https://www.youngju.dev/blog 천천히 올바르게. AI Researcher & DevOps Engineer Youngju's tech blog. GPU/CUDA, LLM, MLOps, Kubernetes AI workloads, distributed training, and data engineering. ko fjvbn2003@gmail.com (Youngju Kim) fjvbn2003@gmail.com (Youngju Kim) Sat, 16 May 2026 00:00:00 GMT https://www.youngju.dev/blog/culture/2026-05-16-ai-hardware-accelerators-2026-nvidia-blackwell-amd-instinct-mi-google-tpu-trillium-cerebras-wse-3-groq-lpu-sambanova-tenstorrent-etched-sohu-furiosa-rebellions-deep-dive.en https://www.youngju.dev/blog/culture/2026-05-16-ai-hardware-accelerators-2026-nvidia-blackwell-amd-instinct-mi-google-tpu-trillium-cerebras-wse-3-groq-lpu-sambanova-tenstorrent-etched-sohu-furiosa-rebellions-deep-dive.en In 2026 AI hardware is no longer an NVIDIA-only story. Blackwell (B100/B200/GB200 NVL72/B300) landed at GTC 2024, Rubin is queued for September 2026, AMD Instinct has marched from MI300X through MI355X to MI400 Helios, Intel Gaudi 3 ships as the last standalone line while Falcon Shores absorption looms, and Apple is set to unveil the AC1 server chip alongside M5 / M5 Pro in spring 2026. Google TPU v5p and Trillium (v5/v6) drive Gemini training, Cerebras WSE-3 packs 4 trillion transistors on a single wafer, Groq LPU rules sequential inference latency, SambaNova SN40L brings Reconfigurable Dataflow, Tenstorrent under Jim Keller bets on a RISC-V open architecture, and Etched Sohu (announced June 2024) is the first transformer-only ASIC. AWS Trainium 2 and Inferentia 3 anchor the in-house era, MatX and Tachyum Prodigy chase from behind, and phone NPUs from Apple A18 Pro, Snapdragon 8 Gen 4, MediaTek Dimensity 9400 and Google Tensor G5 push inference fully on-device. With NVLink 5/6, PCIe Gen 6/7, CXL, HBM3E/HBM4 and liquid cooling, this is the full 2026 landscape — seen through Korean (FuriosaAI, Rebellions) and Japanese (SoftBank Graphcore, Preferred Networks MN-3, Rapidus 2nm) lenses. Sat, 16 May 2026 00:00:00 GMT fjvbn2003@gmail.com (Youngju Kim) ai-hardwaregpuacceleratornvidia-blackwellb100b200gb200b300rubinamd-instinctmi300xmi325xmi355xmi400-heliosintel-gaudi-3falcon-shoresapple-m5apple-ac1google-tputrilliumcerebras-wse-3groq-lpusambanova-sn40ltenstorrentwormholeblackholeetched-sohutrainium-2inferentia-3matxtachyuma18-prosnapdragon-8-gen-4dimensity-9400tensor-g5hbm3ehbm4nvlinkcxlsamsung-hynix-hbmfuriosa-airebellions-sapeongraphcore-softbankpreferred-networks-mn-3rapidus2026deep-diveenglish https://www.youngju.dev/blog/culture/2026-05-16-ai-hardware-accelerators-2026-nvidia-blackwell-amd-instinct-mi-google-tpu-trillium-cerebras-wse-3-groq-lpu-sambanova-tenstorrent-etched-sohu-furiosa-rebellions-deep-dive.ja https://www.youngju.dev/blog/culture/2026-05-16-ai-hardware-accelerators-2026-nvidia-blackwell-amd-instinct-mi-google-tpu-trillium-cerebras-wse-3-groq-lpu-sambanova-tenstorrent-etched-sohu-furiosa-rebellions-deep-dive.ja 2026 年の AI ハードウェアはもはや NVIDIA 一社の物語ではない。Blackwell (B100/B200/GB200 NVL72/B300) が GTC 2024 で登場し、Rubin が 2026 年 9 月に控えるなか、AMD Instinct は MI300X から MI355X を経て MI400 Helios まで進撃した。