Skip to product information
1 of 1

Cisco

Cisco UCSC-GPU-MI210 | Instinct MI210 64GB HBM2e GPU, 300W, PCIe 4.0

SKU:UCSC-GPU-MI210

Request Quote
Sale Sold out
Shipping calculated at checkout.

Stock Status: Enquire

 Trade supply only. Server & GPU quote requests need a company name, ABN and work email address. We do not quote Gmail, Outlook or Hotmail addresses, or supply for home use. 

20-min response

Need a quote or bulk pricing? Answered within 20 minutes, Sydney business hours.

Custom Built Servers

Need a different spec, quantity or configuration?

We build configured-to-order refurbished servers from Cisco, Dell, HPE, NVIDIA Lenovo and Supermicro. Tell us what you need and we'll come back with options within 24 hours.

Configure your server now →

Description

The Instinct MI210 is a passively cooled, double-slot, full-height PCIe 4.0 compute accelerator from the AMD Instinct MI200 Series, built on the CDNA 2 architecture. It is offered by Cisco as a UCS C-Series option and is aimed at HPC and scientific computing as well as large-scale machine-learning training in single-server or multi-GPU data-centre deployments.

The card carries 64 GB of HBM2e memory on a 4096-bit interface, delivering 1.6 TB/s of peak memory bandwidth with full-chip ECC. It provides 104 compute units and 6,656 stream processors, with a peak of 181 TFLOPs for FP16 and bfloat16 and 45.3 TFLOPs for FP64 matrix workloads. Three Infinity Fabric links at 100 GB/s each allow GPU-to-GPU connection in multi-GPU configurations. The board draws 300 W TDP over a PCIe 4.0 x16 bus.

Being passively cooled, the card relies on server chassis airflow, so the host needs a PCIe 4.0 x16 full-height, full-length double-width slot, suitable GPU power cabling and PSU capacity. It runs on a 64-bit Linux OS with the AMD ROCm software stack.

Enquire with XS Network Tech for new or used pricing and availability.

Features

- CDNA 2 architecture optimised for double- and single-precision HPC and large-scale AI training
- 64 GB of HBM2e memory with 1.6 TB/s of peak bandwidth
- Full-chip ECC memory with RAS support and page retirement
- Three Infinity Fabric links for GPU-to-GPU connection in multi-GPU configurations
- Supported by the AMD ROCm software ecosystem
- Programming environments include OpenMP, OpenCL, ISO C++ and CUDA via the HIP conversion tool
- Passive cooling design for chassis-airflow data-centre servers

Warranty

All products sold by XS Network Tech include a 12-month warranty on both new and used items. Our in-house technical team thoroughly tests used hardware prior to sale to ensure enterprise-grade reliability.

All technical data should be verified on the manufacturer data sheets.

View full details

specs-tabs

Collapsible content

Technical Specifications

FAQs

Technical Specifications

- GPU architecture: CDNA 2
- Compute units: 104
- Stream processors: 6,656
- Peak engine clock: 1700 MHz
- Peak FP16 performance: 181 TFLOPs
- Peak bfloat16 performance: 181 TFLOPs
- Peak INT8 performance: 181 TOPs
- Peak INT4 performance: 181 TOPs
- Peak FP32 performance: 22.6 TFLOPs
- Peak FP32 matrix performance: 45.3 TFLOPs
- Peak FP64 performance: 22.6 TFLOPs
- Peak FP64 matrix performance: 45.3 TFLOPs
- Memory: 64 GB HBM2e
- Memory interface: 4096-bit
- Memory clock: 1.6 GHz
- Peak memory bandwidth: 1.6 TB/s
- Memory ECC: Yes, full-chip
- Bus type: PCIe 4.0 x16
- Form factor: PCIe add-in card
- Infinity Fabric links: 3 at 100 GB/s
- Cooling: Passive
- Board size: Full height, 10.5 in (267 mm) long, double slot
- Power: 300 W TDP

FAQs

Q: Does this GPU need a licence or subscription?
A: No licence or subscription is listed for the card itself. Drivers and the ROCm software stack are available from AMD.

Q: How much power does the card draw and how is it cooled?
A: The card has a 300 W TDP and is passively cooled, so it depends on server chassis airflow. GPU power cabling and PSU sizing depend on the host server.

Q: What does the host server need to support this card?
A: The host needs a PCIe 4.0 x16 full-height, full-length double-width slot and a 64-bit Linux OS with the ROCm software stack. It is offered as a Cisco UCS C-Series option.

Q: Can multiple cards be linked together?
A: Yes. The card has three Infinity Fabric links at 100 GB/s each, which allow GPU-to-GPU connection in multi-GPU configurations.

Q: What workloads is it designed for?
A: It targets HPC and scientific computing using double- and single-precision maths, as well as large-scale machine-learning training.

  • Request a Quote

    Looking for competitive pricing? Submit a request, and our team will provide a tailored quote that fits your needs.

  • Contact Us Directly

    Have a question or need immediate assistance? Call us for expert advice and real-time support.

    Call us Now  
  • Contact Us Directly

    Have a question or need immediate assistance? Call us for expert advice and real-time support.

    Contact us