What MIG Is, and When a Fractional GPU Is Enough

MIG (Multi-Instance GPU) is an NVIDIA feature that partitions one physical data centre GPU into as many as seven smaller GPUs, each with its own memory and its own isolated share of the hardware. A MIG GPU slice is enough when your whole job fits inside the slice's memory and you do not need the full card's bandwidth or more than one GPU. Before you rent one, check it against the price of a whole cheap GPU, because a slice of an expensive card is often poor value next to an entire inference or consumer card.

How MIG divides a GPU

NVIDIA's MIG user guide (updated 11 September 2026) says the feature "allows GPUs (starting with NVIDIA Ampere architecture) to be securely partitioned into up to seven separate GPU Instances". Each instance has "separate and isolated paths through the entire memory system", which gives "a defined quality of service (QoS) with fault isolation". In plain terms, this is GPU slicing done in hardware. Your slice has memory that no other tenant can touch, and a crash or a heavy job in the next slice should not reach yours.

The guide's supported-GPU list and NVIDIA's product pages give the following, as read on 21 September 2026. "Not recorded" means our research did not capture a figure, not that none exists.

Two things to notice. First, the A100's 10 GB figure can only describe the 80 GB card, since seven such slices would not fit in the 40 GB version. Second, the slices do not add up to the whole card. Seven 10 GB slices are 70 GB of an 80 GB H100, and seven 18 GB slices are 126 GB of a 141 GB H200. NVIDIA's pages do not explain the difference in the text retrieved, so plan around the published slice size and nothing more.

"Up to seven" also means fewer, larger instances are possible. We did not record NVIDIA's full profile list, but rental listings show the idea. On 21 September 2026 Baseten's pricing page listed an "H100 MIG (40 GiB)", which is half of an 80 GB card, and Runpod's serverless price list included "MIG 48GB" and "MIG 24GB" tiers.

GPUs without MIG

MIG starts with the Ampere architecture, so older data centre cards are out. (Our guide to NVIDIA GPU generations shows which cards sit where.) The T4 is a Turing part from 2018 and the V100 is Volta, which puts both before the cut-off.

Newer does not mean supported either. NVIDIA's L40S spec table has the row "Multi-Instance GPU (MIG) Support: No". GeForce cards are absent from the supported list, which covers the RTX 4090 and RTX 5090. The L4 was not among the GPUs we recorded from NVIDIA's list, and we did not confirm it either way. Check the list before you trust a listing that calls a fraction of an L4 a MIG slice.

That matters when you read a listing. If someone offers you a fraction of an L40S or an RTX 4090, it is not a MIG slice, whatever the listing calls it.

MIG is not time-slicing

There are two ways to let several customers use one GPU.

MIG divides the hardware. Each tenant gets a fixed piece of memory and compute, and NVIDIA's guide says this provides "fault isolation for different clients such as VMs, containers or processes".

Time-slicing shares the GPU in turns. Tenants run one after another on the same silicon, and time-sliced vGPU does not give the hardware isolation that MIG does. Nothing in that arrangement stops a host from placing more tenants on a card than it can serve well at once, which is oversubscription. We have no source showing that any particular provider does this. It is simply the question the design leaves open, and your throughput depends on how busy your neighbours are.

So the time slicing vs MIG question comes down to what is guaranteed. With MIG, the memory and compute on the listing are yours. With a time-shared GPU, what you get depends on who else is running. Our research did not cover NVIDIA's other sharing modes, so this page does not describe them.

How fractional GPUs appear on rental sites

A fractional GPU shows up as a listing with less than one GPU, or as a product name that carries a memory size smaller than the card's. Three examples from 21 September 2026 show how different these can be.

• Runpod's documentation lists a "PRO 6000 MIG 24GB" as a mid-range option for 7B to 13B inference. The RTX PRO 6000 Blackwell supports four MIG instances, and the name says MIG.
• Vultr's plan API lists fractions of a card, such as an A16 plan with 2 GB of VRAM and an A40 plan with 2 GB. Neither card is on the MIG list as we recorded it, so ask how the split is done.
• Hugging Face ZeroGPU "dynamically allocates and releases" RTX PRO 6000 Blackwell GPUs, and its default large size is half a GPU with 48 GB. That is a quota of GPU minutes, not a rented slice.

Some platforms do not share at all. Vast.ai's docs say instances get "exclusive GPU access" and the GPU is "never shared between users".

Before renting a fraction, get three answers from the provider.

1. Is the memory dedicated? A MIG slice has its own. For anything else, ask whether the stated memory is reserved for you.
2. Is it MIG or software sharing? If the underlying card does not support MIG, you have your answer.
3. Is the card oversubscribed? Ask how many tenants can be placed on one GPU and whether that number is fixed.

We are not listing which providers sell MIG slices beyond the examples above, because we have not verified it provider by provider.

When a slice is enough

Start with memory, because a slice has a hard limit. NVIDIA's inference optimisation post (17 November 2023) says a 7 billion parameter model loaded in 16-bit precision takes roughly 14 GB. Runpod's documentation, read on 21 September 2026, gives the same thing as a rule of thumb: about 2 GB of VRAM per billion parameters. At 8-bit the weights need about 1 byte per parameter, and at 4-bit about half a byte. How to read GPU specs for AI walks through the full sizing method.

