Nvidia has announced a PCI-e plug-in card of its A100 accelerator with GA100 GPU based on the Ampère architecture. Several manufacturers have announced server hardware in which the cards are used.
The GA100 GPU on the Nvidia A100 PCI-e plug-in card has the same specifications as the sxm4 variant that Nvidia announced earlier, but due to the modified form factor, the TDP has been reduced from 400W to 250W. This is necessary because systems with plug-in cards can be cooled less efficiently than systems that use the sxm4 modules.
According to Nvidia, the peak performance of the PCI-e card is the same as that of the module, but the plug-in card can maintain the maximum performance for a shorter time. The manufacturer states that the new plug-in card offers 90 percent of the performance compared to the sxm4 module. However, consumption is also a lot lower, due to the lower tdp.
With the announcement, Nvidia mentions a large list of manufacturers that will offer the A100 accelerator in rack servers. These are Asus, Atos, Cisco, Dell, Fujitsu, Gigabyte, HPE, Inspur, Lenovo, One Stop Systems, Quanta/QCT and Supermicro. Manufacturers can combine several of the cards, for example Gibabyte makes a server with ten A100 cards. Presumably, the manufacturers use the A100-pci-e cards in combination with AMD Epyc processors. These are currently the only server processors with PCI-e 4.0 support.
In mid-May, Nvidia announced its Ampère architecture for GPUs. The GA100 is the largest chip and it is intended for use in data centers. The full GA100 is used on the A100 accelerator, which Nvidia itself releases in sxm4 form factor and uses in its DGX systems. The chip is made at 7nm and has a surface of 826mm². The GA100 GPU has 6912 cudacores and is combined with 40GB hbm2. Nvidia will later use the Ampère GPU in other configurations for GeForce video cards.
Nvidia A100 accelerator as sxm4 module