- Home
- Accelerators
- AMD
- AMD MI210
Overview
HPC-Class FP64 in a Standard Server Slot
The AMD Instinct MI210 brings exascale-class double-precision performance to mainstream rack servers. With 45.3 TFLOPS of FP64 matrix performance and 64GB of HBM2e memory at 1.6 TB/s, it delivers over 2X the FP64 throughput of competing PCIe accelerators in the same 300W power envelope.
Built on the CDNA 2 architecture with three Infinity Fabric links for multi-GPU connectivity, the MI210 fits standard PCIe Gen4 x16 slots. No OAM baseboards, no special chassis. It uses the open-source ROCm software stack with PyTorch, TensorFlow, and broad HPC application support.
MI210 vs. A100 PCIe: The HPC Comparison
Both cards draw 300W in a PCIe form factor. The MI210 dominates in FP64. The A100 leads in mixed-precision AI.
| MI210 | A100 PCIe 80GB | |
|---|---|---|
| Architecture | AMD CDNA 2 | NVIDIA Ampere |
| Memory | 64GB HBM2e | 80GB HBM2e |
| Bandwidth | 1.6 TB/s | 2.0 TB/s |
| FP64 | 22.6 TFLOPS | 9.7 TFLOPS |
| FP64 Matrix | 45.3 TFLOPS | 19.5 TFLOPS |
| FP16/BF16 | 181 TFLOPS | 312 TFLOPS |
| TDP | 300W | 300W |
45.3
TFLOPS FP64 Matrix
64GB
HBM2e Memory
1.6
TB/s Memory Bandwidth
300W
PCIe Dual-Slot
Compatible Servers
Dell PowerEdge Servers That Support the MI210
The MI210's 300W, dual-slot FHFL PCIe Gen4 design fits standard 2U GPU-capable servers. Pairs naturally with AMD EPYC platforms.
Dell PowerEdge R760xa
Multi-GPU HPC- 16th Gen GPU-optimized 2U
- Multi-MI210 for parallel HPC workloads
Dell PowerEdge R7625
AMD Platform- 16th Gen AMD EPYC 2U server
- All-AMD stack: EPYC + MI210
Dell PowerEdge R7615
Single-Socket- 16th Gen single-socket AMD EPYC
- Cost-effective MI210 HPC node
Dell PowerEdge R7525
Refurbished AMD- 15th Gen AMD EPYC 2U server
- Budget-friendly MI210 platform
Dell PowerEdge R750xa
Refurbished HPC- 15th Gen GPU-optimized 2U
- Available refurbished for MI210 clusters
Use Cases
Where the MI210 Excels
Scientific Computing & HPC
The MI210's FP64 matrix performance is over 2X the A100 PCIe, making it one of the strongest double-precision accelerators in a standard PCIe form factor. Climate modeling, molecular dynamics, computational fluid dynamics, and genomics workloads benefit from both the raw compute throughput and the 64GB of HBM2e that keeps large datasets in GPU memory. Three Infinity Fabric links enable 2-way and 4-way GPU clusters within a single server.
AI Training & Mixed-Precision Workloads
The MI210 delivers 181 TFLOPS of FP16/BF16 for deep learning training. While the A100 leads in pure AI throughput, the MI210 is a viable option for organizations running HPC and AI convergence workloads on the same hardware. AMD Matrix Cores handle mixed-precision compute efficiently, and ROCm provides upstream framework support for PyTorch and TensorFlow out of the box.
Open Ecosystem & ROCm
AMD's ROCm software stack is open-source with upstream PyTorch and TensorFlow support. The HIP conversion tool helps port CUDA applications. AMD Infinity Hub provides pre-optimized HPC containers. For teams already running AMD EPYC CPUs, the MI210 creates a fully integrated all-AMD compute stack with Infinity Fabric interconnect between GPUs and direct PCIe Gen4 bandwidth to the CPU.
Specifications
AMD Instinct MI210 Accelerator
| Specification | MI210 |
|---|---|
| GPU Architecture | AMD CDNA 2 |
| Process | 6nm FinFET |
| Compute Units | 104 |
| Stream Processors | 6,656 |
| GPU Memory | 64GB HBM2e with ECC |
| Memory Bandwidth | 1.6 TB/s |
| Memory Interface | 4096-bit |
| FP64 | 22.6 TFLOPS |
| FP64 Matrix | 45.3 TFLOPS |
| FP32 | 22.6 TFLOPS |
| FP32 Matrix | 45.3 TFLOPS |
| FP16 / BF16 | 181 TFLOPS |
| INT8 | 181 TOPS |
| Interconnect | 3× Infinity Fabric 3.0 links (100 GB/s each) |
| Host Interface | PCIe Gen4 x16 |
| TDP | 300W |
| Form Factor | Dual-slot, full-height full-length (FHFL) PCIe |
| Thermal | Passive |
| Software | AMD ROCm (open-source) |
Looking for PCIe-Based HPC Acceleration?
The MI210 delivers leadership FP64 in a standard server slot. ServerMonkey can help you build the right configuration.
Request a Quote




