Lecture

AI Hardwares

01 Sep 2025 • Richard Kuo

AI chips, Edge-AI, Benchrmarks, Frameworks

AI chips

Tesla

AI5

特斯拉 AI5 / HW5 下代 FSD 晶片號稱運算性能達 2,000~2,500TOPS（每秒兆次操作），是現款 HW4（約 500TOPS）晶片五倍，可支援更複雜的無監督 FSD 演算法。輝達 RTX5080 和 RTX5090（分別約為 1,500 美元和 3,000 美元）運算性能為 1,800TOPS 和 3,400TOPS。

Tesla Hardware 4 (AI4) – Full Details and Latest News

Hardware 3 (AI3) – FSD Computer

Tesla’s Hardware & E/E Architecture

Nvidia

Nvidia Draws GPU System Roadmap Out To 2028

Analysis of NVIDIA’s Latest Hardware: B100/B200/GH200/NVL72/SuperPod

DGX-Spark

Nvidia GB10 Grace Blackwell Superchip單系統晶片。GB10包含具第5代Tensor Cores及FP4精度支援的Nvidia Blackwell GPU，使DGX具備128GB記憶體，並支援最高1000 TOPS

ASUS Ascent GX10 搭載 NVIDIA® GB10 Grace Blackwell 超級晶片，提供高達 1 petaFLOP 的 AI 運算效能，並配備 128GB 記憶體，足以支援 2000 億（200B）參數模型的微調。

Geforce RTX 50 series

原價屋線上估價

Giga 5090OC 32GB
MSI 5080OC 16GB
Giga 5070Ti 16GB
Zotac 5060Ti 16GB

Built on the 4th Gen AMD CDNA™ architecture, AMD Instinct™ MI355X GPUs deliver leadership AI and HPC performance enabling high density infrastructures with 288GB HBM3E memory, 8TB/s bandwidth, and expanded FP6 and FP4 datatype support.

AMD 7 GHz Zen 6 Clock Speed Target “Confirmed” – Leaker Claims

Intel Gaudi 3

Intel® Gaudi® 3 AI Accelerators on IBM Cloud

Groq LPU

The Groq LPU™ Inference Engine

Paper: A Software-defined Tensor Streaming Multiprocessor for Large-scale Machine Learning

Google TPU v6e

Google Cloud推第6代TPU為基礎的新版Hypercomputer

Taiwan AI Cloud - 台智雲

Taiwan AI Cloud operates based on the strength of Taiwania 2, the supercomputer, which consists of 2,016 NVIDIA Tesla V100 32GB GPUs, delivering 9 PFLOPS of superior performance.

Edge AI

Nvidia Jetson

Kinara

Ara-2
the Kinara Ara-2 AI processor, the leader in Edge AI acceleration. This 40 TOPS powerhouse tackles the massive compute demands of Generative AI and transformer-based models with unmatched cost-effectiveness.

Kneron

KNEO300 EdgeGPT

KL730 AI Soc

Quad ARM® Cortex™ A55 CPU。
內建DSP，可以加速AI模型後處理，語音處理。
Linux和RTOS、TSMC 12 納米工藝。
高達4K@60fps解析度，與主流感測器的無縫 RGB Bayer 接口，多達4通道影像接口。
高達3.6eTOPS@int8 / 7.2eTops@int4。
支持Cafee、Tensorflow、Tensorflowlite、Pytorch、Keras、ONNX框架。
并兼容CNN、Transformer、RNN Hybrid等多種AI模型，有更高的處理效率和精度。

Realtek

AMB82-MINI

MCU
- Part Number: RTL8735B
- 32-bit Arm v8M, up to 500MHz
MEMORY
- 768KB ROM
- 512KB RAM
- 16MB Flash
- Supports MCM embedded DDR2/DDR3L memory up to 128MB
KEY FEATURES
- Integrated 802.11 a/b/g/n Wi-Fi, 2.4GHz/5GHz
- Bluetooth Low Energy (BLE) 5.1
- Integrated Intelligent Engine @ 0.4 TOPS

Nuvoton M55M1

RPi 5 + Hailo-8L

Hailo-8L = 13 TOPs
Broadcom BCM2712 2.4GHz quad-core 64-bit Arm Cortex-A76 CPU, with Cryptographic Extension, 512KB per-core L2 caches, and a 2MB shared L3 cache

Benchmarks

MLPerf

MLPerf Inference v5.0 measures inference performance on 11 different benchmarks, including several large language models (LLMs), text-to-image generative AI, recommendation, computer vision, biomedical image segmentation, and graph neural network (GNN).

MLPerf Training v5.0 measures the time to train on seven different benchmarks: LLM pretraining, LLM fine-tuning, text-to-image, GNN, object detection, recommendation, and natural language processing.