSeeed Studio @seeedstudio on X Voice AI demos are easy— until they meet lobby echo, unstable Wi-Fi, or a robot that can’t wait for the cloud. For this test , Same ears. Same mouth. Different brains. Same site, two architectures. Cloud: usage grows, bill grows. Local:… | Compute | |
| 5:50 AM ET | @seeedstudio | Seeed Studio | @seeedstudio5:50 AM ETSeeed Studio @seeedstudio on X Look at the camera. It wakes up. Speak. It understands. 👀🎙️ Meet Visual Wake + Offline Speech Recognition on reCamera Pro — powered by on-device AI, with no cloud needed for core inference. Demo & guide: https://t.co/6gdnnM7s5u Product:… | Sensing | |
| 5:06 AM ET | @ricci_nov |  SK hynix | @ricci_nov5:06 AM ETSK hynix @ricci_nov on X SK hynix put 100+ in-house H-TPU cores in an SSD controller. Each draws ~2.5% of the power and 4% of the area of a general-purpose CPU core. It feeds Storage Next, which addresses 512-byte units, an eighth of the usual 4KB. Storage is… | Compute | |
| 12:03 PM ET | @pstAsiatech | Huawei | @pstAsiatech12:03 PM ETHuawei @pstAsiatech on X Huawei: Everyone wants the speed to be consistent within the cabinet, between cabinets, between data centers, and even across regions, so that a huge physical space can become a single computer Overall, we are closing the gap between… | Compute | |
| 4:24 AM ET | @seeedstudio | Seeed Studio | @seeedstudio4:24 AM ETSeeed Studio @seeedstudio on X 🔥🔥Multimodal Retail AI Center, running locally on the reComputer RK3588. 👁️ Smart Shelf Vision Monitor product states and restocking events with RKNN object detection. 🎙️ Voice Inventory Assistant Ask questions like “How many bottles… | Compute | |
| 8:01 PM ET | @pequityresearch | NVIDIA | @pequityresearch8:01 PM ETNVIDIA @pequityresearch on X Expert Takeaways: Inference > GPU Rental Rates Up ~20%: Driven by HBM supply constraints. Neo-cloud providers are passing hardware cost increases straight to tenants via shorter 1–3 year contracts rather than taking margin cuts. > HBM4… | Compute | |
| 7:51 PM ET | @kyleichan | @kyleichan7:51 PM ET @kyleichan on X This Chinese company is making custom AI chips embedded in humanoid robot fingertips to process up to 100,000 tactile sensing points. https://t.co/XesvMV6qqY | Humanoids | ||
| 9:36 AM ET | @tslaming |  Neuralink | @tslaming9:36 AM ETNeuralink @tslaming on X ENGINEERING NEWS 🧠 @neuralink has taught its brain chip a clever trick to tell a firing neuron from a chewing jaw 🦷 Chew a bite of your sandwich, and a brain implant might think a neuron just fired. Neuralink's patent application teaches… | Sensing | |
| 3:40 AM ET | @seeedstudio | Seeed Studio | @seeedstudio3:40 AM ETSeeed Studio @seeedstudio on X 🚀 Do you know reComputer RK3576 Module can work with Raspberry Pi CM4 IO Board We're testing compatibility between the #Rockchip #RK3576 Module and the #RaspberryPi CM4 IO Board, with several interfaces already up and running: ✅ HDMI… | Compute | |
| 11:41 PM ET | @seeedstudio | Seeed Studio | @seeedstudio11:41 PM ETSeeed Studio @seeedstudio on X Come and start learning with us! @NVIDIARobotics Full video here: https://t.co/O1p2UrncuT Get your own Jetson device: https://t.co/sZTpltxktl https://t.co/o35M7snt8Q | Sensing | |
