Skip to main content

AI Compute Card History Timeline (2006–2026)

From general-purpose GPU to dedicated AI accelerators, two decades of compute evolution.

2006–2011: The CUDA Era Begins

YearEventSignificance
2006NVIDIA G80 (8800 GTS) launched, CUDA 1.0 releasedYear zero for general-purpose GPU computing — GPUs are no longer just graphics cards
2007NVIDIA Tesla C870 launchedFirst dedicated GPU compute card, no display output
2008NVIDIA Tesla C1060 (GT200)240 CUDA cores, first dual-precision support
2009NVIDIA Fermi (Tesla C2050)ECC memory, error correction, enters HPC market
2010NVIDIA CUDA 3.0Official C++ support, precursor to cuDNN
2011NVIDIA Tesla M2090 (Fermi)665 GFLOPS double precision, supercomputing standard

2012–2016: The Deep Learning Boom

YearEventSignificance
2012AlexNet (CUDA)GPU-accelerated deep learning makes a name for itself
2012NVIDIA K20 (Kepler)15.9 TFLOPS FP32, HPC explosion
2013NVIDIA Tesla K4012GB memory, first 12GB GPU
2014NVIDIA CUDA 6.0Unified Memory, simplifying GPU programming
2015NVIDIA Tesla P100 (Pascal)First HBM memory, HPC memory bandwidth revolution
2016NVIDIA Tesla P100 (production)NVLink 1.0, 160 GB/s GPU-to-GPU interconnect

2016–2020: The Birth of AI-Specific Accelerators

YearEventSignificance
2016Google TPU v1 (inference)First non-GPU architecture AI accelerator, ~92 TOPS INT8
2017Google TPU v2 (training + inference)180 TFLOPS BF16, TPU Pod rack
2017NVIDIA V100 (Volta)Tensor Core debut, dedicated AI training units
2018Google TPU v3420 TFLOPS BF16, liquid-cooled TPU Pod
2019NVIDIA A100 (Ampere)MIG Multi-Instance GPU, 7nm, 40GB HBM2
2020Google TPU v4 (Intrepid)1,086 TFLOPS BF16, sparse architecture

2021–2023: The Large Model Era

YearEventSignificance
2021NVIDIA H100 (Hopper)Transformer Engine, FP8 support, king of LLM training
2022Google TPU v5e (Cydonia)Single-chip inference optimization, 1,028 TFLOPS
2022Google TPU v5p (Cydonia)Training-optimized, 1,444 TFLOPS, 32GB HBM
2023NVIDIA H200141GB HBM3e, 4.8 TB/s, optimized for large model inference
2023AMD MI300X192GB HBM3, competing with H100, open-source ROCm
2023Cerebras WSE-22.6 trillion transistors, 85 PFLOPS, wafer-scale
2023Intel Gaudi 26nm, first product after Habana Labs acquisition

2024–2025: Blackwell & Domestic Alternatives

YearEventSignificance
2024 Q1NVIDIA B200 (Blackwell)9 PFLOPS, 192GB HBM3e, 2nm process
2024 Q2NVIDIA GB2002× B200 + Grace CPU, LLM inference monster
2024 Q3AMD MI325X256GB HBM3e, competing with H200
2024 Q4AMD MI350X (CDNA 4)288GB HBM3e, 9.2 PFLOPS
2024 Q4Google TPU v6e (Trillium)918 TFLOPS, 32GB HBM, GA
2025 Q1Intel Gaudi 31,600 TFLOPS, 128GB SRAM, competing with H100
2025 Q2NVIDIA B300 Ultra14 PFLOPS, 288GB HBM3e, 1,400W
2025 Q2AMD MI355X10.1 PFLOPS (MXFP6), 288GB HBM3e
2025 H2Huawei Ascend 910CDual-die, 780 TFLOPS, domestic alternative workhorse

2026 (Announced / Expected)

ProductVendorKey Details
NVIDIA Rubin R200NVIDIA288GB HBM4, 50 PFLOPS FP4 (sparse), Vera CPU
AMD MI400 + HeliosAMD432GB HBM4, 40 PFLOPS FP4, 260 TB/s UALink
Google TPU Ironwood (v7)Google192GB HBM, ~2,000 TFLOPS BF16
Cerebras WSE-4Cerebras1.4 trillion transistors, 125 PFLOPS FP8
AWS Trainium 3AWS3nm, ~2,000 TFLOPS, 128GB HBM
Huawei Ascend 920Huawei900+ TFLOPS BF16, 4 Tbps inter-chip interconnect

Compute Growth Curve (FP16 Training Compute)

Year Product FP16 Compute
----------------------------------------
2006 Tesla C870 ~0.5 TFLOPS
2010 Tesla C2050 ~1.3 TFLOPS
2016 Tesla P100 ~10 TFLOPS
2017 V100 ~15 TFLOPS
2020 A100 ~312 TFLOPS
2021 H100 ~1,979 TFLOPS
2023 H200 ~1,979 TFLOPS (memory ↑)
2024 B200 ~4,500 TFLOPS
2025 B300 Ultra ~7,000 TFLOPS
2026 Rubin R200 ~10,000 TFLOPS (FP4: 50 PFLOPS)
2026 AMD MI400 ~10,000 TFLOPS (FP4: 40 PFLOPS)

Trend: ~4× compute growth every 2–3 years (Moore's Law × Tensor Core architectural innovation)

Key Architecture Evolution

NVIDIA Architecture Codename

ArchitectureYearKey Innovation
Tesla (G80)2006CUDA 1.0, general-purpose GPU computing
Fermi2010ECC memory, double precision
Kepler2012GPU Boost, dynamic frequency
Maxwell2014Energy efficiency optimization
Pascal2016HBM memory, NVLink 1.0
Volta2017Tensor Core, dedicated AI training
Turing2018RT Core, consumer ray tracing
Ampere2020MIG, multi-instance GPU
Hopper2022Transformer Engine, FP8
Blackwell20242nm, 9 PFLOPS, 192GB HBM3e
Rubin2026HBM4, 50 PFLOPS FP4, Vera CPU

Google TPU Generations

GenerationYearKey Innovation
v12016First ASIC AI accelerator, inference-only
v22017Training + inference, TPU Pod
v32018Liquid cooling, ~420 TFLOPS
v42020Sparse architecture, 1,086 TFLOPS
v5e2022Inference-optimized, single-chip
v5p2022Training-optimized, 32GB HBM
v6e (Trillium)2024918 TFLOPS, 32GB HBM
v6p2024Training-optimized, 64GB HBM
8t / 8i2026Training/inference split architecture
Ironwood (v7)2026192GB HBM, ~2,000 TFLOPS

Contribute & Corrections

Spot an error or missing a key event in the timeline? Submit a PR to help us improve!


← Back to Home | Future Roadmap → | Full Comparison Table →