A100 vs T4

The A100 offers 80GB with 312 TFLOPS for training, while the T4 provides 16GB with 65 TFLOPS for inference. Choose A100 for training, T4 for low-cost inference.

Jump to: Specs Pricing Offers Similar GPUs

Overview

Nvidia A100

The Nvidia A100 is a data center GPU on the Ampere architecture, released Q2 2020, with 80 GB HBM2e and 1,935 GB/s memory bandwidth on the PCIe variant.

High Performance Datacenter

Nvidia T4

The Nvidia T4 is a data center GPU on the Turing architecture, released Q3 2018, with 16 GB GDDR6 and 300 GB/s memory bandwidth.

Mid-Range Datacenter

Memory per GPU

5x more on A100

A100

80 GB HBM2e 40 GB on SXM, PCIe

16 GB GDDR6

Memory bandwidth

6.5x higher on A100

A100

1,935 GB/s 2,039 GB/s on SXM 1,555 GB/s on SXM, PCIe

300 GB/s

GPU interconnect

A100

NVLink 600 GB/s

PCIe only

Peak performance (FP16)

4.8x higher on A100

A100

312 TFLOPS

65 TFLOPS

Vendor peak figures, dense and per GPU. Not measured throughput.

FP16 per watt

1.1x higher on A100

A100

1.04 TFLOPS/W

0.93 TFLOPS/W

Board power

4.3x higher on A100

A100

300 W

70 W

Precision support

A100

FP16 BF16 TF32 FP32 FP64 INT4 INT8

FP16 FP32 INT4 INT8

Highlighted formats are supported by only one of the two.

Architecture

A100

Ampere

Turing

Released

2 years newer on A100

A100

Q2 2020

Q3 2018

Figures are for each model's headline configuration. Where a form factor carries a different number, it is named under the figure.

Source: A100 datasheet and T4 datasheet.

Price comparison

Cheapest on-demand

3.8x higher on A100

A100

$0.47 / GPU / hr

$0.12 / GPU / hr

Typical rate

2.2x higher on A100

A100

$1.77 / GPU / hr

$0.81 / GPU / hr

Median across providers, one vote each, so a large catalog does not outweigh a small one.

Spot floor

2.4x higher on A100

A100

$0.13 / GPU / hr

$0.06 / GPU / hr

Providers listing it

32 more on A100

A100

Price, last 90 days

A100

+2%

−6%

Change in the market-wide median, which moves when providers reprice and when cheaper or more expensive listings enter the market.

Heads up: Prices are estimates from published rates for common setups and vary by region and usage; verify with the provider before provisioning.

Offers and availability

How this list works

Order

Each provider appears once per GPU. We pick the one configuration that speaks for that provider, and rank on it in this order:

Availability: in stock, then waitlist, then not reported, then out of stock. A price you can act on outranks a cheaper one you cannot.
Billing type: on-demand first, then reserved, then spot, then quote-only. On-demand leads because it is the only rate you can hold without committing to anything.
Price: the lowest per-GPU hourly rate the provider lists, converted to USD.

The configuration we pick is chosen the same way, availability before billing before price. The rest of that provider's configurations are on the GPU's own page.

Ties sort alphabetically. Reserved rates are not separated by term length, so a three-year commitment can sit above a one-month one: the label beneath each price says which you are looking at. Providers who publish no rate are listed last within their group.

Transparency and funding

Ads and sponsors: Paid placements sit at the top of the list and are always labeled as sponsored content. Sponsorship never influences the organic ranking itself.
Affiliates: Affiliate links are marked. We may earn a commission if you click them, but commissions never affect the order.
Prices: Shown in USD, converted at daily reference rates where a provider publishes in another currency. A month means 720 hours. How we estimate costs.