Machine Types

CPU Machine Types

For lightweight workloads that don’t require GPU acceleration — routing, preprocessing, API proxies.

Machine Type	RAM	CPU Cores
XS	512 MB	0.5
S	1 GB	1
M	2 GB	2
L	15 GB	4
XL	30 GB	8

GPU Machine Types

Machine Type	VRAM	RAM	CPU	Bandwidth	Video Enc / Dec
GPU-A100	40 GB	60 GB	12	2.0 TB/s	— / 5
GPU-L40	48 GB	100 GB	6	0.9 TB/s	3 / 3 (AV1)
GPU-H100	80 GB	112 GB	12	3.4 TB/s	— / 7
GPU-RTXPRO6000	96 GB	180 GB	14	1.8 TB/s	4 / 4 (AV1)
GPU-H200	141 GB	112 GB	12	4.8 TB/s	— / 7
GPU-B200	192 GB	210 GB	19	8.0 TB/s	— / 7

Video encode/decode counts refer to hardware NVENC/NVDEC engines — dedicated hardware units that encode or decode video independently of the GPU’s compute cores. GPUs with encoders (RTX PRO 6000, L40) can output video frames without using GPU compute time. GPUs marked -- for encode have no hardware encoder and require software encoding on the CPU.

Choosing a GPU

By VRAM requirement — pick the smallest GPU that fits your model:

40 GB (A100): General-purpose training and inference at a lower price point than Hopper GPUs
48 GB (L40): AI inference combined with video transcoding and graphics rendering
80 GB (H100): LLM inference and training, NVLink 4.0 at 900 GB/s for multi-GPU scaling
96 GB (RTX PRO 6000): Diffusion, large-image, and video generation with AV1 hardware encode; ample VRAM headroom for high-resolution and batched workloads
141 GB (H200): Large models and long-context workloads on a single GPU — 76% more memory and 43% more bandwidth than H100
192 GB (B200): Maximum memory and compute for the largest models, FP4/FP6/FP8 precision support

By workload type:

Image generation: L40 or RTX PRO 6000 (good throughput, generous VRAM)
Video generation: RTX PRO 6000 (4 NVENC + 4 NVDEC, AV1 encode, more VRAM) or L40 (3 NVENC + 3 NVDEC)
LLM inference: H100 or H200 (high bandwidth, large VRAM)
Training: A100, H100, or H200 (depending on model size)
Largest models: B200, RTX PRO 6000, or multi-GPU H100/H200

Configuration

Set the machine type in your application:

class MyApp(fal.App):
    machine_type = "GPU-H100"
    num_gpus = 1

Multiple Machine Types

Allow your app to use multiple machine types for a larger pool of available machines:

class MyApp(fal.App):
    machine_type = ["GPU-H100", "GPU-A100"]

Machine types are tried in order. If the first type has no available capacity, the next is used.

Multi-GPU

For models that need more than one GPU:

class MyApp(fal.App):
    machine_type = "GPU-H100"
    num_gpus = 2

Multi-GPU Workloads

Learn how to distribute inference across multiple GPUs

Changing Machine Types

Via Code: Update machine_type and redeploy:

class MyApp(fal.App):
    machine_type = "GPU-A100"

fal deploy

machine_type is a code-specific parameter — it always comes from your code and resets on every deploy. See Scaling Configuration for details.

Setting Up

Model APIs

Serverless

Compute

Organizations

Machine Types

CPU Machine Types

GPU Machine Types

Choosing a GPU

Configuration

Multiple Machine Types

Multi-GPU

Multi-GPU Workloads

Changing Machine Types

​CPU Machine Types

​GPU Machine Types

​Choosing a GPU

​Configuration

​Multiple Machine Types

​Multi-GPU

Multi-GPU Workloads

​Changing Machine Types

CPU Machine Types

GPU Machine Types

Choosing a GPU

Configuration

Multiple Machine Types

Multi-GPU

Changing Machine Types