Pasar al contenido principal

El próximo salto de la infraestructura de IA ya está aquí

Los modelos avanzados de IA actuales están cambiando nuestras vidas rápidamente. La infraestructura de computación acelerada está evolucionando a una velocidad sin precedentes en todos los segmentos del mercado. Una infraestructura flexible, robusta y escalable de forma masiva con GPU de última generación está haciendo posible un nuevo capítulo de la IA.

En estrecha colaboración con NVIDIA, Supermicro ofrece una de las selecciones más amplias de sistemas con certificación NVIDIA que proporcionan el máximo rendimiento y eficiencia desde pequeñas empresas hasta clústeres de formación de IA masivos y unificados con las nuevas GPU NVIDIA H100 y H200 Tensor Core.

Juntos, conseguimos hasta nueve veces el rendimiento de entrenamiento de la generación anterior para algunos de los modelos de IA más desafiantes, reduciendo una semana de tiempo de entrenamiento a sólo 20 horas. Los sistemas Supermicro con las GPU H100 PCIe y HGX H100, así como las recién anunciadas GPU HGX H200, incorporan conectividad PCIe 5.0, NVLink y NVLink Network de cuarta generación para escalabilidad horizontal, y las nuevas tarjetas NVIDIA ConnectX®-7 y BlueField®-3 que potencian GPUDirect RDMA y almacenamiento con NVIDIA Magnum IO y el software NVIDIA AI Enterprise.

Cartera de sistemas de GPU

Supermicro HGX Systems

8U NVIDIA HGX H100/H200 8-GPU System

8U NVIDIA HGX H100/H200 8-GPU System

Adecuado para los modelos de entrenamiento de IA y HPC de mayor escala de la actualidad, con capacidad térmica superior con acústica reducida, más E/S y almacenamiento de gran capacidad

  • GPU: NVIDIA HGX H100/H200 8-GPU with up to 141GB HBM3e memory per GPU
  • GPU-GPU Interconnect: 900GB/s GPU-GPU NVLink interconnect with 4x NVSwitch – 7x better performance than PCIe
  • CPU: Dual 4th/5th Gen Intel Xeon® or AMD EPYC 9004 series processors
  • Memory: Up to 32 DIMM slots: 8TB DDR5-5600
  • Storage: Up to 16x 2.5" Hot-swap NVMe drive bays
  • NIC: High-performance 400 Gb/s NVIDIA BlueField®-3 or NVIDIA ConnectX®-7
  • Networking: Scale out with NVIDIA Quantum-2 InfiniBand or NVIDIA Spectrum-X Ethernet
4U Liquid-Cooled NVIDIA HGX H100/H200 8-GPU System

4U Liquid-Cooled NVIDIA HGX H100/H200 8-GPU System

Highest density and efficiency with D2C liquid cooling for both GPUs and CPUs to optimize performance and energy cost

  • GPU: NVIDIA HGX H100/H200 8-GPU with up to 141GB HBM3e memory per GPU
  • GPU-GPU Interconnect: 900GB/s GPU-GPU NVLink interconnect with 4x NVSwitch – 7x better performance than PCIe
  • CPU: Dual 4th/5th Gen Intel Xeon® or AMD EPYC 9004 series processors
  • Memory: Up to 32 DIMM slots: 8TB DDR5-5600
  • Storage: Up to 16x 2.5" Hot-swap NVMe drive bays
  • NIC: High-performance 400 Gb/s NVIDIA BlueField®-3 or NVIDIA ConnectX®-7
  • Networking: Scale out with NVIDIA Quantum-2 InfiniBand or NVIDIA Spectrum-X Ethernet
  • Direct-to-chip liquid cooling to support CPUs up to 350W and GPUs up to 700W

Supermicro’s NVIDIA MGX™ Systems

2U NVIDIA MGX System with Intel- or ARM-based Processor

NVIDIA MGX Systems with NVIDIA GH200 Grace Hopper Superchip

Integrated CPU, GPU, and Coherent Memory System for Cloud-Scale AI Inference

  • CPU/GPU: Up to 2x 72-core Grace Arm Neoverse V2 CPU + H100 Tensor Core GPU in a single chip via NVIDIA GH200 Grace Hopper Superchip in 1U
  • Memory (onboard): CPU: 480GB LPDDR5X GPU: up to 144GB HBM3e
  • Storage: 8x Hot-swap E1.S drives and 2x M.2 NVMe drives
  • NIC: High-performance 400 Gb/s NVIDIA BlueField®-3 or NVIDIA ConnectX®-7
  • Networking: Scale out with NVIDIA Quantum-2 InfiniBand
2U NVIDIA MGX System with Intel- or ARM-based Processor

NVIDIA MGX PCIe GPU Systems

Modular Building Block Platform Supporting Today’s and Future GPUs, CPUs, and DPUs

  • GPU: Hasta 4 GPU PCIe NVIDIA de doble ancho, incluyendo H100 PCIe, H100 NVL PCIe, L40S y más
  • CPU: NVIDIA Grace™ CPU Superchip, or 4th/5th Gen Intel® Xeon® Scalable processor
  • Memory: Up to 480GB integrated LPDDR5X DRAM (with Grace CPU Superchip) or up to 2TB 4800MT/s ECC DDR5 DRAM (with Intel CPU)
  • Storage: Up to 8 Hot-Swap E1.S NVMe
  • NIC: High-performance 400 Gb/s NVIDIA BlueField®-3 or NVIDIA ConnectX®-7
  • Networking: Scale out with NVIDIA Quantum-2 InfiniBand

