Q: Which GPUs have a hardware video encoder?

A common misconception: the A100, H100 and H200 have no NVENC encoder. They only carry decoders, so FFmpeg transcoding falls back to the CPU. Hardware encoding is available on the RTX 4000 Ada, the RTX PRO 6000 Blackwell and the L4 Ada, all of which encode H.264, HEVC and AV1 in hardware. The Tesla T4 and A10 have NVENC but no AV1. If video transcoding is part of your workload, talk to us first so the card matches it.

Question 1

Which GPU is right for my AI project?

Accepted Answer

That depends on the workload: For inference and rendering, we recommend the RTX 4000 Ada (20 GB GDDR6, 130 W, efficient full-height card). For LLM training and multi-GPU setups, the RTX PRO 6000 Blackwell (96 GB GDDR7) is the better choice. The RTX PRO 6000 Blackwell Server Edition with 600 W is available on request in a Supermicro chassis. We analyze your workload free of charge.

Question 2

What is the difference between RTX PRO 6000 Max-Q and Server Edition?

Accepted Answer

Both variants use the same Blackwell chip with 96 GB GDDR7 and identical memory bandwidth — they differ in power draw and cooling. The Max-Q runs at 300 W with its own blower fan and fits our Dell PowerEdge R760xa, which takes up to four cards. The Server Edition draws 600 W and is cooled passively by chassis airflow, so it requires a chassis designed for it — we offer it in our Supermicro GPU server with up to four cards, two of which are available immediately. Peak compute of the Max-Q is roughly 10 to 15 percent lower. For bandwidth-limited LLM inference the practical difference is smaller, because memory rather than compute is the limit there.

Question 3

Are multi-GPU setups possible?

Accepted Answer

Yes, we implement multi-GPU configurations for demanding AI and ML workloads. The GPUs can also be connected to the INGATE Cloud via Direct Connect for hybrid workloads.

Question 4

What software is pre-installed?

Accepted Answer

Upon request, we pre-install CUDA, cuDNN, TensorRT, and container runtimes for GPU workloads. We support all major ML frameworks including PyTorch, TensorFlow, and JAX.

Question 5

Does my training data stay in Germany?

Accepted Answer

Yes, your data and models remain on sovereign German infrastructure. As an owner-managed German GmbH, we are not subject to the US Cloud Act.

Question 6

Which operating systems are supported?

Accepted Answer

The common choices are Ubuntu 24.04 LTS and 22.04 LTS, Debian 13, Rocky Linux 9 and AlmaLinux 9, each with matching NVIDIA drivers. Windows Server is available against a licence fee, as is installing your own ISO through the iDRAC virtual media function. Other distributions we pre-install at no extra cost.

Question 7

Is a different hardware configuration possible?

Accepted Answer

Yes, we can configure any system to your specifications — including GPU configuration, storage, and networking. Our hardware consulting is included.

Question 8

What is inference and how does it differ from training?

Accepted Answer

Inference refers to using an already trained AI model to make predictions or decisions in real time — such as image classification, speech recognition, or chatbots. Training, on the other hand, is the upstream process where the model learns from large datasets and optimizes its parameters. Training requires massive computing power and is typically performed on GPUs like the NVIDIA H100 or A100, which feature Tensor Cores and high VRAM bandwidth (HBM3). Inference can often be run on more efficient GPUs like the NVIDIA L4, T4, or RTX series, as throughput and energy efficiency matter more than raw computing power.

Question 9

What are the differences between GPUs for rendering and AI workloads?

Accepted Answer

Rendering GPUs like the NVIDIA RTX PRO 6000 are optimized for visualization, CAD, VFX, and simulation. Their focus is on RT Cores (raytracing) and high VRAM for large scenes. AI GPUs like the NVIDIA H100 or A100 prioritize Tensor Cores for matrix operations and offer higher memory bandwidth through HBM3 technology — critical for training large models. Some GPUs like the NVIDIA L4 or the RTX series can handle both tasks well, making them ideal for companies that need both rendering and inference. INGATE advises you individually on which GPU configuration optimally suits your workload.

Question 10

Which GPUs have a hardware video encoder?

Accepted Answer

A common misconception: the A100, H100 and H200 have no NVENC encoder. They only carry decoders, so FFmpeg transcoding falls back to the CPU. Hardware encoding is available on the RTX 4000 Ada, the RTX PRO 6000 Blackwell and the L4 Ada, all of which encode H.264, HEVC and AV1 in hardware. The Tesla T4 and A10 have NVENC but no AV1. If video transcoding is part of your workload, talk to us first so the card matches it.

Question 11

Are continuous workloads allowed?

Accepted Answer

Yes. You rent dedicated hardware, with no fair-use limit, no throttling and no interruption as with spot instances. Training that runs for weeks works exactly like continuous inference in production. There are no egress fees of the kind hyperscalers charge.

Question 12

What is included in the monthly price?

Accepted Answer

Rack space in the data centre, power, cooling, network connectivity at 1 Gbit/s full duplex with 25 TB of transfer volume, plus spare parts and hardware replacement in case of failure. You carry neither procurement cost nor residual value risk.

Question 13

Can I move to more powerful hardware later?

Accepted Answer

Yes, at any point during the term, as far as it is technically feasible. Additional cards or more memory can usually be added while the server keeps running. A move to a different server model is set up in parallel so you can migrate without time pressure.

Question 14

How long does provisioning take?

