The Silicon Zoo
Vetted Specifications for AI Accelerators and Edge Devices
The Silicon Zoo is the authoritative registry for physical hardware in mlsysim. Every specification is typed (pint.Quantity), provenance-tracked, and validated against official datasheets and MLPerf baselines, so a figure such as the A100’s bandwidth (2.04 TB/s) has one agreed value.
Reference these specs when reasoning about bottlenecks. Any device listed here loads directly in Python: hw = mlsysim.Hardware.Cloud.A100. The three columns that matter most for roofline analysis are Peak Performance, Memory BW, and Capacity.
Data Center Accelerators
| Device | Year | Peak Performance | Memory BW | Capacity | TDP |
|---|---|---|---|---|---|
| NVIDIA B200 | 2024 | 2.2 PFLOPs/s | 8.0 TB/s | 193.3 GB | 1,000 W |
| Cerebras CS-3 (WSE-3) | 2024 | 125.0 PFLOPs/s | 21.0 PB/s | 47.2 GB | 23,000 W |
| d-Matrix Corsair | 2024 | 2.4 PFLOPs/s | 150.0 TB/s | 2.1 GB | 600 W |
| NVIDIA GB200 NVL72 | 2024 | 162.0 PFLOPs/s | 576.0 TB/s | 14.8 TB | 120 kW |
| Intel Gaudi 3 | 2024 | 1.8 PFLOPs/s | 3.7 TB/s | 137.4 GB | 900 W |
| AMD Instinct MI325X | 2024 | 1.3 PFLOPs/s | 6,000.0 GB/s | 274.9 GB | 1,000 W |
| Reference Desktop CPU | 2024 | 1.0 TFLOPs/s | 50.0 GB/s | 68.7 GB | 150 W |
| Google TPU v6e (Trillium) | 2024 | 918.0 TFLOPs/s | 1,638.0 GB/s | 34.4 GB | 300 W |
| AWS Trainium 2 | 2024 | 380.0 TFLOPs/s | 2.4 TB/s | 103.1 GB | 500 W |
| NVIDIA H200 | 2023 | 989.0 TFLOPs/s | 4.8 TB/s | 151.4 GB | 700 W |
| AMD MI300X | 2023 | 1.3 PFLOPs/s | 5.3 TB/s | 206.2 GB | 750 W |
| Google TPU v5p | 2023 | 459.0 TFLOPs/s | 2.8 TB/s | 102.0 GB | 300 W |
| Untether AI speedAI240 | 2023 | 2.0 PFLOPs/s | 1.0 PB/s | 249.6 MB | — |
| Intel Gaudi 2 | 2022 | 432.0 TFLOPs/s | 2.5 TB/s | 103.1 GB | 600 W |
| NVIDIA H100 | 2022 | 989.0 TFLOPs/s | 3.4 TB/s | 85.9 GB | 700 W |
| NVIDIA H100 PCIe | 2022 | 756.0 TFLOPs/s | 2.0 TB/s | 85.9 GB | 350 W |
| AMD MI250X | 2021 | 383.0 TFLOPs/s | 3.2 TB/s | 137.4 GB | 500 W |
| Google TPU v4 | 2021 | 275.0 TFLOPs/s | 1,200.0 GB/s | 34.4 GB | 200 W |
| NVIDIA A100 | 2020 | 312.0 TFLOPs/s | 2,039.0 GB/s | 85.9 GB | 400 W |
| NVIDIA A100 40GB | 2020 | 312.0 TFLOPs/s | 1,555.0 GB/s | 42.9 GB | 400 W |
| Graphcore Colossus Mk2 GC200 IPU | 2020 | 250.0 TFLOPs/s | 45.0 TB/s | 943.7 MB | — |
| Groq LPU | 2020 | 188.0 TFLOPs/s | 80.0 TB/s | 241.2 MB | 215 W |
| Intel SGX Enclave (Reference) | 2020 | 1.0 TFLOPs/s | 10.0 GB/s | 134.2 MB | — |
| NVIDIA T4 | 2018 | 65.0 TFLOPs/s | 320.0 GB/s | 17.2 GB | 70 W |
| Google TPU v3 | 2018 | 123.0 TFLOPs/s | 900.0 GB/s | 34.4 GB | 450 W |
| Google TPU v1 | 2017 | 92.0 TFLOPs/s | 34.0 GB/s | 8.6 GB | 75 W |
| Google TPU v2 | 2017 | 46.0 TFLOPs/s | 700.0 GB/s | 17.2 GB | 280 W |
| NVIDIA V100 | 2017 | 125.0 TFLOPs/s | 900.0 GB/s | 34.4 GB | 300 W |
| NVIDIA V100 PCIe | 2017 | 112.0 TFLOPs/s | 900.0 GB/s | 34.4 GB | 250 W |
| NVIDIA Tesla P100 | 2016 | 21.2 TFLOPs/s | 732.0 GB/s | 17.2 GB | 300 W |
| NVIDIA Tesla K80 (one GK210 GPU) | 2014 | 4.4 TFLOPs/s | 240.0 GB/s | 12.9 GB | 150 W |
Workstations
| Device | Year | Peak Performance | Memory BW | Capacity | TDP |
|---|---|---|---|---|---|
| NVIDIA DGX Spark (GB10) | 2025 | 125.0 TFLOPs/s | 273.0 GB/s | 137.4 GB | 200 W |
| MacBook Pro (M3 Max) | 2023 | 14.2 TFLOPs/s | 400.0 GB/s | 137.4 GB | 100 W |
| NVIDIA GeForce GTX 580 (3 GB) | 2010 | 1.6 TFLOPs/s | 192.4 GB/s | 3.2 GB | 244 W |
Mobile Devices
| Device | Year | Peak Performance | Memory BW | Capacity | TDP |
|---|---|---|---|---|---|
| Google Pixel 8 (Tensor G3) | 2023 | 15.0 TFLOPs/s | 60.0 GB/s | 8.6 GB | 5 W |
