NVIDIA Tesla T4: Passive Cooling, Power and Licensing
The NVIDIA Tesla T4 is a low-profile single-slot accelerator designed for servers. Three things about it differ from a graphics card in ways that catch people, and one of them can damage the card.
It has no fan
The most important physical fact about this card.
The T4 is passively cooled. There is a heatsink and no fan, because the card was designed for a server chassis where fans at the front pull air through the whole machine front to back.
Three consequences.
It will overheat in a workstation or tower with gentle airflow, even one that looks well ventilated. The card relies on forced air moving through its heatsink, and a case fan blowing generally around the chassis does not provide that.
Chassis fan configuration matters. Many servers require additional fans or a higher fan profile when an accelerator is fitted. Fitting the card without them means thermal throttling — reduced performance with no error reported, which is genuinely hard to diagnose. Our fans guide covers that failure mode.
Card position affects cooling. Where the chassis documents which slots are supported for accelerators, that is a thermal statement rather than an electrical one.
If your machine is not a server with front-to-back airflow, this is the wrong card — and no amount of adding case fans reliably fixes it.
No display output, and that is deliberate
Worth stating because it surprises people buying their first accelerator.
The T4 has no video ports. It is a compute card — it accelerates workloads and returns results rather than driving a monitor.
So it does not replace a graphics card, and a machine using it still needs whatever display output it had. In a server that is usually the onboard management video, which is fine because nobody sits at a server.
Power: what it does not need
A genuine advantage worth understanding.
The T4 draws its power entirely from the PCIe slot. It needs no supplementary power connector, which is why it fits servers that could not host a larger accelerator.
That matters practically. Larger accelerators need auxiliary power cables that are chassis-specific parts and frequently not present in a machine that did not ship with one — our accelerator guide covers that whole set of requirements. The T4 sidesteps it.
Still check total power headroom. Adding any accelerator raises draw, and in a rack near its circuit limit that matters — our PDU guide covers why each redundant path must carry the whole load alone.
Slot requirements
Two checks.
Lanes. An accelerator wants a full-width electrical connection, and physical slot size is not the same as electrical width — a card in an under-wired slot works at reduced throughput with nothing reported as a fault. On dual-socket boards, some slots are wired to the second processor and are dead with only one fitted. Our PCIe guide covers both.
Bracket. The card is low-profile, which suits 1U servers — but confirm which bracket ships with it, since a card supplied with only one type needs the other sourced separately.
Also confirm your platform supports accelerators at all, and in which slots. Some servers require a firmware level or a specific riser for accelerator support.
The licensing question
The part most often missed, and it can cost more than the card.
Where the T4 is used for virtual desktops or shared GPU across virtual machines, NVIDIA's virtual GPU software requires licences separate from the hardware. Buying the card does not grant them.
Where the card is used for compute in a single machine — inference, transcoding, analytics on bare metal — that licensing generally does not apply.
So the question to settle before ordering is what the card will actually do, and whether the answer involves virtualising it. Our accelerator guide covers this alongside the other requirements, and it is the same class of trap as per-core software licensing on processors.
Where the T4 fits
An honest positioning.
Good fit: inference workloads, video transcoding, virtual desktop acceleration, and analytics where the job fits within the card's memory. Its appeal is that it delivers useful acceleration within a power and cooling budget that ordinary servers already have.
Poor fit: large training workloads, or anything needing more memory than the card carries. Accelerator memory is a hard limit — a job that does not fit does not run slowly, it does not run.
Wrong fit: a workstation or tower without server airflow, and any machine where you actually wanted a graphics card.
Before you order
Your server model and generation, and confirmation it supports accelerators. Which slots are supported, their electrical width, and how many processors are fitted. Whether additional fans or a higher fan profile are required. Your power headroom. And what the card will do, because that decides the licensing question.
Common questions
Can I use a T4 in a desktop or workstation?
Not sensibly. It is passively cooled with no fan, designed for a server chassis that pulls air front to back. It will overheat in a tower with gentle airflow, and adding case fans does not reliably fix it.
Does it have video outputs?
No. It is a compute card that accelerates workloads and returns results rather than driving a monitor, so it does not replace a graphics card. In a server the onboard management video handles display, which is fine because nobody sits at a server.
Does it need an auxiliary power cable?
No, and that is a real advantage — it draws power entirely from the PCIe slot. Larger accelerators need auxiliary cables that are chassis-specific parts and frequently absent from machines that did not ship with one.
Do I need a licence to use it?
For virtual desktops or sharing the GPU across virtual machines, yes — NVIDIA’s virtual GPU software requires licences separate from the hardware. For compute on bare metal such as inference or transcoding, that licensing generally does not apply.
What happens if my job needs more memory than the card has?
It does not run slowly — it does not run. Accelerator memory is a hard limit rather than a performance characteristic, which makes it the first specification to check against your workload.
Send us your server model and what the card will be doing, and we will confirm slot support, cooling requirements and whether licensing applies.
