Do I share the GPU with anyone?

No. One physical machine, one account, and the card passed through to the operating system you installed -- no hypervisor, no MIG partition, no vGPU profile, no time-slicing scheduler, no neighbour's batch job on your silicon. The numbers you benchmark on day one are the numbers you keep.

Will two cards fit in one machine?

Often, and it is quoted rather than ticked: chassis clearance, the card's physical width, the lane budget and the power supply set the count, so an honest answer needs the specific part. Ask, and the reply names the chassis, the count and the lead time. The usual motive is memory rather than speed -- two 48 GB cards hold a model one 80 GB card cannot.

Which driver and CUDA release ships on a Server Room GPU server?

The ones you install. The machine arrives with a clean operating system and root; you pin the driver, the CUDA or ROCm release and the framework build to what your code was validated against, and nothing moves them afterward. If you want the driver in place before handover, put the exact version on the order.

What if the card I want is not on a shelf today?

Then it is fitted from stock or bought in, and you are told which before you pay: stocked parts seat within four to twenty-four hours, bought-in parts run five to ten working days, and allocation-bound accelerators ten to fifteen. When the requirement is a machine today, https://serverroom.net/instant lists what is racked and powered right now.

Can the card change after deployment?

Yes, in the machine you already hold: no migration, no reinstall, no new addresses. It is a scheduled rack visit with a short outage, and the invoice moves by exactly the difference between the two cards. Memory, disks and port speed upgrade under the same rule.

What happens when a graphics card dies?

It is swapped, on a clock the agreement puts in writing: four hours from the moment it is reported by telephone or live chat. Miss the window and service credit follows a published schedule rather than a goodwill request.

Is the bandwidth really unmetered? What does egress cost?

Really: nothing is metered in either direction, at any port speed we sell, and egress costs zero because no meter exists to charge it from. On GPU work this is routinely the biggest line elsewhere -- a training run pulls datasets in and pushes checkpoints out, an inference endpoint answers real traffic, and per-gigabyte egress bills every one of those bytes.

Monthly here versus a GPU by the hour -- where is the crossover?

Arithmetic settles it. Hourly wins while the work is bursty and genuinely stops -- up to a few hundred hours a month. Past roughly two thirds of the month it stops being close, before counting egress, persistent storage between runs, and queue time waiting on capacity. Monthly also means the same machine, disks and addresses next month.

Is mining allowed? Blockchain and decentralized-compute nodes?

Allowed, ordinary, and always have been -- part of this catalog was bought for exactly that duty. https://serverroom.net/mining covers the mining case specifically.

How is it paid, and what does the contract look like?

Card, PayPal, ACH, bank wire or cryptocurrency -- and there is no contract to sign: one month is the default term, three, six and twelve-month cycles take a discount, and renewal does not step the price up. On a larger first order a wire or a stablecoin usually clears fraud review faster than a card.

What if the card turns out to be the wrong one?

Within three calendar days of first activation, after troubleshooting is exhausted, the refund policy applies as written: https://serverroom.net/refund-policy-us. The sizing and bus-generation sections of the page exist precisely to make that clause unnecessary -- it is cheaper for both sides when the right card is bought first.

Do you carry the newest graphics card?

Sometimes -- and it is usually the wrong specification. The newest part is scarce, on allocation, priced as if every transistor were needed, and seated on platforms that cannot feed it the bus generation it was designed for. Most GPU work is serving a model that already fits, encoding video, rendering frames or driving desktops, and for those the right card is often two generations old at a quarter of the money. Name the job; if the answer genuinely is the newest part, we say so, with the cost and the date it can exist.

Add-in graphics cards on bare metal · Server Room · est. 2004

The card is one line on the invoice. The machine is the other.

Every catalog in this industry prices the card and the server as one figure that somebody else composed. This one keeps them apart: the machine carries its own monthly figure, the card carries its own, and the total is their sum. Set the two controls below — what runs on it, and how much card memory the job needs — and everything still standing is a complete server, priced by the same catalog the checkout bills from.

Narrow the shelves Whose card is it?

One account per physical machine. The card is handed to your operating system on PCIe passthrough — your kernel module, your driver version, no hypervisor in the path.

The bench

Two controls. The rest is arithmetic.