| Snapdragon 8 Gen 3 | 2023 | 45.0 TFLOPs/s | 77.0 GB/s | 12.9 GB | 5 W |
| iPhone 15 Pro (A17 Pro) | 2023 | 35.0 TFLOPs/s | 51.2 GB/s | 8.6 GB | 5 W |
| Apple M2 Neural Engine | 2022 | 15.8 TFLOPs/s | 100.0 GB/s | 17.2 GB | 20 W |
Edge & Robotics
| Device | Year | Peak Performance | Memory BW | Capacity | TDP |
|---|---|---|---|---|---|
| NVIDIA DRIVE Thor | 2025 | 1.0 PFLOPs/s | 546.0 GB/s | 68.7 GB | 100 W |
| Edge Server | 2024 | 1.0 TFLOPs/s | 100.0 GB/s | 137.4 GB | 300 W |
| NVIDIA Jetson Orin NX | 2023 | 25.0 TFLOPs/s | 102.0 GB/s | 17.2 GB | 25 W |
| NVIDIA Jetson Orin Nano | 2023 | 10.0 TFLOPs/s | 68.0 GB/s | 8.6 GB | 15 W |
| NVIDIA Jetson AGX Orin | 2022 | 275.0 TFLOPs/s | 204.0 GB/s | 68.7 GB | 60 W |
| RoboTaxi Reference Compute (NVIDIA DRIVE AGX Orin class) | 2022 | 5.2 TFLOPs/s | 200.0 GB/s | 34.4 GB | 60 W |
| Intel NUC + Movidius | 2020 | 1.0 TFLOPs/s | 25.0 GB/s | 17.2 GB | 15 W |
| Google Coral Edge TPU | 2019 | 4.0 TFLOPs/s | 8.0 GB/s | 1.1 GB | 2 W |
TinyML Microcontrollers
| Device | Year | Peak Performance | Memory BW | Capacity | TDP |
|---|---|---|---|---|---|
| Lockstep safety MCU (reference class) | 2024 | 800.0 MFLOPs/s | 1.6 GB/s | 4.2 MB | 1 W |
| Oura Ring 4 (wearable reference profile) | 2024 | 100.0 MFLOPs/s | 0.1 GB/s | 2.1 MB | 0 W |
| ESP32-S3 (AI) | 2022 | 500.0 MFLOPs/s | 0.1 GB/s | 4.2 MB | 0 W |
| Himax WE-I Plus | 2020 | 200.0 MFLOPs/s | 0.1 GB/s | 2.1 MB | 0 W |
| Nordic nRF52840 (Cortex-M4F) | 2018 | 64.0 MFLOPs/s | 0.1 GB/s | 1.0 MB | 0 W |
| STM32F407 (Cortex-M4F) | 2011 | 168.0 MFLOPs/s | 672.0 MB/s | 196.6 KB | — |
How to Read the Silicon Zoo
The Three Numbers That Matter
For roofline analysis, focus on three columns:
Peak Performance (TFLOP/s): the compute ceiling. This determines how fast compute-bound workloads run (e.g., large-batch training, LLM pre-fill).
Memory Bandwidth (TB/s): the memory ceiling. This determines how fast memory-bound workloads run (e.g., small-batch inference, LLM token decoding).
Capacity (GB): the memory wall. If the model plus activations exceed this, the workload is infeasible on a single device.
The Ridge Point
The ratio of Peak Performance to Memory Bandwidth gives the ridge point (in FLOP/byte), such as 153 FLOP/byte on the A100 and 295 FLOP/byte on the H100. Workloads with arithmetic intensity below the ridge point are memory-bound; above it, compute-bound. See the Math Foundations page for the full derivation.
Common Patterns
- Cloud GPUs (A100, H100, H200) have 80–141 GiB of HBM with high bandwidth (2.04–4.80 TB/s). They are designed for throughput.
- Edge devices (NVIDIA Jetson Orin NX at 25 W) trade peak performance for lower power budgets, making TDP per TFLOP a useful comparison metric.
- TinyML MCUs (STM32F4 at 192 KiB, nRF52840 at 1 MiB) have KB- to MB-scale memory, so only the smallest quantized models fit. Use the Model Zoo to find matching workloads.
Textbook Connection
These specifications are used throughout Volumes 1 and 2 of the textbook. The Hardware Acceleration chapter uses them for roofline construction, and the Compute Infrastructure chapter uses them for fleet sizing and TCO analysis.
Custom hardware specs can be defined on the fly in Python, and new vetted specs can be contributed to the registry. See the Contributing Guide for how to add persistent specs, or the Hardware API Reference for defining custom HardwareNode objects.
Browse the Silicon Zoo from the terminal: mlsysim zoo hardware
For full technical specs and validation details, see the API Reference.