AIME A4004 - Deep Learning Server

If you are looking for a server specialized in maximum deep learning training and inference performance with the highest demands in HPC computing, the AIME A4004 multi-GPU 2U rack server takes on the task of delivering.

The AIME 4004 is basend on the ASUS ESC4000A-E12 barebone which is powered by an AMD EPYC™ Genoa processor generation with up to 96 cores which result 192 parallel CPU threads.

Its GPU-optimized design with high air flow cooling allows the use of up to four high-end double-slot GPUs of the latest NVIDIA generation like the NVIDIA H100, A100 or RTX Ada GPU models.

Definable GPU Configuration

Choose the desired configuration among the most powerful NVIDIA GPUs for Deep Learning:

Up to 4x NVIDIA H100

The latest NVIDIA generation the NVIDIA Hopper H100, the most powerful NVIDIA processor with the highest compute density. The NVIDIA H100 is based on the GH-100 processor in 5nm manufacturing with 18.432 CUDA cores, 576 forth-generation Tensor cores and 80 GB HBM2 memory with the highest data transfer rates. A single NVIDIA H100 accelerator brings 1,5 peta-TOPS fp16 performance. Four accelerators of this type add up to more than 3000 teraFLOPS fp32 performance. The NVIDIA H100 is currently the most efficient deep learning accelerator card available.

Up to 4x NVIDIA A100

The NVIDIA A100 is the flagship of the NVIDIA Ampere processor generation and the current successor to the legendary NVIDIA Tesla accelerator cards. The NVIDIA A100 is based on the GA-100 processor in 7nm manufacturing with 8.192 CUDA cores, 512 third-generation Tensor cores and 40 or 80 GB HBM2 memory with the highest data transfer rates. A single NVIDIA A100 accelerator already breaks the peta-TOPS performance barrier. Four accelerators of this type add up to more than 1000 teraFLOPS fp32 performance.

Up to 4x NVIDIA L40S 48GB

The NVIDIA L40S is built on the latest NVIDIA GPU architecture: Ada Lovelace. It is the direct succesor of the RTX A40 and the passive cooled version of the RTX 6000 Ada. The L40S combines 568 fourth-generation Tensor Cores, and 18.176 next-gen CUDA® cores with 48GB GDDR6 graphics memory for unprecedented rendering, AI, graphics, and compute performance.

Up to 4x NVIDIA RTX 6000 Ada

The RTX ™ 6000 Ada is built on the latest NVIDIA GPU architecture: Ada Lovelace. It is the direct succesor of the RTX A6000 and the Quadro RTX 6000. The RTX 6000 Ada combines 568 fourth-generation Tensor Cores, and 18.176 next-gen CUDA® cores with 48GB of graphics memory for unprecedented rendering, AI, graphics, and compute performance.

Up to 4x NVIDIA RTX 5000 Ada

The RTX ™ 5000 Ada is built on the latest NVIDIA GPU architecture: Ada Lovelace. It is the direct succesor of the RTX A5000/A5500 and the Quadro RTX 6000. The RTX 5000 Ada combines 400 fourth-generation Tensor Cores, and 12.800 next-gen CUDA® cores with 32GB of graphics memory for a convincing rendering, AI, graphics, and compute performance.

Up to 4x NVIDIA RTX A5000

With its 8.192 CUDA and 256 Tensor cores of the 3rd generation, the NVIDIA RTX A5000 is similair powerful than a RTX 3090. However, with its 230 watts of power consumption and 24 GB of memory, it is a more efficient accelerator card and especially for inference tasks the RTX A5000 is a very interesting option.

All NVIDIA GPUs are supported by NVIDIA’s CUDA-X AI SDK, including cuDNN, TensorRT which power nearly all popular deep learning frameworks.

EPYC Genoa CPU Performance

The next-gen Genoa AMD EPYC CPU designed for servers delivers up to 96 cores with a total of 192 threads per CPU with support for the latest DDR5 memory standard.

The available 128 PCI 5.0 lanes of the AMD EPYC CPU allow highest interconnect and data transfer rates between the CPU and the GPUs and ensures that all GPUs are connected with full x16 PCI 5.0 bandwidth.

A large amount of available CPU cores can improve the performance dramatically in case the CPU is used for preprocessing and delivering of data to optimaly feed the GPUs with workloads.

Up to 30 TB Interal High-Speed NVME SSD Storage

Deep Learning is most often linked to high amount of data to be processed and stored. A high throughput and fast access times to the data are essential for fast turn around times.

The AIME A4004 can be configured with two exchangeable U.2 NVMe PCIe 4.0 triple level cell (TLC) SSDs with a capacity of up to 15 TB each, which adds up to a total capacity of 30 TB of fastest SSD storage.

Since each of the SSDs is directly connected to the CPU and the main memory via PCI 4.0 lanes, they achieve consistently high 6000 MB/s read and 4000 MB/s write rates.

As usual in the server sector, the SSDs have an MTBF of 2,000,000 hours and a 5-year manufacturer's guarantee with 1 DWPD.

Optional RAID: Up to 60 TB High Speed NVME SSD Storage

For further storage of large datasets and training checkpoints often additional storage capacity is required. The A4004 offers the option to use its four additional drive bays with a hardware reliable RAID configuration. Up to 80 TB HDD SATA storage or 60 TB fastest NVME SSD storage with RAID levels 0 / 1 / 5 / 10 and 50.

High Connectivity and Management Interface

With the available 100/200/400 Gbit/s QSFP28/QSFP56/OSFP options fastest connections to NAS resources and big data collections are achievable. Also for data interchange in a distributed compute cluster the highest available LAN connectivity is a must have.

The AIME A4004 is completely manageable with ASMB9 (out-of-band) and ASUS Control Center (in-band) makes a successful integration of the AIME A4004 into larger server clusters possible.

Optimized for Multi GPU Server Applications

The AIME A4004 offers energy efficiency with 1+1 redundant titanium power supplies, which enable long time fail-safe operation.

Its thermal control technology, for cooling GPU and CPU tracks individualy, provides more efficient power consumption for large-scale environments.

All setup, configured and tuned for perfect Multi GPU performance by AIME.

The A4004 come with preinstalled Linux OS configured with latest drivers and frameworks like Tensorflow, Keras and PyTorch. Ready after boot up to start right away to accelerate your deep learning applications.