The largest HBM per GPU of its generation
256 GB of HBM3E per GPU, 2 TB across an 8-GPU node, holds a 70-billion-parameter model in FP16 on one accelerator and serves 405-billion-parameter models on a single node without spilling to host memory.
// coming soon: cdna 3
AMD Instinct MI325X GPU Server
CDNA 3 bare metal for memory-bound inference and large-model training: 256 GB HBM3E at 6 TB/s per GPU, 1,307 TFLOPS FP16, 256 MB Infinity Cache, and Infinity Fabric linking 8 OAM GPUs per node on the open ROCm stack. Reserve capacity now.
- Coming soon, reservations open
- 256 GB HBM3E, up to 8 OAM GPUs per node
- 20+ Tier III global data centers
- Crypto payments & 24/7 support included
New HardwareComing Soon
Request Sent
Thanks. An engineer will spec your MI325X build and reach out shortly.
Something went wrong. Please try again or email sales@redswitches.com.
// reserve capacity
Reserve AMD Instinct MI325X Capacity
MI325X servers are landing in our data centers soon. Leave your details and our engineers will reserve a configuration for you, confirm lead times, and quote single-GPU, 4x and 8x OAM builds with volume and committed-term discounts.
Request Sent
Thanks for reaching out. An engineer will spec your build and get back to you shortly.
Something went wrong. Please try again or email sales@redswitches.com.
// why redswitches for mi325x
What You Get With Every AMD Instinct MI325X Server
Every RedSwitches AMD Instinct MI325X server is single-tenant bare metal with full root and IPMI access, unmetered 1, 10 or 25 Gbps bandwidth, no setup fee, no bandwidth overage and a flat monthly price. Stocked builds are online in about 1 hour, larger configurations up to 8 GPUs per node are built to order, volume and 6/12-month committed-term discounts apply, payments include crypto, and engineers answer 24/7.
- Free
Reserve at No Cost
Reservations for AMD Instinct MI325X are free and non-binding. An engineer confirms the lead time, holds the hardware for you, and sends a quote before you commit to anything.
- $0
Setup Fee, Flat Monthly
No setup fee and a flat monthly price for the whole server. No per-hour meter and no surprise line items, so GPU spend is forecastable.
- Unmetered
Bandwidth, No Egress Bills
Unmetered 1, 10 or 25 Gbps uplinks are included, and whatever port speed you choose you can use all of it: no overage charges and no egress bills, ever. Moving datasets, checkpoints and model weights in and out costs nothing extra.
- 1 tenant
Bare Metal, Full Control
Single-tenant hardware with root and IPMI access. You choose the OS, drivers, CUDA or ROCm version, and the NVLink or MIG layout, and a private VLAN can link your RedSwitches servers.
- Up to 8x
Multi-GPU, Built to Order
Single and multi-GPU nodes, up to 8 GPUs per server with NVLink where the card supports it. Volume discounts on multi-GPU and multi-server orders, plus 6 and 12-month term savings.
- 24/7
Engineers, Not Bots
Live chat, Telegram and email answered by engineers around the clock. Pay by card, PayPal, bank wire or crypto with no KYC, in 20+ Tier III data centers across the EU, US and Asia.
// compared with cloud
RedSwitches MI325X Server vs a Typical Cloud GPU Instance
Same accelerator, different economics: what changes when the GPU sits in a dedicated server you control instead of a metered instance.
| RedSwitches MI325X server | Typical cloud GPU instance | |
|---|---|---|
| Billing | Flat monthly price per server | Per-hour or per-second metering |
| Bandwidth | Unmetered 1/10/25 Gbps, use the full port, no overage or egress fees | Egress billed per GB |
| Private networking | Private VLAN between your servers on request | Paid VPC and peering constructs |
| Setup fee | $0 | Varies by instance and region |
| Tenancy | Single-tenant bare metal | Shared, virtualised hosts |
| Access | Root and IPMI, your OS and drivers | Hypervisor-managed images |
| Multi-GPU | Up to 8 per node, NVLink where supported, built to order | Fixed instance shapes |
| Payments | Card, PayPal, bank wire, crypto | Card or invoice |
| Support | 24/7 engineers on chat, Telegram and email | Ticket tiers, paid support plans |
AMD Instinct MI325X Key Specifications
CDNA 3 silicon with 256 GB HBM3E, 6 TB/s bandwidth, 1,307 TFLOPS FP16, and Infinity Fabric linking 8 OAM GPUs per node.