Nothing here reserves stock or opens an account. The figure on each row is the whole server with that card seated — processor, memory, mirrored disks, unmetered port, one IPv4 — and the card’s own share is printed underneath, so both halves of the money stay visible.

What runs on it?

How much card memory?

anygigabytes or more

5 GB7-8 billion parameters, 4-bit
8 GB13-14 billion parameters, 4-bit
19 GB30-34 billion parameters, 4-bit
38 GB70 billion parameters, 4-bit

Estimates from the order catalog, not a rate card, and they are the checkout’s own figures. A card whose manufacturer publishes no memory figure prints none here — and drops out as soon as the memory control leaves “any”, because we will not show a part clearing a bar we cannot verify; ask and we read the figure off the card itself. Several parts seat in more than one chassis at more than one price: the row shows the cheapest seat, the configurator lists the rest.

Counted as the page was served

57Cards on the shelves
$20.79Cheapest card line, per month
$91.72Cheapest complete build with a card
5Cities
0Bytes metered

Read from the order catalog at render time, never typed. Withdraw a card from the store tomorrow and this strip is smaller tomorrow. What is racked and powered this morning is a separate question with its own page: instant servers.

Tenancy, before anything else

One tenant. The whole card. Nothing scheduled above you.

Passed through, not carved up.

The card sits in a physical PCIe slot in a machine rented to one account, and the operating system that owns it is the one you installed. You load the kernel module, you pin the driver, and CUDA or ROCm stays at the release your code was tested against. There is no hypervisor layer, no MIG partition, no vGPU profile and no time-slicing scheduler — the latency you measure on the first afternoon is the latency you keep, at four in the afternoon and at four in the morning.

The chassis follows the same rule: every core, every DIMM, both disks, root, and the machine’s own iLO or iDRAC on a separate management network for console, virtual media and power. A driver experiment that takes the network down still leaves you the screen.

No sliceNo MIG partition, no vGPU profile, no fractional allocation. One card, one operating system.
No queueNo scheduler shares the silicon. When your process is idle, the card is idle — and still yours.
No layerBare metal. Your kernel, your modules, your driver version, pinned until you move them.
No meterDatasets in, checkpoints and renders out. Nothing on the port counts the bytes.

The pricing model

Two lines that add. Not a tier.

A GPU “instance” is somebody else’s decision about what a card buyer should pay for the processor, memory, disks and network around the card. Here those are one line, the card is another, and they add — which is why either end of the shelf above seats in the same chassis, and why swapping the card changes the invoice by exactly the difference between the two cards.

The machineIntel Xeon E5-2600 v1/v2Intel Xeon E5-2630L Hex Core 2.00 GHz · 32 GB DDR3 · 2 × SATA 500 GB (RAID 1) · 500 Mbps$70.93
The cardAny part on the shelves above$20.79 to $1,599, cheapest shelf to dearest$20.79
The portUnmetered, both directionsNo allowance, no egress line, at any speedincluded
SetupNothingNo contract; the default term is one month—

Complete machine, card fitted, per month, from$91.72

The same arithmetic at the other end of the list: H100 80GB in a complete machine is $1,682.67 a month — $2.31 an hour if the comparison is against hourly billing, and the traffic adds nothing.

The configurator rounds per-disk pricing and can land a cent or two off these figures. Longer billing cycles discount further — pricing has every figure in one place, and older hardware, lower price is the standing argument for why the bottom of the shelf exists.

Card memory

Fit is binary. Everything else is merely speed.

Cores and clocks set how fast a model runs. Card memory sets whether it runs: the weights are either resident or they are not, and a card two gigabytes short does not run slowly — it exits. Settle this number before any other, because it is the one specification almost nobody prints next to a price.

The arithmetic is parameters times bytes per parameter: two bytes each at 16-bit, one at 8-bit, half a byte at 4-bit. Quantizing to 4-bit is the lever that seats a model on a smaller shelf, and it trades a measurable amount of quality for the seat.

Budget head-room. The table is the weights and nothing else. The runtime, the activations and the key-value cache all live in the same memory, and the cache grows with every token of context — 38 GB of weights does not serve comfortably from a 40 GB card. Allow a quarter to a half again, or tell us the model and the context length you actually run and we will size it against that.