| 8:55 PM ET | @stretchcloud |  Microsoft | @stretchcloud8:55 PM ETMicrosoft @stretchcloud on X $2,599 buys a Windows laptop that Microsoft says renders AI images 4.3 times faster than a 16 inch MacBook Pro on M5 Pro. The marketing is aimed squarely at Apple. The architecture underneath is aimed at something bigger: deciding, task by… | Compute | |
| 9:51 AM ET | @WevolverApp | Mouser Electronics | @WevolverApp9:51 AM ETMouser Electronics @WevolverApp on X Edge AI is moving beyond background analysis into machines that must sense, decide, and act in real time. Robotics, vision systems, and automation increasingly process data locally, reducing latency and improving responsiveness. This… | Physical AI | |
| 7:04 PM ET | @WevolverApp |  BrainChip | @WevolverApp7:04 PM ETBrainChip @WevolverApp on X Wearable devices must deliver more intelligence in smaller, lower-power designs. Continuous sensing and longer battery life limit available compute and energy, while AI workloads demand significant processing power. This creates a key… | Compute | |
| 9:19 AM ET | @techniahqrobot |  Kawasaki | @techniahqrobot9:19 AM ETKawasaki @techniahqrobot on X Kawasaki is bringing Physical AI into elderly care with Home LEO, a home robot planned for FY2028. The robot is designed to move around a house, recognize people and its surroundings, and search for objects. Kawasaki also plans functions… | Humanoids | |
| 6:16 AM ET | @techniahqrobot |  Doosan | @techniahqrobot6:16 AM ETDoosan @techniahqrobot on X Doosan Robotics just got picked for $69M worth of Korean Physical AI projects. One target a dual-arm industrial humanoid running AI directly onboard. The plan also includes 5 kg, 10 kg and 20 kg cobots. A Korean AI SoC with an NPU will… | Humanoids | |
| 8:54 PM ET | @CyberRobooo |  Tesla | @CyberRobooo8:54 PM ETTesla @CyberRobooo on X Tesla is redesigning the memory configuration of its AI chips for Optimus production. Elon said AI5 was initially planned with 144GB of RAM, then cut to 72GB LPDDR5( later nudged up to 96GB). AI6 is set at 144GB LPDDR6. Elon also says… | Humanoids | |
| 4:24 PM ET | @oprydai | @oprydai4:24 PM ET @oprydai on X it’s honestly insane how many pieces of the robotics stack are finally becoming good enough at the same time we now have - cheap high torque actuators - compact motors and gearboxes - better batteries - force and tactile sensing - depth… | Humanoids | ||
| 3:48 PM ET | @ricci_nov |  Mobilint | @ricci_nov3:48 PM ETMobilint @ricci_nov on X Mobilint's second-gen NPU Regulus slips six months to June 2027 after TSMC 12nm optimization with Asicland missed performance targets. The revenue target is halved to 15bn won from 30bn, while the Asicland contract was extended a year and… | Compute | |
| 9:00 AM ET | @seeedstudio |  Rockchip | @seeedstudio9:00 AM ETRockchip @seeedstudio on X 🔥🔥NEW Launch: reComputer Industrial RK3576 is Officially HERE!!!! Powered by Rockchip #RK3576 octa‑core CPU + 6 TOPS NPU, this compact #fanless industrial edge AI box delivers reliable, low‑latency on‑device AI inference without cloud… | Compute | |
| 2:28 AM ET | @seeedstudio |  Alibaba | @seeedstudio2:28 AM ETAlibaba @seeedstudio on X Run #LLMs on @NVIDIARobotics #Jetson without x86 ONNX export. 🚀 #TensorRT-Model-Connect on reComputer Classic J5012 (#AGXOrin 64GB, #JP7.2). @huggingface checkpoint → trtmc build → .bundle on device trtmc run for text gen Tested:… | Compute | |