Then add the KV cache. NVIDIA's example in the same post is Llama 2 7B at FP16, batch size 1 and 4,096 tokens, which needs about 2 GB.

Put those together for a 10 GB slice of an A100 or H100. A 7B model at 16-bit does not fit. At 8-bit it is about 7 GB of weights plus about 2 GB of cache for one sequence, which fits with almost nothing left for batching. The 18 GB slice of an H200 holds the same model at 16-bit with a little room. Our LLM VRAM calculator does this sum for other model sizes.

A slice works well for:

• Inference on small models: embeddings, classifiers, speech, and language models of a few billion parameters.
• Notebooks and development, where the GPU sits idle most of the time while you think.
• CI jobs that need a CUDA device to run tests, not throughput.
• Several small services that must not interfere with each other.

A slice is the wrong tool for:

• Latency-sensitive LLM serving on a larger model. NVIDIA's post describes the decode phase as "a memory-bound operation", limited by how fast weights and cache move from memory. The bandwidth number in a spec table belongs to the whole card. A slice gets its own isolated part of the memory system, not all of it. We have no published per-slice bandwidth figure, so do not assume one.
• Training and full fine-tuning. Hugging Face's model memory guide (read 21 September 2026) gives 6 bytes per parameter for mixed-precision weights, 8 for Adam states and 4 for gradients. That adds up to 18 bytes per parameter before activations, or about 126 GB for a 7B model. No slice holds that.
• Anything that spans GPUs. If the job needs several GPUs exchanging data over NVLink, you need whole cards in one server. See our best GPU for LLM guide for that case.

The whole-GPU price a slice has to beat

A slice is never competing with the card it was cut from. It competes with the cheapest whole GPU that can do the same job. The table below is live, and the L4, T4 and RTX 4090 rows are the ones that matter most here.

GPU	Cheapest $/GPU-hr	Provider	Providers in stock
A100	$0.68	LeaderGPU	10
H100	$2.50	Hyperstack	9
H200	$3.43	QuantaCloud	7
L4	$0.49	RunPod	3
Tesla T4	none in stock
RTX 4090	$0.53	Vast.ai	3

Here is what those whole cards give you, from our datasheet-verified spec table.

Spec	A100	H100	H200	L4	T4	RTX 4090
VRAM	40 to 80 GB	80 to 94 GB	141 GB	24 GB	16 GB	24 GB
Memory bandwidth	2,039 GB/s	3,350 GB/s	4,800 GB/s	300 GB/s	320 GB/s	1,008 GB/s
Architecture	Ampere	Hopper	Hopper	Ada Lovelace	Turing	Ada Lovelace
Launch year	2020	2022	2024	2023	2018	2022
Interconnect	NVLink, PCIe 4.0, InfiniBand	NVLink, PCIe 5.0, InfiniBand	NVLink, PCIe 5.0, InfiniBand	PCIe 4.0	PCIe 3.0	PCIe 4.0

A whole T4 has 16 GB. A whole L4 or RTX 4090 has 24 GB. Each of those is more memory than a 10 GB or 12 GB H100 slice, with no neighbours on the card and the full memory bandwidth of that GPU to yourself. The honest position is that slices are often poor value against a cheap whole inference or consumer GPU. The table lets you check that on the day you rent, instead of taking our word for it.

Whole cheap GPU or slice of an expensive one

Use this rule.

1. Work out the memory you need: weights at your chosen precision, plus cache, plus headroom.
2. Find the cheapest whole GPU in the live table with at least that much memory.
3. Rent the slice only if it is cheaper than that whole GPU, it is a true MIG instance with dedicated memory, and your job is on the "works well" list above.
4. If the prices are close, take the whole GPU. You get more memory, no shared silicon and nothing to ask the provider about.

There are two fair reasons to pick the slice anyway. One is that you need to stay with a provider that only offers data centre cards, and a whole A100 or H100 is far more than the job needs. The other is that you want hardware isolation on data centre infrastructure, and the cheap whole consumer cards available to you come from individual marketplace hosts, which Vast.ai's own docs say "may have less formal security measures". Outside those cases, a whole small GPU is the default, and a slice has to earn its place on price. Browse all GPUs to see what else fits your memory budget, or compare the two big cards in A100 vs H100 if the job has outgrown a slice.

Sources

Source pages were retrieved with automated research tools on the dates shown and each figure was traced back to its source before publishing. Rental prices on this page are not typed: they are read live from GPUperhour's own data.

Every source was accessed on 21 September 2026. A date in an entry is the publication or last-update date.

• NVIDIA MIG User Guide: introduction, updated 11 September 2026
• NVIDIA MIG User Guide: supported GPUs, updated 11 September 2026
• NVIDIA H100 product page
• NVIDIA H200 product page
• NVIDIA L40S product page
• Runpod docs: Choose a Pod
• Runpod pricing
• Baseten pricing
• Vultr cloud GPU plans API
• Hugging Face docs: Spaces ZeroGPU
• Vast.ai docs: instances overview
• Vast.ai docs: security FAQ
• NVIDIA blog: Mastering LLM Techniques, Inference Optimization, published 17 November 2023
• Hugging Face docs: model memory anatomy