One or two graphics cards for local AI? 2026 Multi-GPU Guide

September 22, 2026

Two graphics cards.
Not twice as fast.

Multi-GPU setups are for something else. Here is what, and how many cards to configure based on your usage.

More VRAMyes, that's the real gain

Faster for one userno, or barely

More usersyes, significantly

This is the question everyone moving upmarket asks: rather than buying a more expensive card, why not put in two? The intuition is simple. It is also largely incorrect, for a mechanical reason that must be understood before spending money.

A useful clarification before we begin. On our machines, the number of graphics cards is a configuration choice, not a fixed product. Depending on the platform, our machines accept up to three cards to choose in the configurator, and four or more cards by quote. You adjust this number at the time of ordering. The question, therefore, is not "which machine to buy," but "how many cards to configure for my usage." This guide answers that question.

The myth of two cards

Why your conversation will not be faster

In text generation, a model produces one token at a time. For each token, it re-reads its entire weights from video memory. Speed therefore depends on the bandwidth of a single card, not on how many there are.

When a model is split across two cards, the layers are processed in sequence: the first card calculates its part, transmits the result, and the second takes over. At any given moment, only one card is working. The other is waiting.

Result: a model split across two cards runs at roughly the speed of one, with a slight communication overhead. This is not a configuration defect. It is the nature of the computation.

One card

Two cards, card A

Two cards, card B

Simplified diagram of layer distribution. Each card processes half the model, taking turns. The number of tokens produced per second remains comparable to that of a single card.

Something many overlook: the RTX 4090 and 5090 do not have NVLink. This high-speed inter-card connection has disappeared from consumer ranges since the RTX 3090. Two RTX 5090s therefore communicate via the PCIe bus, which is much slower. For inference, this is acceptable, as data exchanged between cards is measured in megabytes per token. For training, where gradients are constantly synchronized, it is a real bottleneck.

What two cards really do

What multi-GPU provides

Increased memory. Two 32 GB cards allow you to load a 64 GB model. This is the main gain, and often the only one that matters.
Multiplied throughput for shared usage. Multiple users or several agents in parallel: each card processes its own requests. Here, the gain is real.
Longer context. The freed-up VRAM is used for context cache, allowing for long windows on large models.
Distinct tasks in parallel. A language model on one card, image generation on the other.

What multi-GPU does not do

Speed up a single conversation. Throughput in tokens per second follows the bandwidth of a single card.
Share image generation. Tools like ComfyUI and video models do not natively distribute across multiple cards.
Replace a professional card for training. Without NVLink, gradient synchronization via PCIe slows things down significantly.
Work with a standard motherboard. Consumer platforms share their PCIe lanes between the two cards.

How many cards to configure? The decision tree

Does your model and its context fit into 32 GB?

Yes

A single card is enough The simplest and fastest choice for individual use. Configure an RTX 5090 and keep the ability to add more later.

Multiple simultaneous users or a lot of parallel processing?

Yes

Two cards or more 64 GB with two RTX 5090s, 384 GB with four RTX PRO 6000 Blackwells, or even more upon request. Throughput effectively multiplies under shared load.

A higher-capacity card A 96 GB RTX 6000 Blackwell offers more memory than a dual-5090 setup, without any inter-card communication, plus ECC.

The rule that summarizes it all: multi-GPU is an answer to memory limitations or the need for shared throughput. It is never a speed optimization for a single user. If your need is simply "more VRAM for me," a single higher-capacity card is almost always the best choice, and it spares you from platform constraints.

How far can you go? Configuration levels

The number of cards you can install depends on the platform, and especially its PCIe lanes. This is where the real difference between a consumer base and a professional base lies.

Configuration	Total VRAM	Platform	Target usage
1 × RTX 5090	32 GB	Consumer Tower	Individual use, models up to 70B in 4-bit
2 × RTX 5090	64 GB	4U Rack, 2,000 W	Team, long contexts on large models
1 × RTX 6000 Blackwell	96 GB ECC	Workstation or rack	Large models solo, fine-tuning, ECC
2 × RTX 6000 Blackwell	192 GB ECC	4U Rack, Threadripper PRO	Production, many users
3 × RTX 6000 Blackwell	288 GB ECC	Threadripper PRO, full PCIe lanes	Research, training, very large models
4 × RTX PRO 6000 Blackwell	384 GB ECC	Threadripper PRO, more cards upon request	Very large models, multiple models in parallel

Why Threadripper PRO is a game changer. A consumer processor provides 16 to 24 PCIe lanes. With two cards, they are shared as x8 and x8. The Threadripper PRO platform provides far more, which allows for powering up to four RTX PRO 6000 Blackwells, or more upon request, totaling 384 GB of ECC video memory with four cards. This is the only clean way to go beyond two cards without choking the bus.

Which card to choose before deciding on the quantity

Choosing the card model precedes choosing the quantity. Here is how the two references compare, in single-card versus multi-card configurations. Both references are available in our AI machine range, in both single-card and multi-card versions.

Criteria	RTX 5090 (32 GB)	2 × RTX 5090 (64 GB)	RTX 6000 Blackwell (96 GB)
Total VRAM	32 GB	64 GB	96 GB
70B model in 4-bit	Fits, short context	Comfortable	Wide margin
Speed, single user	Maximum	Comparable to one card	High, no bottleneck
Multiple users	Limited	Multiplied throughput	Good, shared memory
Image and video generation	Excellent	One card at a time	Excellent
Fine-tuning	LoRA up to approx. 13B	Limited by PCIe	QLoRA on 70B
ECC memory	No	No	Yes
Expandable up to	3 cards in configurator, 4+ by quote	3 cards in configurator, 4+ by quote	3 cards in configurator on TR PRO, 4+ by quote
GPU power consumption	575 W	1,150 W	600 W

An honest reading of this table. The RTX 5090 alone has made multi-GPU unnecessary for many use cases: its 32 GB can run a 70B model in 4-bit with short context. Two 5090s remain relevant for shared throughput and long contexts. The RTX 6000 Blackwell is the go-to as soon as ECC, fine-tuning, or solo capacity comes into play, and it is also the only one that can be installed in configurations of four or more upon request on the Threadripper PRO platform.

