diff --git a/Makefile b/Makefile new file mode 100644 index 0000000..0b41816 --- /dev/null +++ b/Makefile @@ -0,0 +1,12 @@ +PORT ?= 8000 + +.PHONY: site open-site + +# Serve the website (site/) at http://localhost:$(PORT), as gpushed.dev serves it. +site: + @echo "Serving site/ at http://localhost:$(PORT) (Ctrl+C to stop)" + python3 -m http.server $(PORT) -d site + +# Open the page straight from disk, no server. +open-site: + open site/index.html diff --git a/site/fonts/LICENSE.txt b/site/fonts/LICENSE.txt new file mode 100644 index 0000000..b3f0950 --- /dev/null +++ b/site/fonts/LICENSE.txt @@ -0,0 +1,16 @@ +Meslo LG (MesloLGM-Regular, MesloLGM-Bold), subset to WOFF2 for this site. +Source: https://github.com/andreberg/Meslo-Font (v1.2.1) + +Copyright 2009, 2010, 2013 André Berg + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. diff --git a/site/fonts/MesloLGM-Bold.woff2 b/site/fonts/MesloLGM-Bold.woff2 new file mode 100644 index 0000000..6dd35d7 Binary files /dev/null and b/site/fonts/MesloLGM-Bold.woff2 differ diff --git a/site/fonts/MesloLGM-Regular.woff2 b/site/fonts/MesloLGM-Regular.woff2 new file mode 100644 index 0000000..02319f0 Binary files /dev/null and b/site/fonts/MesloLGM-Regular.woff2 differ diff --git a/site/index.html b/site/index.html index d721629..85bb6ef 100644 --- a/site/index.html +++ b/site/index.html @@ -3,10 +3,15 @@ -gpushed: your dedicated NVIDIA GPU, on your Mac - +gpushed: NVIDIA GPUs for your Mac (A100, L4, T4 with CUDA) + + + + + + - + @@ -15,14 +20,61 @@ - + + + + + +
-
+
+
Mac only · Apple silicon · macOS 26+ +

Your dedicated NVIDIA GPU, on your Mac.

-

A whole NVIDIA A100, L4 or Tesla T4 for your research, one command from your Mac’s terminal. Boots ready to work, billed per second, deleted when you're done.

+

Your Mac can't run CUDA. Now it can. Get a whole NVIDIA A100, L4 or Tesla T4 for fine-tuning, inference and research with one command. Ready in minutes, billed per second, deleted when you're done.

  • Preinstalled
  • NVIDIA driver
  • @@ -308,6 +504,94 @@

    Your dedicated NVIDIA GPU, on your Mac.

    boxuser@gp-b7k2:~$ python train.py▌
+
+ +
+
+
Minutesfrom gpushed up to SSH
+
$0.88/hfor a 24 GB L4, per second
+
16–40 GBdedicated GPU memory
+
5regions across US, Europe, Asia
+
+
+ +
+
+
Use cases
+

What people run on a gpushed box.

+

Anything that needs CUDA and more VRAM than your Mac has. A few of the common ones, with the plan that fits each best:

+
+ +
+
+ + l4 · a100 +
+

Fine-tune open LLMs

+

LoRA and QLoRA on Llama, Mistral, Qwen or Gemma with PEFT, TRL or Unsloth. A 7B–8B model fits on an L4; reach for the A100 for 13B+ or longer contexts.

+
$ pip install torch peft trl bitsandbytes
+$ python finetune.py
+
+ +
+
+ + l4 +
+

Serve and test models

+

Stand up vLLM, TGI or Ollama behind an OpenAI-compatible API and test your model before you pick production hosting.

+
$ pip install vllm
+$ vllm serve Qwen/Qwen2.5-7B-Instruct
+
+ +
+
+ + t4 · l4 +
+

Generate images and video

+

Run Stable Diffusion, SDXL, Flux or ComfyUI workflows with diffusers, train LoRAs on your own style, and batch-render without heating up your laptop.

+
$ pip install diffusers transformers
+$ python sdxl.py
+
+ +
+
+ + any plan +
+

Write CUDA and Triton kernels

+

Apple silicon has no CUDA, so write, compile and run real kernels here with nvcc or Triton, on a real NVIDIA GPU.

+
$ nvcc -O3 matmul.cu -o matmul
+$ ./matmul
+
+ +
+
+ + l4 · a100 +
+

Reproduce papers and run evals

+

Clone the repo, pip install -r requirements.txt, and rerun the experiments. Run lm-eval-harness or your own benchmarks on a clean, known machine.

+
$ git clone github.com/…/paper-code
+$ lm_eval --model hf --tasks mmlu,gsm8k
+
+ +
+
+ + any plan +
+

Jupyter and VS Code, remotely

+

Keep your editor on your Mac and your tensors on the GPU. Forward a port for JupyterLab, or point VS Code Remote-SSH at the box.

