AMD Instinct GPU servers

Deploy the AMD Instinct MI210 on bare metal infrastructure

The AMD Instinct MI210 accelerator, with 64 GB of HBM2e for machine learning and high-performance computing, in a bare metal server on an Intel Xeon Scalable or AMD EPYC platform. Sizing hardware for a language model is a different question: start from LLM hosting instead.

Continue to configuration

64 GB HBM2e with 1,638 GB/s memory bandwidth. 2nd Gen CDNA™ architecture on a PCIe 4.0 card. Three server lines: two on Intel Xeon Scalable, one on AMD EPYC.

AMD Instinct MI210 bare metal servers

Engineered for machine learning and high-performance computing. The MI210 pairs 2nd Gen CDNA™ compute architecture with 64 GB of HBM2e memory for intensive training, inference, and HPC applications.

Maximum compute throughput

2nd Gen CDNA™ architecture delivers breakthrough compute performance for deep learning training and scientific computing workloads requiring maximum throughput.

Standard PCIe form factor

The MI210 brings MI200-series silicon to a standard PCIe 4.0 card, so it seats in an ordinary Xeon Scalable or EPYC server rather than a specialized accelerator platform.

Massive memory capacity

64 GB of HBM2e with ECC enables training and inference of large-scale models and memory-bound HPC workloads, with 1,638 GB/s of bandwidth keeping the compute units fed.

The PCIe member of the MI200 series

The MI250 and MI250X are OAM modules for specialized exascale platforms. The MI210 brings the same 2nd Gen CDNA™ architecture to a standard PCIe 4.0 card — which is why it is the Instinct accelerator on our shelf.

2nd Gen CDNA™ architecture

Purpose-built compute silicon delivering exceptional density and power efficiency for parallel processing workloads at scale.

AI acceleration

Optimized matrix operations and mixed-precision compute for accelerated deep learning training, model fine-tuning, and inference deployment.

Memory coherency architecture

3rd Gen AMD Infinity architecture enables high-bandwidth, low-latency communication between GPU compute units and system memory for maximized data throughput.

GPU interconnect

AMD Infinity Fabric™ links enable high-speed peer-to-peer communication between MI210 cards for multi-GPU training and distributed computing workloads.

Intel Xeon Scalable and AMD EPYC servers for AMD Instinct™ workloads

Three host platforms

The MI210 goes on the Intel Xeon Scalable Silver and Gold line, on the two enterprise lines built to order (Xeon Gold 6330 to 6354 on DDR4, Gold 6438Y on DDR5), and on the AMD EPYC line, up to the 64-core EPYC 7763. EPYC gives the card a PCIe 4.0 slot, its own generation; Silver and Gold give it PCIe 3.0, half the host link, which slows a training loop that streams batches over the bus and changes nothing for an inference server whose weights already sit in HBM.

Flexible expansion

Scale your GPU infrastructure on-demand with fast hardware provisioning. Standard upgrades and additional capacity typically deployed within 24 hours.

Expert support

GPU infrastructure specialists available around the clock via live chat and email to assist with deployment, optimization, and troubleshooting.

	MI210	L40S	A100	H100
GPU Architecture	CDNA 2.0	Ada Lovelace	NVIDIA Ampere	Hopper
GPU Memory	64GB HBM2e	48GB GDDR6	80GB HBM2e	80GB HBM2e
GPU Memory Bandwidth	1638 GB/s	864 GB/s	1935 GB/s	2039 GB/s
FP32	22.63 TFLOPS	91.6 TFLOPS	19.5 TFLOPS	51 TFLOPS
TF32 Tensor Core	—	366 TFLOPS	312 TFLOPS	756 TFLOPS
FP16/BF16 Tensor Core	181 TFLOPS	733 TFLOPS	624 TFLOPS	1513 TFLOPS
Power	Up to 300W	Up to 350W	Up to 300W	Up to 350W
The card alone, per month	Loading...	Loading...	Loading...	Loading...

NVIDIA publishes the L40S, A100 and H100 tensor-core figures with structured sparsity, so halve those cells for dense matrix work. AMD quotes the MI210's 181 TFLOPS with no sparsity factor, which makes the FP16/BF16 row one dense figure beside three sparse ones. The MI210 has no TF32 mode, hence the dash in that row.

Frequently asked questions about AMD Instinct MI210 servers

Get answers to common questions about deploying and operating MI210 GPU-accelerated bare metal servers for training, inference, and high-performance computing applications.

What is the AMD Instinct MI210 and which workloads benefit most?

The AMD Instinct MI210 is an enterprise-grade compute GPU built on 2nd Gen CDNA™ architecture, optimized for compute rather than graphics. It excels at deep learning training and inference, scientific simulations, computational fluid dynamics, molecular modeling, and data analytics requiring massive parallel processing capabilities. If you are choosing hardware for a language model, that question has its own page — see LLM hosting.

How does the MI210 differ from the MI250 and MI250X?

The MI250 and MI250X are OAM modules built for specialized exascale platforms, and are not part of our catalog. The MI210 brings the same 2nd Gen CDNA™ architecture to a standard PCIe 4.0 card with 64 GB of HBM2e, which is why it goes into the same Xeon Scalable and EPYC servers as our other GPU options.

What is the typical deployment time for AMD Instinct servers?

A machine already racked with an MI210 is delivered within about 30 minutes of payment verification. When we fit a card we hold in stock, it takes 4 to 24 hours; a card we have to buy in takes five to ten working days, and you are told which before you pay. All AMD Instinct servers support instant OS reload without requiring support tickets, enabling rapid iteration. Network infrastructure is optimized for sustained high-throughput workloads and low-latency connectivity.

Which software frameworks and tools support AMD Instinct GPUs?

AMD Instinct accelerators run on ROCm (Radeon Open Compute), an open-source GPU computing platform supporting PyTorch, TensorFlow, JAX, ONNX Runtime, and other leading ML frameworks. ROCm includes HIP (Heterogeneous-Compute Interface for Portability) enabling straightforward CUDA code porting, plus optimized libraries for linear algebra, FFT, random number generation, and deep neural networks. Full container support via Docker and Kubernetes enables production-scale AI/ML deployment.

What memory does the MI210 provide?

The MI210 carries 64 GB of HBM2e (High Bandwidth Memory) with ECC, delivering 1,638 GB/s of memory bandwidth. That capacity supports large datasets and memory-bound HPC workloads without constant host-accelerator memory shuffling. System RAM and NVMe storage are configured separately on the order form, so they scale independently of the accelerator.