MIG slicing
One H200, up to seven GPUs.
NVIDIA Multi-Instance GPU splits one H200 NVL into as many as seven slices, each with its own memory and compute, isolated in hardware. Pay for the slice, not the card.
Coming soonEnterprise early access is open
How it works
Three steps.
- 01
Choose a slice size
From one seventh of an H200 NVL up to the whole card.
- 02
Work as on any GPU
Notebooks, training and serving see the slice as a GPU of its own.
- 03
Grow when you need to
Move the same work to whole GPUs when it outgrows the slice.
See it
One GPU, seven ways. Each slice isolated in hardware.
141 GB
One GPU with 141 GB of memory.
What you get
MIG slicing, in detail.
Up to seven slices
As small as 16.5 GB each, or fewer, larger slices when a job needs more memory.
Isolated in hardware
Each slice has its own memory and compute. A neighbour's crash or out-of-memory error never reaches yours.
Sized to the job
Run a notebook or a small model on a slice that fits, so a whole GPU never sits half idle.
Made for classrooms
Give every student or researcher an isolated GPU of their own, from one card.
Under the hood
The key numbers.
- MIG instances, up to
- 7
- Memory per smallest instance
- 16.5 GB
- GPU memory
- 141 GB
- MIG instances, up to
- 7
- Memory per smallest instance
- 16.5 GB
- GPU memory
- 141 GB
- GPU
- NVIDIA H200 NVL
- Memory bandwidth
- 4.8 TB/s
- Currency
- Nepalese rupees (NPR)
- Ways to pay
- eSewa, Khalti, bank transfer or invoice
Billing
Per second, per slice
Commitment: none. Rates are being finalised; every request gets a firm quote in rupees.
How billing worksQuestions
About mig slicing.
What is MIG?
Multi-Instance GPU, an NVIDIA feature that partitions one GPU into separate instances, each with dedicated memory and compute.
How small can a slice be?
One seventh of an H200 NVL: 16.5 GB of memory with its own share of compute.
Start with mig slicing.
Tell us what you want to run. It takes about a minute.