Nvidia attacks AMD and Intel with ARM server processor Vera
Server racks with up to 22,528 CPU cores: Nvidia's Vera CPU with Olympus cores is not only entering the market in conjunction with AI accelerators.
Rack with servers featuring Nvidia Vera CPU
(Image: Nvidia)
Nvidia's ARM server processor "Vera" with self-developed ARM cores aims to compete in the territories of AMD Epyc and Intel Xeon. This is because Nvidia is not only combining Vera with the "Rubin" AI accelerator but also building pure CPU systems with it.
At its GTC trade show, Nvidia also announced a "Vera CPU Rack" for the second half of 2026. The water-cooled version can accommodate modules with a total of 256 Vera processors, each having 88 cores. This means a rack will run a total of 22,528 CPU cores with 45,056 threads, as each of the "Olympus" cores, compatible with ARMv9.2, processes two threads simultaneously.
More cores in x86
When it comes to the sheer number of CPU cores per rack, AMD and Intel offer more. For example, Supermicro sells "Twin" servers that contain two individual servers per rack unit (RU), each with two CPU sockets. This allows for a total of 160 physical processors to be housed in 40 RU. In the water-cooled versions, each can dissipate up to 500 watts, resulting in 75 kW of pure CPU power consumption per rack.
Equipped with 160 AMD Epyc 9965 with 192 Zen 5c cores each, this configuration results in 30,720 CPU cores and 61,440 threads per rack. Even with the 144-core Intel Xeon 6780E, 23,040 cores would be possible, but without Simultaneous Multithreading, as Intel's E-cores do not support it. However, the Xeon 6780E consumes 330 watts, and the 6766E version even uses only 250 watts. Intel also offers preferred customers a Xeon 6 with up to 288 E-cores, meaning up to 46,080 per rack.
(Image:Â Nvidia)
Vera Advantages
Nvidia lists several advantages of Vera compared to the competition and already has customers: Meta, among others, intends to use Vera, as do Alibaba, Cloudflare, Nebius, and Oracle Cloud Infrastructure (OCI).
Nvidia promises particularly high efficiency for its ARM cores. Furthermore, Vera does not connect Registered DIMMs with DDR5 SDRAM but rather eight SOCAMM2 modules with more power-efficient LPDDR5X memory. This allows each Vera CPU to achieve a data transfer rate of around 1.2 TByte/s and control up to 1.5 TByte of main memory.
Vera is also among the first processors to support PCI Express 6.0. However, this is also expected to come with AMD Venice in the second half of the year.
(Image:Â Nvidia)
High Single-Thread Performance
Without releasing specific benchmark results yet, Nvidia promises particularly high single-thread performance for the individual Olympus cores. This is expected to provide advantages for certain computational tasks in AI workflows.
Videos by heise
The Olympus cores also implement Simultaneous Multithreading (SMT) differently than the x86 cores from AMD or Intel. Nvidia refers to this as Spatial Multithreading. This is intended to isolate competing threads more strongly from each other, leading to less performance fluctuation than with other SMT implementations.
In its own servers, Nvidia combines Vera processors with its own ConnectX network adapters or BlueField-4 SmartNICs. This allows Vera racks to be easily combined with Nvidia's AI accelerators, for which Nvidia also sells switches.
Other manufacturers, such as Dell, HPE, Lenovo, and Supermicro, also plan to sell Vera servers. They use the modular Nvidia MGX system. Air-cooled Vera servers and single-processor (single socket) versions are also planned.
(ciw)
