SemiAnalysis

There's a lot of bullshit rumors about TSMC being spread regarding 2nm, Apple, Nvidia, and Qualcomm TSMC 2nm N2 has no delays. It was always H2 2025 for high volume manufacturing, as stated by TSMC publicly in 2023 TSMC switched to >2 year schedule between major nodes in 2021. N2 wafers take ~14 weeks to be fabricated across thousands of steps + couple months for packaging assembly. The absolute soonest a product can come out with N2 is ~Q2 2026. Apple has always known/planned chips codenamed Theras, Tilos, Hidra, Sotra, Baltra, Isonoe to be N3P Apple will release their first products with N2 in 2026, on schedule, as planned. They are increasing use of N-1 nodes with SoIC in their high end APs to keep costs down despite large wafer price increases. Apple's focused on wide IO copackaged memory to keep perf gains going. Nvidia and Qualcomm have no faith in Samsung Samsung is~3 years behind TSMC in PPA at this point, and even ~1 year behind Intel. Nvidia and Qualcomm always test new nodes at Samsung, and now also testing Intel. Neither plans to move their highest volume products away from TSMC.

Huahong, the 2nd largest Chinese foundry just made Dr. Peng Bai the President Dr. Bai is the real deal Intel for 31 years, including CVP in charge of Logic Technology Development for multiple nodes He managed Intel's D1C R&D Fab Was pushed out in end by being put on supply chain

Nvidia’s Christmas Present: GB300 & B300 – Reasoning Inference, Amazon, Memory, Supply Chain Blackwell Delays, Microsoft Orders, GB300 BOM, Nvidia Gross Margin, ConnectX-8, VRMs, Micron, Samsung, SK Hynix, Wistron, FII Foxconn, Aspeed, Axiado https://lnkd.in/dAQFWgnt

MI300X vs H100 vs H200 Benchmark Part 1: Training CUDA Moat Still Alive Our 5-month journey conducting independent analysis & benchmarking of AMD MI300X vs Nvidia H100 + H200 Detailed, open source low-level benchmarks performance vs TCO Comprehensive public recommendations It’s not just immature software, they need to change how they do development User Experience, Usability GEMM + attention InfiniBand vs Spectrum-X vs RoCEv2 Ethernet https://lnkd.in/du4Pqym3

SemiAnalysis at PTC'2025🌴 Jeremie Eliahou Ontiveros and Daniel Nishball, CFA are heading to Hawaii for PTC'25 next January! Want to meet? Contact us to schedule a one-to-one meeting or join our Networking Event co-organized with Hydra Host Register here for our networking event: https://hubs.la/Q0302LH20 #PTC2025 #Networking #DataCenter #SemiAnalysis

MI300X vs H100 vs H200 Benchmark Part 1: Training CUDA Moat Still Alive Our 5-month journey conducting independent analysis & benchmarking of AMD MI300X vs Nvidia H100 + H200 Detailed, open source low-level benchmarks performance vs TCO Comprehensive public recommendations It’s not just immature software, they need to change how they do development User Experience, Usability GEMM + attention InfiniBand vs Spectrum-X vs RoCEv2 Ethernet https://lnkd.in/du4Pqym3

Scaling Laws – O1 Pro Architecture, Reasoning Training Infrastructure, Orion and Claude 3.5 Opus “Failures” AI Lab Synthetic Data Infrastructure, Inference Tokenomics of Test Time Compute, The Data Wall, Evaluation's are Broken, RLAIF, Inference Time Search, Scale Needed More Than Ever https://lnkd.in/d9HfyvPi

Scaling Laws – O1 Pro Architecture, Reasoning Training Infrastructure, Orion and Claude 3.5 Opus “Failures” AI Lab Synthetic Data Infrastructure, Inference Tokenomics of Test Time Compute, The Data Wall, Evaluation's are Broken, RLAIF, Inference Time Search, Scale Needed More Than Ever https://lnkd.in/d9HfyvPi

Intel on the Brink of Death Board Short Circuits, Cultural Rot, Product Focus Is Flawed, x86 Has No Moat, Sell The PC Business, Foundry Must Survive Nvidia's PC CPU Push - 3 Chip Tegra Strategy Sierra Forest, Granite Rapid, Clearwater Forest, Diamond Rapids, Rogue River Forest https://lnkd.in/dCv5Ha4F

$ALAB COO misspoke at the UBS conference yesterday when talking about the company's Trainium2 content. As per our detailed architecture breakdown, there are 2 OSFP XD paddle cards per Trainium2 chip. This equates to 128 AEC paddle cards per Trn2-Ultra super rack (64 AEC cables). https://lnkd.in/eH9kewJw