Ampere
Nvidia's 2020 GPU architecture behind GeForce RTX 30 and the A100 accelerator
Ampere is the GPU microarchitecture Nvidia announced on 14 May 2020 as the successor to both Volta and Turing. Built on TSMC 7 nm for the A100 and a custom Samsung 8 nm process for consumer parts, it introduced third-generation Tensor Cores, second-generation ray tracing cores and multi-instance GPU virtualisation across GeForce RTX 30, RTX A and data centre products.
- Announced 14 May 2020, named after André-Marie Ampère
- A100: 6912 FP32/INT32 CUDA cores, 40 GB HBM2, 1.6 TB/s bandwidth
- Multi-instance GPU splits one A100 into up to seven instances
- Consumer GA10x chips use Samsung 8N; A100 uses TSMC 7 nm
- First GeForce generation with hardware AV1 decode
What it is
Ampere is the codename for a family of graphics processing unit designs from Nvidia, covering both consumer graphics cards and data centre accelerators. It was revealed on 14 May 2020 and named after the French mathematician and physicist André-Marie Ampère. The design carries CUDA Compute Capability 8.0 in the A100 and 8.6 in GeForce 30 parts, and supports DirectX 12 Ultimate, Vulkan 1.4, OpenGL 4.6 and OpenCL 3.0. Chips span the GA100, GA102, GA103, GA104, GA106, GA107 and GA10B dies, powering everything from laptops to the DGX A100 server.
How it came about
Nvidia positioned Ampere as the follow-on to two earlier architectures at once, Volta for compute and Turing for consumer graphics. The A100 accelerator arrived on the announcement date, 14 May 2020. Consumer cards followed at a GeForce Special Event on 1 September 2020, and an 80 GB version of the A100 was shown at SC20 that November. Mobile RTX parts and the RTX 3060 were unveiled on 12 January 2021. Hopper was presented as the next architecture at GTC 2022, with a further successor named at GPU Technology Conference 2021.
Inside the chip
Each streaming multiprocessor (SM) holds third-generation Tensor Cores handling FP16, bfloat16, TensorFloat-32 and FP64, plus sparsity acceleration. On GA100 a Tensor Core does 256 FP16 fused multiply-adds per clock, four times the earlier generation, while GA10x reaches twice that; the count drops to one Tensor Core per SM. Consumer GA10x chips double the FP32 cores per SM and add second-generation ray tracing cores that run alongside shading and compute. Cache is 128 KB of L1 per SM on consumer parts, 192 KB on professional ones, with 2 MB to 6 MB of L2. NVLink 3.0 delivers 50 Gbit/s per pair, and PCIe 4.0 is supported.
How it was made
Two processes were used. The A100 was built on TSMC's 7 nm FinFET node, while GeForce 30 consumer parts used a custom variant of Samsung's 10 nm-class process branded 8N. Memory options include HBM2 on the A100 40 GB and 80 GB, GDDR6X on the RTX 3090, 3080 Ti, 3080 and 3070 Ti, and GDDR6 elsewhere. Dies range from the large GA100 down to the small GA107 and the GA10B used in Tegra SoCs.
Where it went
Ampere reached desktops through the GeForce RTX 30 series, from the RTX 3050 up to the RTX 3090 Ti, and laptops through matching mobile parts plus the GeForce MX570 and RTX 2050. Workstation cards carry the RTX A naming, from the RTX A1000 to the RTX A6000. Data centre products include the A2, A10, A16, A30, A40 and A100. The A100 first shipped inside the third-generation DGX server, which pairs eight of them with two 64-core AMD Rome 7742 CPUs, 1 TB of RAM, 15 TB of PCIe gen 4 NVMe storage and HDR InfiniBand. Tegra parts GA10B and GA10F appear in AGX Orin, Orin NX, Orin Nano and the T239.
Versions and family
The die line-up runs GA100, GA102, GA103, GA104, GA106, GA107, GA10B and GA10F. GA100 serves the A30 and A100, GA102 the top GeForce cards and A10, A40, RTX A4500 to A6000, with smaller dies filling mid-range and laptop slots. An A100 80 GB variant doubled the memory of the original 40 GB part. A100X and A30X are also listed. On the timeline Ampere sits between Turing and Volta before it, and Hopper and Ada Lovelace after it.
Why it matters
Ampere unified Nvidia's compute and consumer lines under one architecture and brought data centre features such as multi-instance GPU partitioning, letting one A100 be split into as many as seven isolated instances. It also introduced hardware AV1 decoding and a five-core JPEG decode block, NVJPG, alongside NVENC video encode. The DGX A100 combined eight accelerators in one system, continuing a chain of DGX platforms built around successive Nvidia parts.
Questions people ask about Ampere
What is Ampere?
Ampere is the codename for a Nvidia GPU microarchitecture announced on 14 May 2020. It covers consumer GeForce RTX 30 graphics cards, RTX A workstation cards and data centre accelerators such as the A100, and it follows both Volta and Turing.
Who manufactures Ampere chips?
Nvidia designed the architecture and used two foundries. The A100 was made on TSMC's 7 nm FinFET process, while the GeForce 30 consumer parts were built on a custom version of Samsung's 8 nm process known as 8N.
What are Tensor Cores in Ampere?
Tensor Cores are units aimed at matrix maths used in AI. Ampere's third generation handles FP16, bfloat16, TensorFloat-32 and FP64 and speeds up sparse models. On GA100 each core does 256 FP16 fused multiply-adds per clock, four times the previous generation.
What is multi-instance GPU?
Multi-instance GPU, or MIG, is a virtualisation feature on the A100 that partitions one physical GPU into up to seven separate instances. Each instance gets its own share of the chip, so several users or workloads can run on a single accelerator.
Which products use Ampere?
The GeForce RTX 30 series and mobile variants, GeForce MX570 and RTX 2050, RTX A workstation cards, data centre A2, A10, A16, A30, A40 and A100, and Tegra parts in AGX Orin, Orin NX, Orin Nano and the T239.
Good reading and sources
- NVIDIA Ampere Architecture In-DepthNvidia, Article
- Checking your browserAnandTech, Article
- Nvidia’s first Ampere GPU is designed for data centers and AI, not your PCThe Verge, Article
- Diving Deep Into The Nvidia Ampere GPU ArchitectureThe Next Platform, Article
The facts on this page come from the Wikipedia article "Ampere (microarchitecture)", written by its contributors and used under the CC BY-SA 4.0 licence, and from its Wikidata record (CC0). EveryChip rewrote the text in its own words, with AI assistance, and compares every page with the article to make sure no passage is copied. It may be reused under the same CC BY-SA 4.0 licence with a link back.
It may contain mistakes, and figures change over time. Nothing here is investment or buying advice. If something is wrong, tell us. Last refreshed 2026-11-14.
Like Ampere
All graphics processorsMore from Nvidia
All of them
🇺🇸 United StatesApple M1
ARM-based system-on-a-chip for Macs and iPads, designed by Apple and manufactured by TSMC.SoC · Arm · 2020
🇺🇸 United StatesRDNA 2
AMD GPU microarchitecture for gaming and consolesGPU · GPU architecture · 2020