Weights alone. Leave head-room for the runtime and the context window before settling on a card.
Model	16-bit	8-bit	4-bit
7-8 billion parameters	16 GB	8 GB	5 GB
13-14 billion parameters	28 GB	14 GB	8 GB
30-34 billion parameters	68 GB	34 GB	19 GB
70 billion parameters	140 GB	70 GB	38 GB

The chassis underneath

Four platforms take a card, and the bus is what separates them.

Each hands the slot a different link generation and lane budget, and the column is printed because it changes results and nobody else publishes it: a customer once swapped one consumer generation for the next on a Gen 3 platform and rendered slower, and had worked out why before we had. The number belongs on the page.

Intel Xeon E5-2600 v1/v2

$70.93per month, before the card

Intel Xeon E5-2630L Hex Core 2.00 GHz · 32 GB DDR3 · 2 × SATA 500 GB (RAID 1) · 500 Mbps

Cards accepted: 28
Bus to the card: PCIe 3.0
Lanes: 40 lanes per socket
Sockets: 2
Memory ceiling: 256 GB

Open it in the configurator →

Intel Xeon E5-2600 v3/v4

$73.47per month, before the card

Intel Xeon E5-2620 v4 Octo Core 2.10 GHz · 32 GB DDR4 · 2 × SATA 500 GB (RAID 1) · 1 Gbps

Cards accepted: 44
Bus to the card: PCIe 3.0
Lanes: 40 lanes per socket
Sockets: 2
Memory ceiling: 1 TB

Open it in the configurator →

Intel Xeon Silver / Gold

$83.67per month, before the card

Intel Xeon Silver 4110 8 Core 2.10 GHz · 16 GB DDR4 · 2 × SATA-SSD 240 GB (RAID 1) · 1 Gbps

Cards accepted: 48
Bus to the card: PCIe 3.0
Lanes: 48 lanes per socket
Sockets: 2
Memory ceiling: 1 TB

Open it in the configurator →

AMD EPYC

$217.59per month, before the card

AMD EPYC 7413 24 CORE 2.65 GHz 128MB L3 CACHE · 32 GB DDR4 · 1 × SATA-SSD 240 GB · 1 Gbps

Cards accepted: 10
Bus to the card: PCIe 4.0
Lanes: 128 lanes per socket
Sockets: 2
Memory ceiling: 1 TB

Open it in the configurator →

When the link generation matters, measured by workload

A card built for a newer bus runs correctly on an older one; it simply cannot move data over the link at its design rate. Whether that costs you anything depends entirely on where the data sits.

It costs you when traffic crosses the bus continuously — a training loop streaming batches from disk or system memory, a renderer feeding scene data every frame, any working set that spills out of card memory.
It costs you nothing once the weights are resident and stay resident — inference on a model that fits, video transcode, most simulation. The link then carries requests and results, and those are small.
Name the workload if you are unsure. We will say plainly whether the platform bottlenecks it — including when the honest spec is a cheaper card on a faster link rather than a dearer card on a slower one.

In the figure already

What every machine ships with before you add a thing.

Root, and a driver stack that is yours

A clean operating system and the keys. You pick the driver release, the CUDA or ROCm version and the framework build, and you pin them — nothing upgrades underneath a deadline. Want the driver in place before handover? Name the version on the order and it will be.
Out-of-band console on its own network

iLO on the HPE chassis, iDRAC on the Dell: keyboard-and-screen console, virtual media, power control, sensor and hardware logs. A module that wedges the display or the NIC is a reboot you perform yourself, watching the POST — not a ticket.
A port nothing counts

Unmetered in both directions at the included speed and at every speed above it. No transfer allowance, no egress line on any invoice — datasets in, checkpoints out, a model served to real traffic, all at the price of the port. Unmetered bandwidth is the ladder upward, and it is the largest line this page deletes from a hyperscaler bill.
Reinstalls without a ticket

Operating system images in one click, as often as you like, at no charge — a driver experiment that goes wrong costs twenty minutes. Reverse DNS is yours to edit, IPv6 is included, and additional IPv4 is a catalog line rather than a negotiation.
Engineers on shift, around the clock

Telephone, live chat and tickets on every machine, and a failed component is replaced within four hours of the report by phone or chat. Support publishes the response times and the credit schedule in writing.
Five rooms on two continents

New York, Miami, San Francisco, Amsterdam and Bucharest. Where the machine stands sets your latency to whatever calls it and the data-protection regime it answers to; data centers describes each building.