| 11:51 AM ET | @NobleMachines |  Noble Machines | @NobleMachines11:51 AM ETNoble Machines @NobleMachines on X Isaac ROS 5.0 launched today at #ROSCon in Toronto. 🚀 Noble Machines is using #NVIDIA Isaac ROS on #Jetson to accelerate development of our general-purpose robot, Moby — adopting ready-to-use AI and perception rather than building it from… | Physical AI | |
| 9:48 AM ET | @NVIDIARobotics | NVIDIA | @NVIDIARobotics9:48 AM ETNVIDIA @NVIDIARobotics on X NVIDIA Isaac ROS 5.0 brings AI agents to robotics development. 🤖 Announced at #ROSCon, the release adds agentic workflows, Isaac Skills, ROS 2 Lyrical support, GPU-accelerated libraries and NVIDIA Jetson support from Orin Nano to Thor.… | Physical AI | |
| 8:58 PM ET | @stretchcloud | Alibaba | @stretchcloud8:58 PM ETAlibaba @stretchcloud on X The pattern I keep seeing this month: once a model layer commoditizes, the fight moves one layer down to whoever serves it fastest. RunAnywhere, a YC W26 company, launched Wally this week, an inference stack for open frontier models that… | Compute | |
| 3:22 PM ET | @ricci_nov |  Google | @ricci_nov3:22 PM ETGoogle @ricci_nov on X Google Cloud's Thomas Kurian puts AI server payback under two years overall, and under one year for servers using in-house silicon such as TPUs, per DigiTimes. That gap, not benchmark wins, is what funds the next custom chip. Order… | Compute | |
| 12:20 AM ET | @pstAsiatech | Huawei | @pstAsiatech12:20 AM ETHuawei @pstAsiatech on X The Ascend 960DT is Huawei’s first publicly known accelerator with High Bandwidth Memory (HBM). 288 GB are distributed across eight memory stacks, providing a total transfer rate of 9.6 TB/s. It remains unclear where the memory originates… | Compute | |
| 1:01 PM ET | @SemiAnalysis_ | Google | @SemiAnalysis_1:01 PM ETGoogle @SemiAnalysis_ on X Many Google TPU customers are already asking Google for SemiAnalysis AgentX performance results, as customers are saying that it is a representative workload of their agentic inference traffic. We are excited to announce that we are… | Compute | |
| 8:16 PM ET | @tphuang |  CXMT | @tphuang8:16 PM ETCXMT @tphuang on X Or that 960 is a lot better. 512k NPU would be 128 Atlas-960 SuperNodes w/ 128 PB of HBM & maybe another another 200 PB+ of other DDR5. They just keep talking about this 10T param model training & inference. Makes me wonder how much of… | Compute | |
| 4:13 PM ET | @pequityresearch |  AMD | @pequityresearch4:13 PM ETAMD @pequityresearch on X J.P. Morgan: ASIC/XPU units expected to reach 55% of total units in 2028E, up from 32% in 2025. $AMD $NVDA $GOOGL $META $MSFT $AMZN $AVGO | Compute | |
| 12:17 PM ET | @pequityresearch | Google | @pequityresearch12:17 PM ETGoogle @pequityresearch on X HSBC: Outside of Google, the % of ASIC compute among major CSPs remains steady from 2025 and into Q2 2026. $GOOGL $AMZN $META $MSFT $ORCL | Compute | |
| 6:15 AM ET | @tphuang | @tphuang6:15 AM ET @tphuang on X Comparing Atlas 950 to 960 SuperNode. Training MFU is up 10%. Training tps is 2.3x. Inference tps s 2.5x & delay is roughly just 70% 1024 to 4096 card & Hi-ONE NPO means 48k to 0 pluggable transceiver. UnifiedBus means just 2us RTT. 1PB is… | Compute | ||