+
$ ssh -i ~/.gpushed/state/box_ssh_key \
+    -L 8888:localhost:8888 boxuser@<ip>
+→ open localhost:8888 on your Mac
+
+ +
+
+
@@ -334,7 +618,25 @@

Get a GPU

-
+
+
+
Choosing a GPU
+

Which NVIDIA GPU fits your workload?

+

VRAM is usually what decides it. When in doubt, start on the L4: it's the cheapest and you can switch any time.

+ + + + + + + + + +
WorkloadPlanWhy
Learning CUDA, notebooks, small modelst4 / l416–24 GB is plenty, from $0.88/hour.
QLoRA / LoRA on 7B–8B LLMsl424 GB fits a 4-bit 8B model with room for batches.
Serving a 7B–8B model with vLLMl4fp16 weights fit, with headroom for the KV cache.
Stable Diffusion, SDXL, ComfyUIt4 / l4Both run SDXL in fp16; the L4 is noticeably faster.
13B+ fine-tuning, long contexts, big batchesa10040 GB of HBM2 and far more memory bandwidth.
+
+
+ +
Built for research

A GPU that's yours, only while you need it.

@@ -374,16 +676,17 @@

Private by default

-
+
Pricing
-

Simple hourly prices, charged by the second.

+

Simple hourly GPU prices, charged by the second.

Each price includes the machine, a 100 GB SSD and a public IP. Prepay from $10 and top up whenever you like. US prices shown; Europe and Asia cost a little more.

-
+
Budget

Tesla T4

16 GB · 8 vCPU · 30 GB RAM
+

Notebooks, Stable Diffusion, learning CUDA.

$0.91 / hour
gpushed up t4
@@ -391,6 +694,7 @@

Tesla T4

Default · best value

L4

24 GB · 4 vCPU · 16 GB RAM
+

7B–8B LoRA, vLLM serving, SDXL.

$0.88 / hour
gpushed up l4
@@ -398,6 +702,7 @@

L4

Large models

A100 40 GB

40 GB · 12 vCPU · 85 GB RAM
+

13B+ fine-tuning, long contexts.

$4.44 / hour
gpushed up a100
@@ -410,12 +715,24 @@

A100 40 GB

-
+
FAQ

Questions

+
+ Can I run CUDA on a Mac? +

Not locally: Apple silicon has no NVIDIA GPU, and CUDA needs one. gpushed gives your Mac a dedicated NVIDIA GPU in the cloud with CUDA 12.6 ready, one command away, so PyTorch, vLLM, Triton and your own kernels run from your Mac's terminal.

+
+
+ Which GPU should I pick for fine-tuning? +

For LoRA or QLoRA on 7B–8B models, the L4 (24 GB) is the best value. Pick the A100 (40 GB) for 13B+ models, longer contexts, bigger batches, or full fine-tuning of smaller models. See choosing a GPU.

+
+
+ Can I use Jupyter or VS Code? +

Yes. Connect with plain SSH using the key in ~/.gpushed/state/box_ssh_key (the IP is in gpushed status) and forward a port for Jupyter, or add the box as a host in VS Code Remote-SSH.

+
Why Mac only?

gpushed's client runs inside Apple's native container runtime, which needs a Mac with Apple silicon on macOS 26 or later. Focusing on one platform keeps setup to a single command.

@@ -428,12 +745,12 @@

Questions

What's on the box?

Ubuntu 22.04 with the NVIDIA driver and CUDA Toolkit 12.6, plus Python 3.11 (pip and uv), Node.js 24 LTS, Rust, git and build tools. pip install torch and go, with full root access for anything else.

+
+
Where does it run?

On Google Cloud, in the region nearest you: Iowa, South Carolina or Oregon in the US, the Netherlands, or Singapore. If it's out of GPUs, gpushed tries its other zones and then the next nearest region automatically. Pick one yourself with gpushed up --region europe-west4. Prices differ a little by region; gpushed plans shows yours.

-
-
What happens when my credit runs out?

You get an email when your balance drops to $2. Just before it reaches zero your box is stopped, with its disk kept. Top up and run gpushed start to resume. A box that stays unfunded for 24 hours is deleted.

@@ -455,17 +772,20 @@

Questions

-
-
-

Start your next experiment on a real GPU.

-

For Macs with Apple silicon on macOS 26 or later.

-
- $ - curl -fsSL gpushed.dev/install.sh | bash - +
+ +
+
+

Start your next experiment on a real GPU.

+

For Macs with Apple silicon on macOS 26 or later.

+
+ $ + curl -fsSL gpushed.dev/install.sh | bash + +
-
-
+
+