Lead time

Half an hour to three weeks, set by where the card is standing today.

One number would be a lie in one direction or the other, because it has never been one number: a card already seated in a racked machine is a handover, a card that must be bought is a purchase order. What we commit to is telling you which rung you are on before you pay.

~30 minutes
The card is already seated in a racked machine
Nothing is fitted and nothing moves. The machine is imaged and handed over. Instant servers lists what is standing there right now.
4–24 hours
A card from our own stock is fitted
The ordinary case for this page. The part is in the store room; a technician seats it, builds the array and installs the operating system. A racked machine that is close but not exact — memory added, a drive swapped, the array rebuilt to your layout — is the same visit, done in place.
5–10 working days
The card is bought in
A part we do not hold is ordered, then the build proceeds as above. Build to order is the page when the whole specification is yours.
10–15 working days
The card is on allocation
Current-generation accelerators ship on the distributor’s date, not ours, and we name that long pole before you commit rather than after. Hardware bought in for one customer is quoted with three months up front, because it is a machine we cannot re-let if the account closes in week two.

Every rung starts when payment and fraud review clear — the one step we will not put a promise on: most orders clear within a couple of hours, a first order from a new account can take longer. Card, PayPal, ACH and wire are accepted, and so is cryptocurrency, which on a larger first order usually clears fastest of all.

Asked before, answered here

The twelve questions the page exists to settle.

Collected from chat and tickets, in roughly the order buyers ask them. The first has cost real orders, which is why it opens the list.

Do I share the GPU with anyone?: No. One physical machine, one account, and the card passed through to the operating system you installed — no hypervisor, no MIG partition, no vGPU profile, no time-slicing scheduler, no neighbour’s batch job on your silicon. The numbers you benchmark on day one are the numbers you keep.
Will two cards fit in one machine?: Often, and it is quoted rather than ticked: chassis clearance, the card’s physical width, the lane budget and the power supply set the count, so an honest answer needs the specific part. Name the card, and the reply comes back with the chassis, the count and the lead time. The usual motive is memory rather than speed — two 48 GB cards hold a model one 80 GB card cannot — and AI dedicated servers is the adjacent page when the second card starts looking like a second machine.
Which driver and CUDA release ships on it?: The ones you install. The machine arrives with a clean operating system and root; you pin the driver, the CUDA or ROCm release and the framework build to what your code was validated against, and nothing moves them afterward. If you want the driver in place before handover, put the exact version on the order. If a particular card must support a particular CUDA release, ask before paying and we check it rather than guess.
The card I want is not on a shelf today — then what?: Then it is fitted from stock or bought in, and you are told which before you pay: stocked parts seat within four to twenty-four hours, bought-in parts run five to ten working days, and allocation-bound accelerators ten to fifteen. This page deliberately prints no stock counts — a card is a part we fit, not a product that comes and goes — but when the requirement is a machine today, instant servers lists what is racked and powered right now.
Can the card change after deployment?: Yes, in the machine you already hold: no migration, no reinstall, no new addresses. It is a scheduled rack visit with a short outage, and the invoice moves by exactly the difference between the two cards. Memory, disks and port speed upgrade under the same rule.
What happens when a card dies?: It is swapped, on a clock the agreement puts in writing: four hours from the moment it is reported by telephone or live chat. Miss the window and service credit follows a published schedule rather than a goodwill request. Support carries the schedule and the phone number.
Is the bandwidth really unmetered? What does egress cost?: Really: nothing is metered in either direction, at any port speed we sell, and egress costs zero because no meter exists to charge it from. On GPU work this is routinely the biggest line elsewhere — a training run pulls datasets in and pushes checkpoints out, an inference endpoint answers real traffic, and per-gigabyte egress bills every one of those bytes. Unmetered bandwidth is the ladder above the included speed.
Monthly here versus a GPU by the hour — where is the crossover?: Arithmetic settles it. Hourly wins while the work is bursty and genuinely stops — up to a few hundred hours a month. Past roughly two thirds of the month it stops being close, before counting egress, persistent storage between runs, and queue time waiting on capacity. Monthly also means the same machine, disks and addresses next month, which production workloads care about more than benchmarks do.
Is mining allowed? Blockchain and decentralized-compute nodes?: Allowed, ordinary, and always have been — part of this catalog was bought for exactly that duty. Mining servers covers the mining case specifically, including the parts that only make sense for it.
How is it paid, and what does the contract look like?: Card, PayPal, ACH, bank wire or cryptocurrency — and there is no contract to sign: one month is the default term, three, six and twelve-month cycles take a discount, and renewal does not step the price up. On a larger first order a wire or a stablecoin usually clears fraud review faster than a card, which is worth knowing before a build waits on it.
What if the card turns out to be the wrong one?: Within three calendar days of first activation, after troubleshooting is exhausted, the refund policy applies as written. The sizing table and the bus-generation column above exist precisely to make that clause unnecessary — we would rather argue you down to a smaller card than process the return on a bigger one.
Do you carry the newest card?: Sometimes — and it is usually the wrong specification. The newest part is scarce, on allocation, priced as if every transistor were needed, and seated on platforms that cannot feed it the bus generation it was designed for. Most GPU work here is not frontier training: it is serving a model that already fits, encoding video, rendering frames, driving desktops — jobs where the right card is two generations old at a quarter of the money. Name the job; if the answer genuinely is the newest part, we say so, with the cost and the date it can exist.