The hidden costs of a second card

The price of the card is only the beginning. A serious dual-GPU configuration involves four items that component lists often forget.

PCIe lanes

Consumer processors offer only 16 to 24 PCIe lanes. With two cards, a standard motherboard splits them into x8 and x8. This is why a dual-card configuration from us is a rack format, and why three cards require a Threadripper PRO platform, the only one capable of providing enough lanes to power them fully.

Power supply

One RTX 5090 consumes 575 W. With two cards, you need a 1,500 W minimum power supply, and preferably 2,000 W Platinum. Beyond that, for a three-professional-card configuration, electrical and thermal sizing is an integral part of the study we conduct with you.

Cooling and chassis

Two cards under load produce twice the heat, in a space where they interfere with one another. You need a chassis with a well-thought-out airflow, and often a rack format that physically separates the cards.

System memory

The model passes through RAM before being loaded into VRAM. Rule of thumb: at least twice the total VRAM in system memory. For 64 GB of VRAM, account for 128 GB of RAM.

How software distributes the model

Two approaches coexist, and they are not equivalent.

Layer-based distribution. This is what Ollama and llama.cpp do automatically. Each card receives a slice of the model and processes them in sequence. Little communication between cards, therefore PCIe-tolerant, but no speed gain solo. This is the default mode and the right choice for consumer hardware.

Tensor parallelism. This is what vLLM offers. Each operation is broken down between the cards, which work simultaneously. A real speed gain is possible, but it requires high bandwidth between cards. Without NVLink, the benefit remains modest and depends heavily on available PCIe lanes.

On our dual-GPU machines, we configure the distribution mode based on your usage: layer-based for simplicity and stability, tensor-based for shared throughput when the platform allows it. These settings are preconfigured and tested before delivery.

Our platforms, configurable in number of cards

Each machine is assembled by hand in Auriol (13390) and is fully configurable, including the number of graphics cards. Adjust the configuration in our online configurator, or describe your usage to us for a custom quote at contact@radiancesystems.eu.

Workstation, one card

CoreIA 64: RTX 5090 32 GBThe default choice for individual use. Graphics card configurable according to your needs. €8,642

Rack, two cards

CoreIA Rack: 2 × RTX 5090, 64 GB4U Rack, 128 GB RAM, 2,000 W. Shared throughput and long contexts. €17,887

CoreIA 128 Rack: 2 × RTX 6000, 192 GB ECC70B model in native precision, dozens of users, training. €42,452

Threadripper PRO, from one to four cards, or more on request

Pro AI Ultra: up to 4 × RTX PRO 6000 Blackwell96 to 288 GB ECC VRAM in the configurator (1 to 3 cards), 4 cards or more by quote. RAM expandable to 2 TB. €26,890

The most scalable platform in our range. The Pro AI Ultra is built on Threadripper PRO, whose PCIe lanes can power up to four RTX PRO 6000 Blackwell cards, or even more on request, providing 384 GB of ECC video memory with four cards. You can start with a single card and add more later without changing your machine. This is the main advantage of this base for those anticipating future scaling needs.

The single-GPU alternative

GB10 Mini Server: 128 GB UnifiedSingle chip, 128 GB unified memory, 240 W. No inter-card bus management required. €6,299

We guide you toward the right configuration, not the most expensive one. If your use case is individual and your model fits within 32 GB, we will tell you that a single card is sufficient. Multi-GPU setups address a specific need, which we verify with you before providing any quote. And since the number of cards is configurable, you can start small and upgrade the machine later.

In short

Do two graphics cards make local AI faster?
Not for a single user. Generation speed follows the bandwidth of a single card, and the cards work in turns. The real gain is elsewhere: more memory and multiplied throughput when several people or agents are using the server simultaneously.

Do the RTX 5090s support NVLink?
No. NVLink has been absent from consumer cards since the RTX 3090. Two RTX 5090s communicate via PCIe, which is suitable for inference but slows down training.

Should I choose two RTX 5090s or one RTX 6000 Blackwell?
For solo capacity, the RTX 6000 is better: 96 GB on a single card with no inter-card data exchange or platform constraints, plus ECC. For shared throughput among multiple users, two 5090s remain a viable option.

What is the maximum number of cards I can install?
It depends on the platform. A consumer tower supports two cards with shared PCIe lanes. Our Pro AI Ultra station, built on Threadripper PRO, supports up to four RTX PRO 6000 Blackwell cards, or more, providing 384 GB of ECC video memory with four cards: up to three cards can be selected in the configurator, four or more by quote.

Is the number of cards fixed at the time of order?
No. All our machines are configurable, including the number of graphics cards. You select the configuration in the online configurator or via a quote, and you can start with one card and add more later if the platform allows it.

Can I put two cards in a standard PC?
Physically sometimes, but PCIe lanes are then split into x8/x8, and the power supply, cooling, and system memory must be sufficient. This is why our configurations with two or more cards use a rack format or the Threadripper PRO platform.

Does image generation benefit from two cards?
Not natively. ComfyUI and video models use one card at a time. However, two cards allow you to run a language model on one and image generation on the other.

The indications in this guide are estimates based on public documentation and feedback from the ecosystem. Real-world performance depends on the inference engine, distribution mode, platform, and model used. For precise sizing, please contact us.

Back to the blog

Country/region

Language