CPU
General computing, data preparation and inference.
THE ENGINES BEHIND AI
AI processors.
Know the capacity.
Explore CPUs, GPUs, TPUs, NPUs and edge AI chips. See who makes them, what their specifications mean and where they fit in an AI workflow.
Explore 9 processors ↓GPU
Parallel computation for training and inference.
TPU
Google-designed tensor acceleration in the cloud.
NPU / SoC
Integrated, efficient on-device AI processing.
Edge AI
Compact devices for robots, cameras and local inference.
Compare like for like. TOPS means trillion operations per second; TFLOPS means trillion floating-point operations per second. Precision (INT8, FP4, BF16), sparsity, software and power limits change results. These examples are not a performance ranking or an exhaustive list of the newest chips. Memory needs also depend on model size, quantization, context and runtime overhead. Specifications checked 20 Oct 2026; follow each official source for updates.
9 processors
No processors match. Try another type, manufacturer or search.
Image source ↗
CPU · AMD
AMD Ryzen 9 9950X
Advanced Micro Devices, Inc.
Designs Ryzen CPUs, Ryzen AI processors and Instinct GPU accelerators. Manufacturer website ↗
- Architecture
- Zen 5; desktop AM5
- Memory capacity
- System DDR5 RAM; installed capacity depends on motherboard
- Compute / bandwidth
- 16 CPU cores / 32 threads; boost up to 5.7 GHz
- AI use
- Data preparation, development and CPU inference. Pair with a compatible GPU for accelerated training.
- Software
- x86-64 CPU frameworks; GPU software depends on the added GPU
CPU cores and clock speed are not GPU AI TOPS. No dedicated Ryzen AI NPU in this model.
Official specifications ↗▶ YouTube: AMD Ryzen 9 9950X — tutorials & explanations ↗Opens YouTube search results
Image source ↗
CPU · Intel
Intel Core Ultra 9 285K
Intel Corporation
Develops Core and Xeon CPUs and integrated AI acceleration technologies. Manufacturer website ↗
- Architecture
- Core Ultra 200S; desktop
- Memory capacity
- System DDR5 RAM; motherboard and configuration dependent
- Compute / bandwidth
- 24 CPU cores (8 performance + 16 efficient); NPU 13 peak INT8 TOPS
- AI use
- Desktop AI development, preprocessing and supported on-device AI tasks.
- Software
- Intel OpenVINO; x86-64 frameworks
The 13 TOPS figure belongs to the integrated NPU, not the CPU or a discrete GPU.
Official specifications ↗▶ YouTube: Intel Core Ultra 9 285K — tutorials & explanations ↗Opens YouTube search results
Image source ↗
GPU · NVIDIA
NVIDIA GeForce RTX 5090
NVIDIA Corporation
Designs GeForce GPUs, data-center accelerators and Jetson edge AI platforms. Manufacturer website ↗
- Architecture
- Blackwell; desktop graphics card
- Memory capacity
- 32 GB dedicated GDDR7 VRAM
- Compute / bandwidth
- 3,352 advertised AI TOPS (peak FP4 with sparsity)
- AI use
- Local model inference, image generation and compatible model fine-tuning.
- Software
- CUDA, TensorRT; compatible PyTorch builds
Desktop specification. Laptop RTX 5090 specifications differ. Peak sparse FP4 is not directly comparable with dense INT8 or BF16.
Official specifications ↗▶ YouTube: NVIDIA GeForce RTX 5090 — tutorials & explanations ↗Opens YouTube search results
Image source ↗
GPU · NVIDIA
NVIDIA H200
NVIDIA Corporation
Designs GeForce GPUs, data-center accelerators and Jetson edge AI platforms. Manufacturer website ↗
- Architecture
- Hopper; data-center accelerator
- Memory capacity
- 141 GB HBM3e per GPU
- Compute / bandwidth
- 4.8 TB/s memory bandwidth
- AI use
- Server-based large-model inference and training.
- Software
- CUDA, TensorRT-LLM, NVIDIA AI Enterprise
Requires a supported server platform. Bandwidth measures data movement, not operations per second.
