7¡Á better efficiency in HPC applications, up to 9¡Á faster AI training, and up to 30¡Á faster AI inference than the NVIDIA HGX A100.
| GPU Model | H100 SXM5 |
| GPU Memory | 80 GB HBM3 |
| Bandwidth | 3.35 TB/s |
| FP8 Tensor | 3,958 TFLOPS |
| NVLink | 4th Gen (900 GB/s) |
| NVSwitch | 4¡Á |
| Power | 700W per GPU |
| Cooling | Air / Liquid |
| GPU Model | H100 PCIe |
| GPU Memory | 80 GB HBM2e |
| Bandwidth | 2 TB/s |
| FP8 Tensor | 3,026 TFLOPS |
| NVLink | N/A (PCIe Gen5) |
| Power | 350W per GPU |
| Cooling | Passive |
We are aware of the recent article published by Bloomberg which has raised concerns regarding our operations. We would like to clarify that the article contains misleading information and incorrect insinuations, including suggestions that we may have been involved in illegal chip transfers.
We want to be absolutely clear: over the past six months, our company has undergone multiple on-site inspections and reviews by key institutions, including the Bureau of Industry and Security of the U.S. Commerce Department, the Ministry of Investment, Trade and Industry of Malaysia, and NVIDIA. These thorough investigations have confirmed that there has been no evidence of any violations concerning the illicit transfer of chips or any illegal activities, and this demonstrates that we are a compliant and legitimate company.
We operate fully within the bounds of all applicable export control regulations and maintain the highest standards of legal and ethical conduct. Our company has always adhered to the law, and we take these accusations seriously. As such, we do not rule out pursuing legal action to protect our reputation and to hold Bloomberg accountable for the misleading nature of the article.
We appreciate the continued trust and support from our clients and partners.