- Architecture
- AMD CDNA 3, 304 compute units, 19,456 stream processors, 1,216 matrix cores; successor to Instinct MI300X
- Memory
- 256 GB HBM3E, 6 TB/s bandwidth per GPU, 256 MB AMD Infinity Cache
- FP8
- 2,614.9 TFLOPS (5,229.8 TFLOPS with structured sparsity)
- FP16/BF16
- 1,307.4 TFLOPS (2,614.9 TFLOPS with structured sparsity)
- INT8
- 2,614.9 TOPS (5,229.8 TOPS with structured sparsity)
- FP32 vector
- 163.4 TFLOPS
- FP64 vector / FP64 matrix
- 81.7 / 163.4 TFLOPS
- Form Factor & TBP
- OAM module in the AMD Instinct MI325X platform, 1,000 W peak board power
- Infinity Fabric
- AMD Infinity Fabric links for all-to-all connectivity across 8 GPUs
- PCIe
- Gen 5 x16 to the host
- Platform
- 8x MI325X per node, 2 TB of HBM3E per node
- Software
- ROCm 6.2+, PyTorch, TensorFlow, JAX, vLLM, SGLang, Triton, Hugging Face; hipify for CUDA porting
Computeper GPU, peak
Interconnects
// why mi325x
Why Choose MI325X
The largest HBM capacity of its generation: 256 GB per GPU at 6 TB/s, CDNA 3 matrix cores, and the open ROCm stack for inference and training.
6 TB/s for memory-bound inference
Token generation is limited by how fast weights and KV cache can be read; 6 TB/s per GPU, 13% more than MI300X and 25% more than H200, lifts tokens per second on long-context and large-batch serving.
1,307 TFLOPS FP16 and 2,615 TFLOPS FP8
CDNA 3 matrix cores deliver 1,307.4 TFLOPS dense FP16/BF16 and 2,614.9 TFLOPS dense FP8 per GPU, doubling with structured sparsity, for training and fine-tuning as well as inference.
Infinity Fabric 8-GPU platform
Infinity Fabric links join 8 OAM GPUs all-to-all in the MI325X platform, so tensor-parallel serving and data-parallel training scale across the node with 2 TB of HBM3E behind them.
Open ROCm software stack
ROCm 6.2+ runs PyTorch, TensorFlow, JAX, vLLM, SGLang, Triton and Hugging Face natively; hipify ports existing CUDA code, so you are not locked into a single vendor toolchain.
Bare metal, not shared cloud
Single-tenant servers with full root and IPMI access, NVMe storage, unmetered 1/10/25 Gbps uplinks, and no noisy neighbours, with the GPU topology you control.
// use cases
Ideal Use Cases
From memory-bound LLM serving and MoE models to ROCm training and FP64 HPC, where 8x MI325X nodes pay for themselves.
Large-Model LLM Inference
Serve Llama, DeepSeek, Mixtral and proprietary models with vLLM or SGLang; 256 GB per GPU keeps weights and KV cache resident for more concurrent requests per accelerator.
Long-Context and Large-Batch Serving
128K-token contexts and high-batch endpoints are memory-bound; 6 TB/s of HBM3E bandwidth turns directly into higher tokens per second per GPU.
Mixture-of-Experts Models
MoE models load every expert's weights even when only a few are active per token; 2 TB of HBM3E per 8-GPU node hosts large MoE checkpoints on fewer nodes.
Training and Fine-Tuning with ROCm
Pre-train mid-size models and fine-tune large ones in BF16 or FP8 with PyTorch on ROCm, scaling across 8 Infinity Fabric-linked GPUs.
FP64 HPC and Scientific Computing
163.4 TFLOPS FP64 matrix and 81.7 TFLOPS FP64 vector per GPU for CFD, molecular dynamics, climate and seismic simulation that cannot use low precision.
Private AI and Cloud Repatriation
Replace metered cloud GPU instances with dedicated MI325X nodes: full root access, unmetered bandwidth, no setup fee, and committed-term discounts.
Trusted by Enterprise Teams Worldwide
// customer reviews
Trusted in Production
Real reviews from Google, HostAdvice, and Cryptwerk.
- 25Gbps Bare Metal
Port speeds on cloud were capped and shady. With RS I actually get the 25Gbps they say. No throttle bs.
- Validators
Set up multiple Solana + Avalanche validators through them. Hardware was clean, latency was low (especially in Europe), and uptime’s been 100% so far.
- Full Root Access
As a sysadmin, I care more about control than flashy dashboards. RedSwitches gives me root access, IPMI, and actual hardware specs I can configure. Good for serious users.
- Cloud Exit
RedSwitches gave us full control over our infrastructure without the vendor lock-in we kept running into with cloud hosts. Customizable builds, fast provisioning, and actual humans handling support tickets. It’s refreshing.
- 99.99% Uptime
My website works smoothly, thanks to their 99.99% uptime guarantee. The support team is another plus for me, as I always have been able to get help whenever I needed it in just 5 minutes on average.
- ETH & BTC Nodes
Deployed a few Ethereum and Bitcoin nodes here. Uptime’s been flawless, and sync speed was great thanks to their storage config.