| 5:47 AM ET | @seeedstudio | Alibaba | @seeedstudio5:47 AM ETAlibaba @seeedstudio on X 🔥🔥#RK1828 vs. #RK1820: 32K context. Fully local. We pushed long-context AI inference further on the #reComputer RK3576 Dev Kit with #Qwen3 running locally: ⚡ RK1828 + Qwen3-4B → 32K NIAH: 12/12 PASS → 74.75 tok/s generation → ~300 ms… | Compute | |
| 7:40 PM ET | @NVIDIARobotics | Alibaba | @NVIDIARobotics7:40 PM ETAlibaba @NVIDIARobotics on X Agentic AI is moving to the edge. On NVIDIA Jetson AGX Thor, TensorRT Edge-LLM completed MLPerf Edge Agentic 6.4x faster than the llama.cpp reference, reaching 52.33 token/sec for Qwen 3.6 27B model. NVFP4, ~96% of prompt tokens served… | Compute | |
| 1:29 PM ET | @tphuang | @tphuang1:29 PM ET @tphuang on X Pair this up w/ Atlas-960E liquid-cooled SuperPoD for 2027Q3 launch, industry's 1st NPO SuperPoD using 7.2T NPO transceiver. SuperPoD size reduced to 4096 (vs 16k planned in 2025/9). Expecting huge gain in MFU, trainin, decoding &… | Compute | ||
| 1:05 PM ET | @tphuang | @tphuang1:05 PM ET @tphuang on X Ascend Timeline - 2025/9 vs 2026/9 1Q+ jump in 960 availability Huge change in focus from FP8 -> FP4 FP4 used to be 2x FP8, now its 3-4x FP8 Big jump from 970 to 980 in memory speed (14.4 -> 38.4 TB/s) & interconnect speed (4.4 -> 8 TB/s)… | Compute | ||
| 12:50 PM ET | @pstAsiatech | Huawei | @pstAsiatech12:50 PM ETHuawei @pstAsiatech on X Big... At HUAWEI CONNECT 2026, Huawei Rotating Chairman David Wang highlighted the Atlas 960E SuperPoD, the industry's first NPO-based SuperPoD, an agentic SuperCluster that can support up to one million NPUs, and progress on NPO… | Compute | |
| 11:45 AM ET | @thenewsink | Huawei | @thenewsink11:45 AM ETHuawei @thenewsink on X Huawei vs Nvidia https://t.co/LTWp72U4Ad | Compute | |
| 4:34 AM ET | @seeedstudio | Alibaba | @seeedstudio4:34 AM ETAlibaba @seeedstudio on X 🔥🔥RUN #7B #VLM on reComputer #RK3576 Dev Kit + #RK1828 We tested #Qwen 2.5-VL-7B_prune on the reComputer RK3576 Dev Kit + RK1828 AI Accelerator: ⚡ TTFT: 418.93 ms 🚀 Streaming Output: 330.68 chars/s ⏱️ End-to-End: 2.84 s From image… | Compute | |
| 4:35 PM ET | @NVIDIARobotics | Alibaba | @NVIDIARobotics4:35 PM ETAlibaba @NVIDIARobotics on X Double the throughput on the same edge platform. ⚡ See Qwen3.5 9B NVFP4 run with DFlash speculative decoding turned on and off, reaching 68.1 tokens/s with DFlash enabled on NVIDIA Jetson. https://t.co/VeDRiQDmz8 | Compute | |
| 10:16 AM ET | @CKCapitalxx |  Broadcom | @CKCapitalxx10:16 AM ETBroadcom @CKCapitalxx on X $AVGO is getting too cheap to ignore. Forward PE is 20x. A year ago it was 32x. Forward EV/EBITDA is 16x, down from 26x. The multiple has been cut by more than a third while the business did this: Q3 revenue $29.6 billion, up 86%. AI… | Compute | |
| 8:51 PM ET | @pstAsiatech |  MediaTek | @pstAsiatech8:51 PM ETMediaTek @pstAsiatech on X MediaTek launches new mobile chip using TSMC's most advanced technology MediaTek said the Dimensity 9600 Pro is its first mobile processor built using Taiwan Semiconductor Manufacturing Co's cutting-edge 2-nanometre process. It also… | Compute | |
