Graviton, a Russia-based server supplier, has announced its first AI and HPC server powered by processors developed domestically. This device can support up to eight compute GPUs to process artificial intelligence and supercomputer workloads, but the maker does not recommend any particular GPUs (though they can be easily guessed) probably because getting them amid sanctions is illegal. Furthermore, whether the machine can achieve competitive performance numbers is unclear.
The Graviton S2124B server is based on two undisclosed 48-core CPUs running at 2 GHz and featuring DDR4-3200 memory, according to ServerNews . The basic specification of the processor likely suggests the Baikal Electronics BE-S1000 server-grade chip that packs 48 Arm Cortex-A75 cores and supports 2-way and 4-way symmetric multiprocessor (SMP) configurations. Yet, this particular version of the BE-S1000 seems to clock the CPU at 500 MHz below its original frequency , which is likely a result of porting its design from TSMC's 16FFC to a different production node at a different foundry. It is also possible that Baikal reduced operating clocks to increase yields or reduce power consumption.
Dividing these performance numbers by eight, we can determine which GPUs Graviton intends to install in its S2124B machine. Per-accelerator performance numbers — 60 FP64 TFLOPS of compute power for supercomputing and 3340 FP8/INT8 TFLOPS/TOPS performance for AI — point to Nvidia's H100 PCIe. In all cases, we are referring to Tensor core performance. That said, those who use the S2124B will have to rely on Nvidia's CUDA ecosystem. However, without support from Nvidia, it is unlikely that peak performance will be achieved. Furthermore, considering the platform is Arm-based and relatively unknown, it remains to be seen how much performance can actually be extracted from these Hopper accelerators.
In addition to two CPUs and eight GPU accelerators, Graviton's S2124B can integrate 12 SATA drives or 12 NVMe U.3 SSDs.
The Graviton S2124B server is currently available for order, with customers also invited to apply for testing opportunities, but its pricing is unknown. Additionally, it is unclear whether Graviton can supply Nvidia H100 GPUs.
"We take pride in consistently offering IT solutions that meet market demands in a timely manner," said Alexander Filchenkov, head of server and network systems at Graviton. "This time, we successfully developed and manufactured servers critical for complex computations using domestic processors. This product represents a significant step in advancing domestic computing technologies and will enable our clients to efficiently address data processing challenges."
