How do I check which GPU I'm renting?

Run nvidia-smi. It shows each card's name, memory, temperature, power use and driver version. In Python, torch.cuda.get_device_name(0) gives the name.

How do I tell if a GPU is overheating or power-limited?

Run nvidia-smi -q -d PERFORMANCE while it works. Active slowdown reasons such as thermal or power cap mean the card is holding itself back.

What memory bandwidth should a healthy card reach?

A simple copy test usually reaches about 80 to 90% of the card's published bandwidth: roughly 800 to 900 GB/s on an RTX 4090 (1,008 GB/s on paper).

For developers · Updated October 13, 2026

Check a rented GPU in five minutes: card, memory, speed and network

KernelMachine is preparing an invite-only beta; no launch date is set. Where this guide describes KernelMachine, it's how it works at launch. Join the waitlist

Short answer: run nvidia-smi to confirm the card and its memory, run the two short PyTorch tests below to measure memory bandwidth and matrix speed, check for heat or power limits while they run, and time a large download. A healthy card reaches about 80 to 90% of its published bandwidth. Anything far below that, or a different card than you paid for, is a reason to stop and switch.

Marketplace renters report cards that were quietly slowed down, machines that lost a GPU mid-rental, and downloads at a fraction of the listed speed (why rentals fail). Five minutes of checks at the start catch most of it before you've spent much.

1. Is it the card you paid for?

$ nvidia-smi
$ nvidia-smi --query-gpu=name,memory.total,memory.used,temperature.gpu,power.draw,power.limit,pcie.link.gen.current,pcie.link.width.current --format=csv

Check the name, the total memory, and that little memory is in use before you start. A low power limit (well under the card's usual figure) or a narrow PCIe link (x1 or x4) will slow some jobs.

2. Measure memory bandwidth

Memory bandwidth decides how fast chat models answer. This copies 2 GB back and forth and reports the speed:

import time, torch

a = torch.empty(512 * 1024**2, dtype=torch.float32, device="cuda")  # 2 GiB
b = torch.empty_like(a)
for _ in range(3):
    b.copy_(a)
torch.cuda.synchronize()
start = time.time()
for _ in range(20):
    b.copy_(a)
torch.cuda.synchronize()
seconds = time.time() - start
print(f"{2 * a.numel() * 4 * 20 / seconds / 1e9:.0f} GB/s")

Compare with the published figure; expect about 80 to 90% of it:

Card	Published bandwidth
RTX 4090	1,008 GB/s
RTX 5090	1,792 GB/s
RTX 3090	936 GB/s
RTX 4080	717 GB/s
RTX 3090 Ti	1,008 GB/s
RTX 5080	960 GB/s
RTX 5070 Ti	896 GB/s
RTX 4070 Ti SUPER	672 GB/s
RTX 5060 Ti 16GB	448 GB/s
RTX 4060 Ti 16GB	288 GB/s

3. Measure matrix speed

Training and image generation lean on raw compute. This times a large 16-bit matrix multiply:

import time, torch

n = 8192
x = torch.randn(n, n, device="cuda", dtype=torch.float16)
y = torch.randn(n, n, device="cuda", dtype=torch.float16)
for _ in range(3):
    x @ y
torch.cuda.synchronize()
start = time.time()
for _ in range(20):
    x @ y
torch.cuda.synchronize()
print(f"{2 * n**3 * 20 / (time.time() - start) / 1e12:.0f} TFLOPS")

There's no single right answer here, because results depend on the PyTorch version and settings. Run the same test on another card of the same model, or on a later rental, and compare. A result far below others usually means heat or a power limit.

4. Look for heat and power limits

While a test or your job runs, in a second terminal:

$ nvidia-smi -q -d PERFORMANCE

Under "Clocks Event Reasons" (older drivers say "Clocks Throttle Reasons"), "Active" next to a thermal or power item means the card is holding itself back. A little power capping under full load is normal; constant thermal slowdown isn't.

5. Check disk and download speed

$ df -h /workspace
$ curl -s -o /dev/null -w '%{speed_download}\n' 'https://speed.cloudflare.com/__down?bytes=25000000' | awk '{printf "%.0f Mbit/s\n", $1 * 8 / 1e6}'

Big models mean big downloads: at 100 Mbit/s a 20 GB model takes about half an hour. If the machine is far slower than you need, switch before downloading.

If something's off

Stop early: the first minutes are the cheapest time to switch. Note what you measured, and report it to the provider.

On KernelMachine, every card passes an automatic hardware test before renters see it (most hosts' machines also a 48-hour trial), and billing starts only when your app is ready. If a rental still isn't right, run km report "what's wrong" or tap "It's broken". The rental stops and is refunded automatically: the current 10 minutes, or the whole rental if it ends within 30 minutes of becoming ready (up to 3 times in 30 days). All km commands

The first check also works as a one-off job, from your own computer, with the card stopping as soon as it answers:

$ km run --size 24GB -- nvidia-smi --query-gpu=name,memory.total,driver_version --format=csv