- Game Servers
Been using RedSwitches for 6 months now for my small game server biz. Uptime has been great, and I haven’t run into any hidden charges.
- AI Storage
Using the storage servers to archive logs and snapshots from our AI pipeline… I also love that I could pick the datacenter closest to our team.
- Support
Support is super responsive. I had an issue with an OS reinstall and they jumped in within 10 minutes… Transparent pricing = win.
- Instant Delivery
They have an instant delivery section… they delivered it within 120 mins with all my requirements fulfilled (OS/RAID/Software configured etc).
- Global Locations
With servers available in numerous strategic locations, RedSwitches offers exceptional versatility and performance for our company’s diverse hosting needs. Plus, their no setup fee policy really helps keep costs down.
- Bare Metal Cloud
The dedicated server I got from RedSwitches has been incredibly reliable and fast. Their bare metal cloud solutions offer excellent performance, and the cloud VPS options are perfect for scaling. Highly recommended!
- Easy Onboarding
Absolutely delighted with RedSwitches! The setup was quick and free, and the fact that they accept all major payment gateways made the process seamless.
- Bare Metal
Very good experience using their bare metal servers. Their customer service is one of the finest I have experienced - always prompt at resolving troubles. Highly recommend.
// faqs
Deep Dive & FAQs
Availability and reservations, MI325X vs MI300X and H200, 8x Infinity Fabric nodes, pricing, ROCm compatibility, power, and drivers.
When will MI325X servers be available and how does reservation work?
MI325X nodes are landing in our data centers soon. Submit the reservation form with your preferred configuration and region; our engineers confirm the lead time for that build, hold the hardware for you, and send a quote. Reservations are free and non-binding until you approve the quote.
MI325X vs MI300X: when does the extra memory matter?
MI325X keeps the same CDNA 3 compute as MI300X (1,307.4 TFLOPS FP16, 2,614.9 TFLOPS FP8) but moves from 192 GB HBM3 at 5.3 TB/s to 256 GB HBM3E at 6 TB/s, with peak board power rising from 750 W to 1,000 W. The difference shows when a model or its KV cache does not fit: a 70-billion-parameter model in FP16 with long contexts, large-batch serving, or MoE checkpoints that spill across GPUs on MI300X can stay resident on MI325X. Compute-bound training at short sequence lengths gains little, so MI300X remains the value option there.
MI325X vs NVIDIA H200: how do they compare?
MI325X carries 256 GB HBM3E at 6 TB/s per GPU against H200's 141 GB HBM3e at 4.8 TB/s, so one MI325X holds 1.8x the model weights and KV cache. H200 lists 1,979 TFLOPS FP16 with sparsity against 2,614.9 TFLOPS sparse FP16 on MI325X, and draws 700 W against 1,000 W. H200 runs the CUDA ecosystem; MI325X runs ROCm with PyTorch, vLLM and SGLang support, so the choice usually comes down to software stack and memory per GPU.
How is MI325X priced and are there discounts?
Pricing is quoted per configuration because GPU count, host CPU, memory, storage and uplink choices vary. Volume discounts apply to multi-GPU and multi-server orders, and every GPU server qualifies for committed-term discounts on 6 and 12-month billing. There is no setup fee, and crypto, card, PayPal and bank wire are accepted.
Does MI325X run my PyTorch or CUDA code?
PyTorch, TensorFlow and JAX ship ROCm builds, and vLLM, SGLang, Triton and Hugging Face libraries run on MI325X through ROCm 6.2 and later, so most framework-level code runs without changes. Custom CUDA kernels are ported with hipify, which translates CUDA source to HIP; in most projects the remaining work is validating performance rather than rewriting logic.
Can I get an 8x MI325X node with Infinity Fabric?
Yes. MI325X ships as an OAM module in the AMD Instinct MI325X platform, where Infinity Fabric links connect 8 GPUs all-to-all with 2 TB of HBM3E per node, so 8x builds are the native form factor. We also quote single-GPU and 4x configurations. State your target topology in the reservation form and we will size the host CPUs, RAM, NVMe and uplinks around it.
What about power and cooling for 1,000 W GPUs?
Our GPU racks are provisioned for high-density, high-power accelerators; an 8x MI325X node at 1,000 W peak board power per GPU is deployed with the power feeds and cooling it needs. You do not need to plan facilities, only the workload.
Which operating systems and drivers do you install?
Ubuntu, Debian, Rocky/AlmaLinux or your own ISO, with ROCm 6.2+ and a container runtime pre-installed on request, so the node is ready for your containers on day one.
// need help
Not Sure Exactly What You Need
No problem. Our talented engineers will consult, architect, migrate, manage, and do whatever it takes to help your business grow and succeed.