AMD EPYC 9455P48 cores · 128 GB DDR5 ECC · 2 × 1.92 TB NVMe
48 Zen 5 cores with full-width AVX-512 and 12 DDR5 channels: the value choice for CPU inference and CI.
AI & CPU compute
CPU servers for AI & CI. Serious compute, no GPU queue.
Not every AI or compute job needs a GPU. Quantised language models, embeddings, CI pipelines, compilation and rendering run well on many modern cores with wide vector units — at a fraction of the price, with no waiting list.
- Up to 64cores per socket
- AVX-512full width on Zen 5
- AMXon Xeon 6
- $255from, per month
What this workload needsWorkload profile
- Wide vector unitsZen 5 executes AVX-512 at full 512-bit width, and Xeon 6 adds AMX matrix engines for BF16 and INT8 inference.
- Memory bandwidthToken generation is bound by memory bandwidth. EPYC Turin feeds 12 DDR5 channels per socket.
- Many parallel jobsCI runners, compile farms and render nodes scale almost linearly with cores. 48 or 64 of them change a build queue.
- Predictable costA flat monthly price instead of per-minute billing: run the queue flat out, day and night.
Recommended
Best dedicated servers for AI & CPU compute
Chosen by the engineers who size these workloads every week. Every one of them is in stock.
Intel Xeon 6761P64 cores · 256 GB DDR5 ECC · 2 × 1.92 TB NVMe
64 Granite Rapids cores with AMX, the best CPU path for BF16/INT8 inference frameworks that support it.
AMD EPYC 9555P64 cores · 256 GB DDR5 ECC · 2 × 3.84 TB NVMe
64 Zen 5 cores and 256 GB for the biggest render and build queues.
Advice from our engineers
How to set up a server for AI & CPU compute
The first hour on a new server decides how the next year goes.
Where to host it
Compute is location-agnostic: pick the cheapest stock near your data. Frankfurt and Amsterdam carry the largest EPYC and Xeon inventory.
Compare latency from your users- Use quantised models (Q4–Q8) with llama.cpp or ONNX Runtime; match threads to physical cores, not threads.
- On Xeon 6, enable AMX in frameworks that support it (PyTorch with oneDNN, OpenVINO).
- Give CI runners their own tmpfs for build directories to spare the NVMe.
- Be honest about scale: training large models still belongs on GPUs — we will tell you when that is the case.
FAQ
Questions about AI & CPU compute
Still have a question?
Customers ask the engineers who run the servers, in a ticket from the client area — median first reply under 15 minutes, 24/7.
How support worksDo you offer GPU servers?
Not at the moment — our range is CPU-focused. For many inference and batch workloads, 48 to 64 modern cores are the better-value option.
How big a model can I run on CPU?
With 256 GB of memory, quantised models of several tens of billions of parameters fit comfortably; generation speed then depends on memory bandwidth.
Can I run CI for many repositories?
Yes. Runners in containers or VMs, each with CPU limits, keep dozens of pipelines busy on one server.
Other workloads
All workloads