Intel Gaudi 3 が最後の独立ラインとして出荷され、Falcon Shores の統合がささやかれ、Apple は M5 / M5 Pro とともに AC1 サーバチップを 2026 年春に投入する。Google TPU は v5p と Trillium (v5/v6) で Gemini の学習を支え、Cerebras WSE-3 は 4 兆トランジスタのウェハースケール、Groq LPU は逐次推論の速度、SambaNova SN40L は Reconfigurable Dataflow、Tenstorrent は Jim Keller 率いる RISC-V オープンアーキテクチャを掲げた。Etched Sohu は transformer 専用 ASIC を 2024 年 6 月に公開、AWS Trainium 2 と Inferentia 3 は自社チップ時代を開き、MatX と Tachyum Prodigy のような新興が続く。スマートフォン NPU 陣営の Apple A18 Pro、Snapdragon 8 Gen 4、MediaTek Dimensity 9400、Google Tensor G5 まで — NVLink 5/6、PCIe Gen 6/7、CXL、HBM3E/HBM4、そして液冷データセンタの風景を韓国 (FuriosaAI、Rebellions) と日本 (SoftBank Graphcore、Preferred Networks MN-3、Rapidus 2nm) の視点で総合的にまとめる。 Sat, 16 May 2026 00:00:00 GMT fjvbn2003@gmail.com (Youngju Kim) ai-hardwaregpuacceleratornvidia-blackwellb100b200gb200b300rubinamd-instinctmi300xmi325xmi355xmi400-heliosintel-gaudi-3falcon-shoresapple-m5apple-ac1google-tputrilliumcerebras-wse-3groq-lpusambanova-sn40ltenstorrentwormholeblackholeetched-sohutrainium-2inferentia-3matxtachyuma18-prosnapdragon-8-gen-4dimensity-9400tensor-g5hbm3ehbm4nvlinkcxlsamsung-hynix-hbmfuriosa-airebellions-sapeongraphcore-softbankpreferred-networks-mn-3rapidus2026deep-dive日本語 https://www.youngju.dev/blog/culture/2026-05-16-ai-hardware-accelerators-2026-nvidia-blackwell-amd-instinct-mi-google-tpu-trillium-cerebras-wse-3-groq-lpu-sambanova-tenstorrent-etched-sohu-furiosa-rebellions-deep-dive https://www.youngju.dev/blog/culture/2026-05-16-ai-hardware-accelerators-2026-nvidia-blackwell-amd-instinct-mi-google-tpu-trillium-cerebras-wse-3-groq-lpu-sambanova-tenstorrent-etched-sohu-furiosa-rebellions-deep-dive 2026년 AI 하드웨어는 더 이상 NVIDIA 한 곳의 이야기가 아니다. Blackwell(B100/B200/GB200 NVL72/B300)이 GTC 2024에서 등장하고, 2026년 9월 Rubin이 예고된 가운데 AMD Instinct는 MI300X에서 MI355X를 거쳐 MI400 Helios까지 진격했다. Intel Gaudi 3가 마지막 별도 라인으로 출하되고 Falcon Shores 합류가 루머링되며, Apple은 M5/M5 Pro와 함께 AC1 서버 칩을 2026년 봄에 띄운다. Google TPU는 v5p와 Trillium(v5/v6)으로 Gemini 학습을 지탱하고, Cerebras WSE-3는 4조 트랜지스터의 웨이퍼 스케일을, Groq LPU는 순차 추론 속도를, SambaNova SN40L은 Reconfigurable Dataflow를, Tenstorrent는 Jim Keller가 이끄는 RISC-V 오픈 아키텍처를 들고 나왔다. Etched Sohu는 transformer 전용 ASIC을 2024년 6월에 공개했고, AWS Trainium 2와 Inferentia 3는 자체 칩 시대를 열었으며, MatX와 Tachyum Prodigy 같은 신예가 뒤따른다. 폰 NPU 진영의 Apple A18 Pro, Snapdragon 8 Gen 4, MediaTek Dimensity 9400, Google Tensor G5까지 — NVLink 5/6, PCIe Gen 6/7, CXL, HBM3E/HBM4, 그리고 액체 냉각 데이터센터의 풍경을 한국(FuriosaAI, Rebellions)과 일본(SoftBank Graphcore, Preferred Networks MN-3, Rapidus 2nm)의 시각에서 종합 정리한다. Sat, 16 May 2026 00:00:00 GMT fjvbn2003@gmail.com (Youngju Kim) ai-hardwaregpuacceleratornvidia-blackwellb100b200gb200b300rubinamd-instinctmi300xmi325xmi355xmi400-heliosintel-gaudi-3falcon-shoresapple-m5apple-ac1google-tputrilliumcerebras-wse-3groq-lpusambanova-sn40ltenstorrentwormholeblackholeetched-sohutrainium-2inferentia-3matxtachyuma18-prosnapdragon-8-gen-4dimensity-9400tensor-g5hbm3ehbm4nvlinkcxlsamsung-hynix-hbmfuriosa-airebellions-sapeongraphcore-softbankpreferred-networks-mn-3rapidus2026deep-dive