A Visual Timeline of NVIDIA AI GPUs: From GTX 580 to Vera Rubin
TL;DR: Fourteen years, six architectures, and one consistent pattern — every generation buys you more memory and more bandwidth, and the precision keeps getting smaller. AlexNet ran on 3 GB. A single Rubin R200 carries 288 GB. The chart below is the whole story in one screen.
Almost every "which GPU should I use" argument is really an argument about history: which generation you're on, and what that generation was built to do. Here's the map.
The timeline
Each bar runs from launch until the next generation took over — roughly, the window in which that part was the newest thing you'd buy.
The specs behind it
| Year | GPU | Memory | Bandwidth | What it unlocked |
|---|---|---|---|---|
| 2012 | GTX 580 | 3 GB GDDR5 | 192 GB/s | AlexNet — proof that CUDA beat CPU clusters |
| 2014 | Tesla K80 | 24 GB GDDR5 | 480 GB/s | The first widely rented cloud AI GPU |
| 2016 | Tesla P100 | 16 GB HBM2 | 732 GB/s | HBM2 and NVLink — multi-GPU stops being PCIe-bound |
| 2017 | Tesla V100 | 32 GB HBM2 | 900 GB/s | First Tensor Cores; mixed-precision training goes mainstream |
| 2018 | Tesla T4 | 16 GB GDDR6 | 320 GB/s | INT8 inference at 70 W — serving gets cheap |
| 2020 | A100 | 80 GB HBM2e | 2.0 TB/s | TF32, structured sparsity, MIG partitioning |
| 2022 | H100 | 80 GB HBM3 | 3.4 TB/s | Transformer Engine and FP8 |
| 2024 | H200 | 141 GB HBM3e | 4.8 TB/s | Long context and bigger KV caches on one GPU |
| 2025 | B200 | 192 GB HBM3e | 8 TB/s | NVFP4 and rack-scale NVL72 domains |
| 2025 | B300 / GB300 | 288 GB HBM3e | 8 TB/s | 15 PFLOPS NVFP4 — reasoning-model inference |
| 2025 | RTX PRO 6000 Blackwell | 96 GB GDDR7 | 1.8 TB/s | The most VRAM on any discrete card — workstation training without HBM prices |
| 2025 | DGX Spark (GB10) | 128 GB unified | 273 GB/s | Blackwell on a desk — capacity without a datacenter |
| 2026 | Rubin R200 | 288 GB HBM4 | ~20 TB/s | 50 PFLOPS NVFP4; HBM4's first outing |
| 2026 | RTX Spark (N1X) | 128 GB unified | 273 GB/s | The same GB10 silicon in a Windows box — ships this fall |
Figures come from NVIDIA's own product pages where the part still has one. Rubin R200 entered full production in June 2026 and ships in volume through the second half of the year, so its numbers are NVIDIA's launch disclosures rather than field measurements.
Memory is the axis that actually moved
Compute got faster, but memory capacity is what decided which models you could run at all:
Note the flat spots — and the interlopers. Memory shrank from K80 to P100 when NVIDIA traded GDDR5 capacity for HBM2 bandwidth, and it sat at 80 GB across two generations while Hopper spent its transistor budget on FP8 instead.
Then look at the three bars that aren't datacenter parts at all: RTX PRO 6000 at 96 GB, DGX Spark and RTX Spark at 128 GB. Each one carries more memory than an H100 — the part that defined frontier training two years earlier. That's the real story of 2025–26: capacity stopped being a datacenter privilege. Bandwidth very much did not, which is why a 128 GB desktop and an 80 GB H100 are good at completely different things.
Three things changed, over and over
- Precision shrank. FP32 → FP16 (Volta) → TF32 (Ampere) → FP8 (Hopper) → NVFP4 (Blackwell). Each halving roughly doubles throughput and halves the memory a model occupies. Most of the "10× faster" generational claims are a precision change wearing a costume.
- Memory got closer. GDDR5 → HBM2 → HBM3e → HBM4, and 192 GB/s → 20 TB/s. For decode-bound inference this is the only number that matters.
- The unit of compute got bigger. A GPU, then an NVLink pair, then an 8-GPU node, now a 72-GPU NVL72 rack addressed as one accelerator.
The consumer line, briefly: GTX 1080 (2016, 8 GB) → RTX 2080 Ti (2018, 11 GB) → RTX 3090 (2020, 24 GB) → RTX 4090 (2022, 24 GB) → RTX 5090 (2025, 32 GB GDDR7). Fast, but capacity-capped — which is why 70B-class local work keeps landing on unified-memory machines instead.
Two recent ways around that cap. The RTX PRO 6000 Blackwell, announced at GTC on 18 March 2025 and on sale through PNY that April, puts 96 GB of GDDR7 on one card — the largest frame buffer NVIDIA has ever shipped on a discrete GPU, and the reason a lot of workstation fine-tuning stopped needing datacenter hardware. And the RTX Spark, announced 31 May 2026 at GTC Taipei and shipping this fall in ASUS, MSI, Acer and GIGABYTE laptops and mini-desktops, is the same GB10/N1X silicon as the DGX Spark in a Windows box — same chip, different machine.
Where this leaves you in 2026
The timeline splits into two useful questions:
- Do you need throughput? You want Hopper or Blackwell in a datacenter — H100, H200, or B200-class — and you'll pay per hour for bandwidth you fully saturate.
- Do you need capacity? DGX Spark's 128 GB of unified memory holds models an 80 GB H100 can't, at 273 GB/s. Slower per token, but it fits — and that's the difference between running an experiment and not running it.
Fourteen years of this timeline says the same thing each generation: the constraint moves, but it's almost always memory first.
FAQ
Which NVIDIA GPU was used to train AlexNet?
Two GeForce GTX 580s, each with 3 GB of GDDR5. The 2012 paper split the network across both cards specifically because it did not fit in one — the memory constraint that has shaped every generation since.
What is the newest NVIDIA AI GPU?
The Rubin R200, which entered full production in June 2026. It carries 288 GB of HBM4 at roughly 20 TB/s and 50 PFLOPS of NVFP4 — about 2.5× the memory bandwidth of a B200 and 6.6× that of an H100.
How much memory does each NVIDIA data center GPU have?
V100 32 GB, A100 80 GB, H100 80 GB, H200 141 GB, B200 192 GB, B300/GB300 288 GB, and Rubin R200 288 GB. Capacity did not rise smoothly: it fell from the K80's 24 GB to the P100's 16 GB when NVIDIA traded GDDR5 capacity for HBM2 bandwidth, and it sat flat at 80 GB across both Ampere and Hopper.
Do I need an H100, or is a DGX Spark enough?
It depends on whether your bottleneck is bandwidth or capacity. An H100 moves 3.4 TB/s against 80 GB; a DGX Spark moves 273 GB/s against 128 GB. Models that don't fit in 80 GB run on the Spark and not on the H100, but they run far slower per token — so Spark suits development, evaluation and memory-bound experimentation, and H100-class hardware suits production throughput.
Is the RTX Spark the same as the DGX Spark?
The silicon is identical — NVIDIA's CEO confirmed the GB10 and the N1/N1X are one chip, with 128 GB of unified memory and about 1 PFLOP of FP4. The machines differ: DGX Spark is a Linux developer box with a ConnectX-7 port for clustering two units, while RTX Spark is a single Windows machine shipping in fall 2026. See same chip, different machine for the full comparison.