Accepted Answer

Usually five to ten working days after we receive your order. That figure is deliberately non-binding because GPU supply fluctuates — we give you a firm date with the order confirmation. Cards we hold in stock are considerably faster. For models with longer procurement times, such as the H200 NVL, plan for several weeks.

Question 15

Do the cards support NVLink?

Accepted Answer

The RTX PRO 6000 Blackwell has no NVLink, the cards couple over PCIe 5.0 x16, which in practice gives us around 55 GB/s between two cards. The H100 NVL and H200 NVL have real NVLink at roughly 600 GB/s, about ten times as much. If each card holds its own model this makes no difference. If you split one model across several cards using tensor parallelism, for instance with vLLM or SGLang, the faster interconnect means noticeably more throughput and lower latency — the larger the model, the more it matters.

Question 16

How much system memory does a GPU server need?

Accepted Answer

As a rule of thumb, system memory should at least match the combined video memory. A host with two 96 GB cards should therefore not run on 128 GB of RAM — when loading, quantising or converting models the weights sit in system memory for a while, and it becomes the bottleneck. For pure inference with models that stay resident in VRAM you can get by with less. We will tell you plainly when saving money at this point makes no technical sense.

Question 17

Do you offer the B200 or B300 as a PCIe card?

Accepted Answer

No, and that is not down to us: with Blackwell, NVIDIA dropped the PCIe card format for its large data centre GPUs. The B200 and B300 exist only as SXM modules mounted on an HGX board carrying eight GPUs. A single card cannot be separated out. If you need that performance class we build it on an HGX platform with eight GPUs per server. As a PCIe card, the RTX PRO 6000 Blackwell is the strongest option.

Question 18

Are your staff bound to confidentiality?

Accepted Answer

Yes. Every member of staff with potential access to customer systems is expressly bound to confidentiality and instructed on the criminal consequences, as are any subcontractors we use. With a dedicated server we have no technical access to the running system: physically exclusive hardware with no virtualisation layer, and privileged access is yours alone.

Question 19

We are bound by professional secrecy — do you offer an agreement under § 43e BRAO and § 203 StGB?

Accepted Answer

Yes, and we sign it as standard. The agreement follows the Bitkom template and contains the duty of silence without time limit, the instruction on criminal liability under § 203 (4) of the German Criminal Code, and the obligation of every person we bring in. It is a standard annex to the data processing agreement and is signed before provisioning.

Question 20

Does CUDA peer-to-peer work between the cards?

Accepted Answer

On a dedicated bare metal system, direct peer-to-peer access is possible with the right BIOS and driver configuration. We cannot guarantee it as a standard feature, and it is not available in a virtualised environment. Tell us in advance which framework you use and whether you split one model or load one per card, and we will check it for your case.

Question 21

Can I reach the BIOS through KVM over IP?

Accepted Answer

Yes. iDRAC access includes the console with BIOS and UEFI, so you can change PCIe and IOMMU settings yourself. The virtual media function also lets you mount your own ISO. On request we restrict iDRAC access to a fixed IP address, which makes it unreachable from outside.

Question 22

Do you handle backups?

Accepted Answer

Not on a dedicated GPU server. You are responsible for the operating system and therefore for backup and restore as well. We provide backups with managed servers, where we take over administration. Tell us beforehand if you need it and we will design the environment accordingly.

Question 23

Is INGATE certified?

Accepted Answer

Yes. INGATE is ISO/IEC 27001:2022 certified by British Assessment Bureau, a UKAS-accredited certification body. INGATE is also certified to ISO/IEC 27017:2015 for information security in cloud services. The ISMS covers all INGATE products and services. Certificate numbers and verification are on our About page.

Model	GPU RAM	Tensor TFLOPS	Tensor Cores	Use Case
NVIDIA H100 NVL	94 GB HBM3	~3,958	640 (FP8)	LLM Training, HPC
NVIDIA RTX PRO 6000 Blackwell Server Edition	96 GB GDDR7	~4,000	752 (FP4)	LLM Training, Inference, Multi-GPU
NVIDIA RTX PRO 6000 Blackwell	96 GB GDDR7	~3,511	752 (FP4)	LLM Training, Multi-GPU
NVIDIA L40s	48 GB GDDR6	~1,466	568 (FP8)	Training, Inference, VDI
NVIDIA RTX 6000 Ada	48 GB GDDR6	~1,322	568 (FP8)	Rendering, Training
NVIDIA L4 Ada	24 GB GDDR6	~485	240 (FP8)	Inference, Video, VDI
NVIDIA RTX 4000 Ada	20 GB GDDR6	~307	192	Inference, Rendering
NVIDIA A2	16 GB GDDR6	~36	128	Inference (Entry)

GPU Server

Supermicro · RTX PRO 6000 Server Edition

Supermicro · RTX PRO 6000 Server Edition

Dell R750xa · RTX 4000 Ada

Dell R7615 · RTX PRO 6000 Blackwell

Dell R760xa · RTX PRO 6000 Blackwell

Dell R7725 · H100 NVL & L40S

Custom GPU Infrastructure

Dell PowerEdge AI Server — All Models Available

Your Advantages at a Glance

GPU Comparison

More Services

Personal Hardware Consulting

CUDA Support & Data Sovereignty

Scalable GPU Clusters

Network & Connectivity

INGATE Premium Support

Managed Option

Technical Highlights

Redundant Power Supply

High-Efficiency Cooling

Fire Protection

High-Speed Backbone

Physical Security

Sustainability

Certified Security

INGATE

Data Center EMC Home of Data, Munich

Frequently Asked Questions

Technology Partners & Memberships