Official specifications ↗▶ YouTube: NVIDIA H200 — tutorials & explanations ↗Opens YouTube search results
Image source ↗
GPU · AMD
AMD Instinct MI300X
Advanced Micro Devices, Inc.
Designs Ryzen CPUs, Ryzen AI processors and Instinct GPU accelerators. Manufacturer website ↗
- Architecture
- CDNA 3; data-center accelerator
- Memory capacity
- 192 GB HBM3 per accelerator
- Compute / bandwidth
- Up to 5.3 TB/s memory bandwidth
- AI use
- Large-model inference and training on compatible servers.
- Software
- ROCm; supported PyTorch builds
The capacity is for one accelerator, not the total memory of an eight-GPU platform. Check ROCm model and operating-system support.
Official specifications ↗▶ YouTube: AMD Instinct MI300X — tutorials & explanations ↗Opens YouTube search results
Image source ↗
TPU · Google
Google Cloud TPU v5e
Google
Designs Tensor Processing Units and offers them through Google Cloud. Manufacturer website ↗
- Architecture
- Tensor Processing Unit; cloud accelerator
- Memory capacity
- 16 GB HBM per chip
- Compute / bandwidth
- 197 peak BF16 TFLOPS per chip
- AI use
- Cloud model training and inference with TPU-compatible code.
- Software
- JAX, TensorFlow and PyTorch/XLA
Cloud service rather than a desktop add-in card. Per-chip figures differ from multi-chip slice or pod totals.
Official specifications ↗▶ YouTube: Google Cloud TPU v5e — tutorials & explanations ↗Opens YouTube search results
Image source ↗
NPU / SoC · AMD
AMD Ryzen AI 9 HX 370
Advanced Micro Devices, Inc.
Designs Ryzen CPUs, Ryzen AI processors and Instinct GPU accelerators. Manufacturer website ↗
- Architecture
- Zen 5 / Zen 5c CPU; XDNA 2 NPU; Radeon 890M GPU
- Memory capacity
- Shared system memory; laptop configuration dependent
- Compute / bandwidth
- 12 CPU cores; NPU up to 50 TOPS
- AI use
- Power-efficient on-device AI features and supported inference workloads.
- Software
- AMD Ryzen AI Software; supported ONNX models
50 TOPS is the NPU rating. Integrated graphics shares system RAM and does not have separate dedicated VRAM.
Official specifications ↗▶ YouTube: AMD Ryzen AI 9 HX 370 — tutorials & explanations ↗Opens YouTube search results
Image source ↗
NPU / SoC · Apple
Apple M4
Apple Inc.
Designs Apple silicon combining CPU, GPU and Neural Engine in one system on a chip. Manufacturer website ↗
- Architecture
- Apple silicon; CPU + GPU + Neural Engine
- Memory capacity
- Unified memory shared by CPU and GPU; device configuration dependent
- Compute / bandwidth
- Up to 10 CPU cores; 10 GPU cores; Neural Engine 38 trillion operations/s
- AI use
- Local inference, AI-assisted creative work and supported Apple-device applications.
- Software
- Metal, Core ML; MLX on supported Macs
Base M4 example; M4 Pro and Max are different chips. Neural Engine performance does not represent GPU performance.
Official specifications ↗▶ YouTube: Apple M4 — tutorials & explanations ↗Opens YouTube search results
Image source ↗
Edge AI · NVIDIA
NVIDIA Jetson Orin Nano Super
NVIDIA Corporation
Designs GeForce GPUs, data-center accelerators and Jetson edge AI platforms. Manufacturer website ↗
- Architecture
- Ampere GPU + Arm CPU; developer kit
- Memory capacity
- 8 GB shared LPDDR5 memory
- Compute / bandwidth
- Up to 67 sparse INT8 TOPS; 102 GB/s memory bandwidth
- AI use
- Robotics, camera inference and small edge AI applications.
- Software
- JetPack, CUDA, TensorRT
Super performance requires the supported software and power mode. Shared memory also serves the operating system.
Official specifications ↗▶ YouTube: NVIDIA Jetson Orin Nano Super — tutorials & explanations ↗Opens YouTube search results