| 4:12 PM ET | @techniahqrobot |  AnalogAI | @techniahqrobot4:12 PM ETAnalogAI @techniahqrobot on X AnalogAI is putting on-device learning into its first Physical AI processors. The Korean company selected Microchip SST’s memBrain SAGE analog compute-in-memory IP, targeting operation at 1 watt or below. The processor is designed to train… | Physical AI | |
| 9:00 AM ET | @seeedstudio | Alibaba | @seeedstudio9:00 AM ETAlibaba @seeedstudio on X 🔥🔥Testing 7B #LLM on #RK3576+#RK1828 We put Qwen2.5-7B (W4A16) to the test on the reComputer RK3576 Dev Kit with RK1828 AI Accelerator: ⚡ Prefill: 121.12 tok/s 🚀 Generation: 60.42 tok/s ⏱️ TTFT: 309.64 ms But this isn’t just another… | Compute | |
| 6:18 PM ET | @SemiAnalysis_ | @SemiAnalysis_6:18 PM ET @SemiAnalysis_ on X Vera Rubin NVL72 Agentic Inference: 67x better Performance per Dollar Jensen Sandbagging Performance Again, 2x more Annual Profit Per GigaWatt, The More you Buy, The More you Earn, AgentX, InferenceX, Extreme Co-Design… | Compute | ||
| 6:13 PM ET | @SemiAnalysis_ | @SemiAnalysis_6:13 PM ET @SemiAnalysis_ on X Vera Rubin NVL72 Agentic Inference: 67x better Performance per Dollar Jensen Sandbagging Performance Again, 2x more Annual Profit Per GigaWatt, The More you Buy, The More you Ear, AgentX, InferenceX, Extreme Co-Design… | Compute | ||
| 5:14 AM ET | @smsehy | @smsehy5:14 AM ET @smsehy on X Analog front-ends and power management are merging with ultra-low-power neural processing units. In physical AI, sensor fusion must execute within milliwatt power envelopes at the edge rather than shipping raw telemetry to central compute.… | Physical AI | ||
| 3:30 AM ET | @seeedstudio | NVIDIA | @seeedstudio3:30 AM ETNVIDIA @seeedstudio on X 🚀🚀 reComputer AI Lab Just Got a Major Upgrade! With 90+ AI models, 100+ tutorials, and an ever-growing collection of projects, we’ve redesigned #reComputerAILab to make finding and deploying AI models easier than ever. 🔍 More ways to… | Compute | |
| 3:30 AM ET | @seeedstudio | Alibaba | @seeedstudio3:30 AM ETAlibaba @seeedstudio on X 👁️ Testing Qwen2.5-VL-3B on the reComputer RK3576 Dev Kit with RK1820 AI Accelerator ⚡ 361.98 ms Time to First Token 🚀 227.56 chars/s Streaming Output 🧠 Local Vision-Language Model 🔍 Image Understanding · Visual Q&A 🔒 Fully Local… | Compute | |
| 5:50 PM ET | @ricci_nov |  Positron AI | @ricci_nov5:50 PM ETPositron AI @ricci_nov on X Positron AI raised $875M at a $5B valuation, in two tranches led by NEA and Jim Clark. Its pitch is inference silicon built on commodity memory instead of scarce HBM. Capital is now flowing to designs whose main claim is avoiding the… | Compute | |
| 5:20 PM ET | @tphuang |  Anthropic | @tphuang5:20 PM ETAnthropic @tphuang on X 30k requests/day really isn't a lot. This to me seems like what you would do for benchmarking. The big issue here is the indiscriminate nature by Kimi which possibly compromised users who would clearly not want their data to flow to… | Compute | |
| 12:13 PM ET | @pequityresearch | Google | @pequityresearch12:13 PM ETGoogle @pequityresearch on X Evercore ISI: ASICs account for about 10-30% of internal workloads today and in 5 years will account for 30-50%, but enterprise driven workloads will remain >90% with Nvidia. $GOOGL $AMZN $META $MSFT $NVDA | Compute | |
