11:28Machine-translated from Azerbaijani by AIRead in Azerbaijani
ASUS Unveils AI Workstation With 748 GB of Memory
ASUS has unveiled the ExpertCenter Pro ET900N G3 workstation, which it developed for artificial intelligence computing in data centers. The device is built around the Nvidia GB300 Grace Blackwell Ultra Superchip. The official price has not been announced, but the system is expected to cost 100 000-120 000 dollars (approximately 170 000-204 000 manats). It is 584 mm high and 232 mm wide. Its depth is 565 mm, and it weighs 27 kg.
The ExpertCenter Pro ET900N G3 contains 72 Arm cores and a Blackwell Ultra graphics chip. The system is also equipped with 748 GB of unified memory. Unified memory is a shared memory pool for the processor and graphics chip. This allows large artificial intelligence models to run without moving data between separate memory pools. This configuration reduces latency and speeds up the simultaneous execution of more tasks. The workstation has a silver metal case and a mesh front panel. The case has 2 large handles to make it easier to carry. The processor, memory and network card are cooled by a closed-loop liquid system. There is also a separate fan for the networking hardware. Power is supplied by an 80 PLUS Titanium-certified 1 600 W power supply. For storage, it has 2 NVMe drives, each with 2 TB. There are 3 PCIe slots and 2 bays for additional hardware.
In the MiniMax M2.7 test, the system generated 4 793 output tokens per second across 128 parallel streams. A token is a small unit of text that an artificial intelligence model processes and generates. The GLM-5.2 model processed 215 GB of data directly in unified memory. As a result, the memory latency that slows down large artificial intelligence models on conventional computers was reduced.
The device has 2 network ports, each rated at 400 gigabits. Their combined data transfer speed is 800 Gbit/s. These ports allow 2 workstations to be combined into a single system. The connected devices have a total of 1 496 GB of unified memory. This makes it possible to run even larger artificial intelligence models that do not fit in the memory of a single workstation. In the GPT-OSS-120B test, the output speed was 9 280 tokens per second. The overall processing speed reached more than 45 000 tokens per second. The system ran 128 artificial intelligence agents simultaneously without serious slowdowns or failures. With dozens of agents handling coding, research and text generation, power consumption approached 1 000 W. Although the case has been visually compared to the classic Apple Mac Pro, this workstation is significantly more powerful than conventional high-end computers.
The ExpertCenter Pro ET900N G3 contains 72 Arm cores and a Blackwell Ultra graphics chip. The system is also equipped with 748 GB of unified memory. Unified memory is a shared memory pool for the processor and graphics chip. This allows large artificial intelligence models to run without moving data between separate memory pools. This configuration reduces latency and speeds up the simultaneous execution of more tasks. The workstation has a silver metal case and a mesh front panel. The case has 2 large handles to make it easier to carry. The processor, memory and network card are cooled by a closed-loop liquid system. There is also a separate fan for the networking hardware. Power is supplied by an 80 PLUS Titanium-certified 1 600 W power supply. For storage, it has 2 NVMe drives, each with 2 TB. There are 3 PCIe slots and 2 bays for additional hardware. In the MiniMax M2.7 test, the system generated 4 793 output tokens per second across 128 parallel streams. A token is a small unit of text that an artificial intelligence model processes and generates. The GLM-5.2 model processed 215 GB of data directly in unified memory. As a result, the memory latency that slows down large artificial intelligence models on conventional computers was reduced. The device has 2 network ports, each rated at 400 gigabits. Their combined data transfer speed is 800 Gbit/s. These ports allow 2 workstations to be combined into a single system. The connected devices have a total of 1 496 GB of unified memory. This makes it possible to run even larger artificial intelligence models that do not fit in the memory of a single workstation. In the GPT-OSS-120B test, the output speed was 9 280 tokens per second. The overall processing speed reached more than 45 000 tokens per second. The system ran 128 artificial intelligence agents simultaneously without serious slowdowns or failures. With dozens of agents handling coding, research and text generation, power consumption approached 1 000 W. Although the case has been visually compared to the classic Apple Mac Pro, this workstation is significantly more powerful than conventional high-end computers.