Supermicro PCIe GPU Systems

Supermicro 5U PCIe 10-GPU System

Supermicro 4U or 5U PCIe 10-GPU System

Suited for Today’s Largest Scale AI Training Models and HPC, featuring superior thermal capacity with reduced accoustics, More I/O, and Vast Storage

  • GPU: Up to 10 PCIe 5.0 GPUs including NVIDIA H100 PCIe, H100 NVL, or NVIDIA L40S
  • CPU: Dual 4th/5th Gen Intel Xeon® or AMD EPYC 9004 series processors
  • Memory: Up to 32 DIMM slots: 8TB DDR5-5600
  • Cooling: Optimized thermal capacity to support CPUs up to 350W and GPUs up to 700W with air cooling
  • NIC: High-performance 400 Gb/s NVIDIA BlueField®-3 or NVIDIA ConnectX®-7
  • Networking: Scale out with NVIDIA Quantum-2 InfiniBand or NVIDIA Spectrum-X™ Ethernet
Supermicro 4U 4-GPU Workstation

Supermicro 4U 4-GPU Workstation

Optimizado para la colaboración en el metaverso 3D, científicos de datos y creadores de contenidos. Disponible en formato para montaje en bastidor y estación de trabajo

  • GPU: Up to 4x double width, full length GPUs, including NVIDIA L40S and NVIDIA RTX 6000 Ada Generation
  • CPU: Dual 4th/5th Gen Intel Xeon® Scalable Processors
  • Memory: 16 DIMM slots, up to 4TB DDR5-5600
  • Expansion: 4 PCIe 5.0 x16 (double-width) slots, 3 PCIe 5.0 x16 (single-width) slots
  • Storage: 8x 3.5" Hot-swap SATA/NVMe/SAS drive bays
Ethernet and InfiniBand for Scale-Out AI SuperClusters

Supermicro: Your One-Stop Shop for AI Infrastructure, powered by NVIDIA Technology

Supermicro is a global leader in architecting and deploying comprehensive rack-scale solutions tailored for AI and HPC. Our solutions are designed to meet the most demanding customer requirements to ensure optimal performance, scalability, and efficiency. NVIDIA Networking Platforms serve as the nervous system for compute infrastructure. Both NVIDIA Spectrum-X and Quantum-2 InfiniBand are available today to meet your demands.

1:1 GPU to NIC 400Gb/s with Non-blocking

Liquid-Cooled 4U GPU Rack

NVIDIA Quantum-2 InfiniBand

NVIDIA Quantum-2 (QM9700)

NVIDIA Quantum-2 (QM9700)

NVIDIA ConnectX®-7

NVIDIA ConnectX®-7

NVIDIA Spectrum-X Ethernet

NVIDIA Spectrum-4 (SN5600)

NVIDIA Spectrum-4 (SN5600)

NVIDIA BlueField®-3

NVIDIA BlueField®-3

Sistemas con certificación NVIDIA de Supermicro

Con el continuo despliegue de aplicaciones y cargas de trabajo avanzadas, los clientes necesitan servidores gestionables, seguros y escalables para sus centros de datos. La atractiva línea de servidores de alto rendimiento de Supermicro compatibles con GPU y DPU NVIDIA incluye un número creciente de sistemas con certificación NVIDIA, y muchos más se encuentran actualmente en proceso de certificación. Cada configuración de servidor/GPU obtiene su propia certificación.

NVIDIA Certified Solutions logo

Más información

Featured Supermicro NVIDIA Certified and Qualified Systems

SYS-821GE-TNHR

NVIDIA-Certified with HGX H100 8-GPU
NVIDIA ConnectX®-7 and BlueField®-3 support
5th/4th Gen Intel® Xeon® Scalable processors
32 DIMM slots, up to 8TB DDR5-5600

AS -4125GS-TNRT

NVIDIA-Qualified with H100 PCIe GPUs
NVIDIA ConnectX®-7 and BlueField®-3 support
Dual-Socket, AMD EPYC™ 9004 Series Processors
24 DIMM Slots, up to 6TB DDR5-4800

SYS-421GE-TNHR2-LCC

NVIDIA-Certified with HGX H100 8-GPU
NVIDIA ConnectX®-7 support
5th/4th Gen Intel® Xeon® Scalable processors
Direct-to-chip Liquid Cooling

AS -8125GS-TNHR

NVIDIA-Certified with HGX H100 8-GPU
NVIDIA ConnectX®-7 and BlueField®-3 support
Dual-Socket, AMD EPYC™ 9004 Series Processors
24 DIMM Slots, up to 6TB DDR5-4800

SYS-521GE-TNRT

NVIDIA-Qualified with H100 NVL, H100 PCIe, and L40S
NVIDIA ConnectX®-7 support
5th/4th Gen Intel® Xeon® Scalable processors
32 DIMM slots, up to 8TB DDR5-5600

ARS-221GL-NR

NVIDIA-Qualified with H100 NVL and L40S
NVIDIA ConnectX®-7 support
NVIDIA Grace™ CPU Superchip with 144 Cores
Up to 480GB LPDDR5X onboard memory

Certain products may not be available in your region