| 10:50 AM ET | @MilkRoadAI | Broadcom | @MilkRoadAI10:50 AM ETBroadcom @MilkRoadAI on X AI inference is about to create a second trillion dollar supply chain because inference is expected to rise from 30% of demand in 2023 to 75% by 2030 (Save this). That means every chatbot, coding assistant, search engine and AI agent will… | Compute | |
| 4:34 AM ET | @seeedstudio | Alibaba | @seeedstudio4:34 AM ETAlibaba @seeedstudio on X 🔥🔥RK3576+RK1820: 178.86 tokens/s. Local LLM, Running Fast. We put the #reComputer #RK3576 Dev Kit with #RK1820 AI Accelerator to the test. Running #Qwen 2.5-0.5B, it achieved an impressive 178.86 tokens/s in our test. 🧠 Rockchip RK3576… | Compute | |
| 3:52 AM ET | @smsehy | @smsehy3:52 AM ET @smsehy on X Heterogeneous compute architectures are mandatory for physical systems. Balancing deterministic real-time RTOS microcontrollers with dense neural processing units is the baseline requirement for functional safety qualification.… | Physical AI | ||
| 11:33 PM ET | @pequityresearch | NVIDIA | @pequityresearch11:33 PM ETNVIDIA @pequityresearch on X AI inference expected to grow to 75% of data center TAM by 2030. GPU share of AI server shipment will decline from 69.7% in 2026 to 58.5% in 2030, but still controlling the lead. $NVDA $AMD $GOOGL $META $MSFT $AMZN https://t.co/vzckmAd4Wo… | Compute | |
| 10:34 PM ET | @SemiAnalysis_ |  vLLM | @SemiAnalysis_10:34 PM ETvLLM @SemiAnalysis_ on X ALERT🚨 GB300 NVL72 NVLINK BACKPLANE ENABLES 13x BETTER PERF PER DOLLAR THAN HOPPER ON AGENTIC INFERENCE Through extreme co-design, GB300 NVL72 uses an innovative copper backplane to expand the scale-up domain to 72 versus 8 on Hopper.… | Compute | |
| 9:54 PM ET | @pequityresearch | @pequityresearch9:54 PM ET @pequityresearch on X Single-lane 200G NPO solutions are expected to begin ramping in 2027, reaching roughly one million units, and scaling to roughly ten million units by 2028. Via @tuolaji2024 at CIOE 2026 https://t.co/lbFKqDibSD | Compute | ||
| 6:08 PM ET | @stretchcloud | NVIDIA | @stretchcloud6:08 PM ETNVIDIA @stretchcloud on X NVIDIA's competitive advantage has never really been about chips alone. It's been about the software moat - CUDA, cuDNN, TensorRT, the full stack that makes their hardware irreplaceable in production ML workloads. SoL-Pi is an interesting… | Compute | |
| 5:58 PM ET | @SemiAnalysis_ | @SemiAnalysis_5:58 PM ET @SemiAnalysis_ on X One of the most useful frameworks in semiconductor research is to think in derivatives. The first derivative is often recognized early. For example, the insatiable demand for more AI accelerators has been well established. And therefore,… | Compute | ||
| 5:11 PM ET | @MilkRoadAI | OpenAI | @MilkRoadAI5:11 PM ETOpenAI @MilkRoadAI on X If compute is the biggest moat in AI, shouldn't we be buying OpenAI and Anthropic? Dylan Patel puts OpenAI at roughly 2 gigawatts entering 2026 and Anthropic slightly below, with both above 5 by year end, a 3 to 4x jump each. Together… | Compute | |
| 1:33 PM ET | @ricci_nov | OpenAI | @ricci_nov1:33 PM ETOpenAI @ricci_nov on X OpenAI's first inference ASIC, Jalapeno, is made at TSMC with Samsung HBM4. OpenAI Korea now says next-generation chips could also run at Samsung foundry and packaging. Samsung and Broadcom already have a $200B+ partnership through 2030.… | Compute | |
