Hardware partitioned

GPU instances with MIG on a 96 GB RTX 6000

Not everyone needs a whole card. With Multi-Instance GPU we split a 96 GB RTX 6000 Server Edition into hardware-isolated instances, each with its own memory, compute units and bandwidth. You pay for the fraction you use, not the card.

Talk to an engineer See dedicated GPU servers

96 GB

GDDR7 ECC vRAM per card

€509

per month from, with 24 GB

Madrid

Data4 · Tier III+

What is a GPU instance with MIG?

An EasyDataHost GPU instance is a physical partition of an NVIDIA RTX 6000 Server Edition 96 GB, created with Multi-Instance GPU (MIG). Unlike the time-slicing of vGPU, MIG divides the card in silicon: each instance gets its own slice of memory, its own compute units and its own share of bandwidth, and shares none of it. Rates current as of October 2026.

check_circle

Hardware isolation: another tenant cannot steal your compute or memory, because there is no scheduler handing out turns between you.

check_circle

Predictable performance: a 24 GB instance always delivers the same, whether the card is full or empty. That is what lets you size without a safety margin.

check_circle

The card’s real profiles: 1g.24gb, 2g.48gb and 4g.96gb, which are exactly the partitions NVIDIA supports on this GPU.

Pick your instance size

Three sizes, all on the same hardware: NVIDIA RTX 6000 Server Edition 96 GB, fifth-generation AMD EPYC 9555, DDR5 ECC Registered at 6400 MT/s and NVMe 5.0. No minimum term, with your data hosted in our Madrid data center.

GPU instance

24 GB vRAM

4 instancias por GPU · 1g.24gb

8 vCPU AMD EPYC 9555
32 GB DDR5 ECC Reg 6400
200 GB NVMe 5.0
2 Gbps unmetered

€509/month

GPU instance

48 GB vRAM

2 instancias por GPU · 2g.48gb

16 vCPU AMD EPYC 9555
64 GB DDR5 ECC Reg 6400
400 GB NVMe 5.0
2 Gbps unmetered

€942/month

GPU instance

96 GB vRAM

GPU completa · 4g.96gb

32 vCPU AMD EPYC 9555
128 GB DDR5 ECC Reg 6400
800 GB NVMe 5.0
2 Gbps unmetered

€1,891/month

Instance summary. Monthly prices excluding VAT, current as of October 2026.
Instance	vRAM	vCPU	RAM	NVMe 5.0	MIG profile	Price
24 GB	24 GB	8	32 GB	200 GB	`1g.24gb`	€509/month
48 GB popular	48 GB	16	64 GB	400 GB	`2g.48gb`	€942/month
96 GB	96 GB	32	128 GB	800 GB	`4g.96gb`	€1,891/month

What hardware is underneath?

The host is not filler. Each node runs an AMD EPYC 9555 with 64 cores and 128 threads at 3.2 GHz and 256 MB of L3 cache, DDR5 ECC Registered memory at 6400 MT/s and NVMe 5.0 storage, so the GPU never waits on the CPU or the disk while data loads.

memory

AMD EPYC 9555

Zen 5 · 64 cores, 128 threads at 3.2 GHz · 256 MB L3

developer_board

DDR5 ECC Registered

6400 MT/s · error correction in memory

storage

NVMe 5.0

PCIe Gen5 · datasets and checkpoints without a bottleneck

lan

2 Gbps unmetered

Dedicated public IP · private VLAN available

dns

NVIDIA RTX 6000 Server Edition

96 GB GDDR7 ECC · Blackwell architecture

shield

Data4 Madrid data center

Tier III+ · ISO 27001 and ENS High · GDPR

MIG, vGPU or a dedicated server?

The question is not which is better but how much GPU you need and whether you can tolerate sharing it. This table sums up the three options we offer, so you do not pay for a whole card when a quarter will do.

The three ways to get a GPU at EasyDataHost.
	MIG instance	Time-sliced vGPU	Dedicated GPU server
Isolation	In hardware	In software	Total
Performance	Constant	Varies with neighbours	Constant
Granularity	24, 48 or 96 GB	Flexible	Whole cards
Host access	No	No	Yes, root
From	€509/month	Not offered	€4,550/month

What is each size for?

24 GB

Inference on quantised models up to 13B, internal AI APIs, development and testing, embedding generation and RAG.

48 GB

Inference from 30B to 70B quantised, LoRA fine-tuning, computer vision pipelines and rendering.

96 GB

Light training, 70B models without aggressive quantisation, workloads that will not fit in half a card.

Frequently asked questions

How is a MIG instance different from an ordinary shared GPU? expand_more

MIG splits the card in hardware. A software-shared GPU hands out turns between processes: if your neighbour launches a heavy job, your performance drops. With MIG each instance is assigned its own memory, compute units and bandwidth, so what others do cannot reach you. It is the difference between a flatshare and your own flat.

Can I start with 24 GB and grow later? expand_more

Yes. All three sizes run on the same card model, so moving from 24 to 48 or 96 GB is a profile change, not a migration. There is no minimum term and the change is scheduled in an agreed window; the environment is usually recreated in under an hour.

What if I need more than one whole GPU? expand_more

Then what you want is a dedicated GPU server, with 2, 4 or 8 RTX 6000 Server Edition cards and root access to the machine, from €4,550 per month. MIG instances go up to one complete 96 GB card; beyond that the server pays off.

Does the data leave Spain? expand_more

No. Instances are hosted in our data center on the Data4 campus in Alcobendas, Madrid, certified to ISO 27001:2022 and ENS High, and the data stays on Spanish soil under GDPR. It is usually the deciding argument in public sector and healthcare.

What does the price include? expand_more

The GPU instance with its dedicated vRAM, the vCPUs on AMD EPYC 9555, DDR5 ECC memory, NVMe 5.0 storage, 2 Gbps of unmetered connectivity, a dedicated public IP, remote access and infrastructure support. NVIDIA AI Enterprise licences are optional and quoted separately. VAT not included.