Graviton, a server provider based in Russia, recently unveiled its inaugural AI and HPC server equipped with domestically produced processors. This new server supports up to eight GPUs for handling tasks related to artificial intelligence and high-performance computing. Graviton has not specified which GPUs will be used, likely due to complications in acquiring them under current sanctions. It remains uncertain if this server will deliver competitive performance metrics.
The Graviton S2124B server features two unspecified 48-core CPUs operating at 2 GHz, and it utilizes DDR4-3200 memory, as reported by ServerNews. These specifications seem to align with the Baikal Electronics BE-S1000 chip, which includes 48 Arm Cortex-A75 cores and is capable of 2-way and 4-way symmetric multiprocessor configurations.
It appears that the CPU’s frequency in this server model is reduced by 500 MHz from the original, possibly due to a shift from TSMC’s 16FFC to an alternate manufacturing node. Baikal might have lowered the frequency to improve chip yields or lessen power consumption.
By analyzing the performance claims and dividing by eight, we can infer the type of GPUs Graviton plans to incorporate in the S2124B server — specifically focusing on tensor core performance. The expected per-accelerator performance — 60 FP64 TFLOPS for supercomputing and 3340 FP8/INT8 TFLOPS/TOPS for AI — suggests the use of Nvidia’s H100 PCIe GPU. Users of the S2124B will need to depend on Nvidia’s CUDA ecosystem, though achieving peak performance may be challenging without Nvidia’s direct support. The server’s performance potential, especially given its relatively obscure Arm-based CPUs, is yet to be fully determined with these Hopper accelerators.
Beyond the CPUs and GPU accelerators, the Graviton S2124B allows for the integration of either 12 SATA drives or 12 NVMe U.3 SSDs.
Graviton is currently accepting orders for the S2124B server and is also offering opportunities for customers to test the server. However, the price of the server has not been disclosed, and it’s not clear if Graviton can secure a supply of the Nvidia H100 GPUs.
Alexander Filchenkov, head of server and network systems at Graviton, stated, “We are dedicated to delivering IT solutions that align with market needs promptly. With this launch, we have successfully developed and produced servers essential for intricate computations using local processors. This achievement marks a significant advancement in our domestic computing technology and will help our clients effectively tackle data processing challenges.”