| 4:36 AM ET | @seeedstudio | Seeed Studio | @seeedstudio4:36 AM ETSeeed Studio @seeedstudio on X 🔥🔥Real DATA: 30 FPS YOLO11n on the reComputer RK3576 Dev Kit Running #YOLO11n on the #reComputer #RK3576 Dev Kit, we achieved 30 FPS and even higher in our real-world testing. With further software optimization, there’s still room to… | Compute | |
| 9:05 PM ET | @SemiAnalysis_ | @SemiAnalysis_9:05 PM ET @SemiAnalysis_ on X Where Does a Robot Think On-Device vs Datacenter Inference The Embodiment Problem, Planning vs Action Layers, Glass-To-Glass Budgets, Wafers & DRAM Constraints, One B300 vs 56 Thors TCO, Factories To Caves https://t.co/ABD9HsM7B7 | Physical AI | ||
| 7:08 PM ET | @ricci_nov |  Arm | @ricci_nov7:08 PM ETArm @ricci_nov on X Arm's C2 cluster can carry two SME2 matrix units instead of one, at 5-7 TOPS, cutting agentic workload time up to 24%. Neoverse CSS N4 claims over 2x per-socket performance and 75% more memory bandwidth. Arm cites token cost as the reason… | Compute | |
| 10:37 AM ET | @CKCapitalxx |  Meta | @CKCapitalxx10:37 AM ETMeta @CKCapitalxx on X $META is my favorite mega cap play right now, and Muse is exactly why. Muse is a personal AI agent that actually takes actions, not just answers questions. Sends emails, books travel, negotiates bills, fills out forms, keeps working after… | Compute | |
| 8:23 AM ET | @ADI_News |  Analog Devices | @ADI_News8:23 AM ETAnalog Devices @ADI_News on X Analog Devices to Acquire Alif Semiconductor, Adding an AI-Native Processing Platform to Advance Physical Intelligence for the Next Generation of Real-World Systems. https://t.co/KyJo4B9EfK https://t.co/SM0Po9UgbO | Physical AI | |
| 9:03 PM ET | @SemiAnalysis_ | @SemiAnalysis_9:03 PM ET @SemiAnalysis_ on X TPUv7 has a specialized hardware-accelerated unit called the SparseCore. TPU's specialized SparseCore handles data movement, gathering each expert’s tokens into contiguous groups, while the TensorCore is left to run the expert matrix… | Compute | ||
| 3:23 PM ET | @pequityresearch | Google | @pequityresearch3:23 PM ETGoogle @pequityresearch on X Google says their payback period on AI servers in aggregate is less than 2 years, and on their own silicon is 1 year. $GOOGL | Compute | |
| 2:21 PM ET | @pequityresearch | Google | @pequityresearch2:21 PM ETGoogle @pequityresearch on X Goldman Sachs: ASICs to account for 50%/52%/55% of total AI chips in 2026/27/28E. $GOOGL $AMZN $MSFT $META | Compute | |
| 1:18 PM ET | @SemiAnalysis_ |  Cerebras | @SemiAnalysis_1:18 PM ETCerebras @SemiAnalysis_ on X Sat down with @milksandmatcha at @Cerebras Supernova to talk chips Everyone is making different inference bets Some want cheap tokens Some want speed NVIDIA is hard to beat on throughput, but there's huge room on speed Cerebras is a… | Compute | |
| 9:01 AM ET | @SemiAnalysis_ | Google | @SemiAnalysis_9:01 AM ETGoogle @SemiAnalysis_ on X ALERT🚨🚨: On apples-to-apples InferenceX Comparisons, TPUv7 Ironwood achieves 50% better perf per dollar than Blackwell Ultra on the new TorchTPU external inference stack! TorchTPU brings native PyTorch to TPUs, & along with Google… | Compute | |