Adjacent pages

Everything one step from this page.

This page owns one question — which cards we fit and what each costs. The questions on either side of it each have their own page, and this is the full set.

Still weighing two shelves? — whoever answers has sized one of these before, and will say when the cheaper card is the right specification.

For the record

The whole page, reduced to figures.

Read from the order catalog at the moment this page was served. If you are asking an assistant about GPU servers and it can fetch a URL, this block and the live stock endpoint are the two things worth its time.

Cards we fit: 57 add-in cards on 3 shelves: 15 training and inference accelerators, 23 workstation and video parts, 19 consumer parts.
The card, as a line: $20.79 to $1,599 a month, added to the machine it seats in. The machine’s own figure never moves with the card.
A complete machine with a card in it: $91.72 to $1,682.67 a month — processor, memory, mirrored disks, unmetered port, one IPv4 and the out-of-band console all in the figure. Nothing left to add.
The machine alone: $70.93 to $217.59 a month before any card is fitted.
Cheapest build that can train: Instinct MI50 16GB, $169.93 a month complete.
Dearest: H100 80GB, $1,682.67 a month complete — $2.31 an hour for the whole machine if you are comparing against hourly billing, with no egress charge arriving later.
Tenancy: One account per physical machine, and the card is passed through to that account’s own operating system. No hypervisor in the path, no MIG partition, no vGPU profile, no time-slicing.
Bandwidth: Unmetered in both directions at every port speed we sell. No transfer allowance and no egress line on any invoice — datasets, checkpoints and rendered output move for the price of the port.
Where: 5 cities: New York, Miami, San Francisco, Amsterdam, Bucharest.
Term: Monthly, no contract. Three, six and twelve-month cycles discount up to 15%.
Multiple cards: Possible, quoted rather than ticked: chassis width, lane budget and power decide the count, so the answer names the specific card. Ask and the reply names the chassis and the lead time.
Payment: Card, PayPal, ACH, wire and cryptocurrency.

Two controls. The rest is arithmetic.

One tenant. The whole card. Nothing scheduled above you.

Two lines that add. Not a tier.

Fit is binary. Everything else is merely speed.

Four platforms take a card, and the bus is what separates them.

Intel Xeon E5-2600 v1/v2

Intel Xeon E5-2600 v3/v4

Intel Xeon Silver / Gold

AMD EPYC

When the link generation matters, measured by workload

What every machine ships with before you add a thing.

Root, and a driver stack that is yours

Out-of-band console on its own network

A port nothing counts

Reinstalls without a ticket

Engineers on shift, around the clock

Five rooms on two continents

Half an hour to three weeks, set by where the card is standing today.

The card is already seated in a racked machine

A card from our own stock is fitted

The card is bought in

The card is on allocation

The twelve questions the page exists to settle.

Everything one step from this page.

The whole page, reduced to figures.