| 12:35 AM ET | @pequityresearch |  Intel | @pequityresearch12:35 AM ETIntel @pequityresearch on X UBS: On Intel’s EMIB-T, industry feedback is positive regarding the yield improvement and overall execution, consistent with what we highlighted in July. We are constructive on EMIB-T’s supply supporting >2m substrate volume for MediaTek’s… | Compute | |
| 4:09 PM ET | @SemiAnalysis_ | @SemiAnalysis_4:09 PM ET @SemiAnalysis_ on X TPU Inference Externalization Full Steam Ahead - InferenceX, Up to 50% Better Performance per Dollar, Rapid Externalization of TPU stack, Growing Customer Base, Ironwood, TPUv8i, Reducing CUDA Moat https://t.co/9CFY8YDRsI | Compute | ||
| 4:04 PM ET | @SemiAnalysis_ | @SemiAnalysis_4:04 PM ET @SemiAnalysis_ on X TPU Inference Externalization Full Steam Ahead InferenceX, Up to 50% Better Performance per Dollar Rapid Externalization of TPU stack, Growing Customer Base Ironwood, TPUv8i, Reducing CUDA Moat https://t.co/9CFY8YDRsI | Compute | ||
| 5:01 AM ET | @NEURARobotics |  SECO | @NEURARobotics5:01 AM ETSECO @NEURARobotics on X A new step forward in scaling Physical AI from Europe. We are glad to welcome a new partner to our ecosystem. We're joining forces with SECO, an Italian company with deep expertise in embedded and edge AI computing, to help bring our… | Humanoids | |
| 7:41 PM ET | @smsehy | @smsehy7:41 PM ET @smsehy on X 15 gigawatt shortfalls in data center power by 2027 will push inference directly into embedded zonal controllers and edge silicon. When megawatt utility interconnects take four years to approve, running smaller quantized models on factory… | Compute | ||
| 5:22 AM ET | @tchsignal | Microsoft | @tchsignal5:22 AM ETMicrosoft @tchsignal on X Microsoft Project Zenith Sets a New Windows Baseline for Local AI Development Microsoft is building Project Zenith around a very different idea of what a developer PC should be. The ready-to-code Windows experience targets developer-class… | Compute | |
| 3:30 AM ET | @seeedstudio | Seeed Studio | @seeedstudio3:30 AM ETSeeed Studio @seeedstudio on X 🔥🔥Officially Launched: reComputer #RK3576 Dev Kit with #RK1820 AI Accelerator — 26 TOPS of Edge AI Power The reComputer RK3576 Dev Kit combines the #RK3576 Compute Module + IO Board + Optional #RK1820 AI Accelerator into a complete… | Compute | |
| 9:23 PM ET | @ricci_nov |  SambaNova | @ricci_nov9:23 PM ETSambaNova @ricci_nov on X SambaNova on Llama 3.1 70B: 16 SN40L chips, versus its own estimate of 336 Cerebras dies for 445 TPS and 576 Groq LPUs for 250 TPS. Each socket holds 520 MiB SRAM, 64 GiB HBM and 1.5 TiB DDR. Vendor comparison, not a benchmark. Inference… | Compute | |
| 3:30 AM ET | @seeedstudio | Seeed Studio | @seeedstudio3:30 AM ETSeeed Studio @seeedstudio on X 🔥🔥Now Available: reComputer RK3576 Module — Compact Industrial Edge AI Platform The module delivers the compute needed for demanding edge workloads, from real-time #CV to local #LLM, #VLM, and offline speech AI. ⚡ 8-Core #CPU: 4×… | Compute | |
| 8:22 PM ET | @tphuang | @tphuang8:22 PM ET @tphuang on X Comparing Kirin 2026 to 9030 Pro for NPU, GPU & CPU single Core - 9030 is Grey, 2026 matching 9030 perf is green & 2026 Full is Orange. Most notable is NPU going from 29 to 70 TOPS or consuming just 34% power @ 29 TOPS. GPU @ same Volt &… | Compute |