Recently I was provided remote access to a NVIDIA Vera server with the dual socket Vera CPU Superchip providing 176 Olympus cores plus with NVIDIA's Spatial Multithreading "SMT" implementation providing 352 total threads. With this benchmarking was free rein for testing, including CPU power and frequency monitoring. Over the next few weeks will be a number of articles looking at different areas of NVIDIA Vera performance. One of the areas I was most eager to test was looking at Vera's Spatial Multithreading. Here is a first look at the impact of NVIDIA Vera SMT.

Spatial Multithreading is new with NVIDIA Vera and their SMT design differs from Simultaneous Multi Threading from the likes of AMD and Intel processors. Vera SMT relies on physical partitioning of core hardware resources rather than dynamic time-slicing of shared execution units. Vera SMT provides two threads per core similar to AMD and Intel processors.

NVIDIA believes their SMT implementation delivers improved utilization and more predictable behavior, especially for agentic AI workloads. Those wishing to go in-depth over NVIDIA Spatial Multithreading can learn more via the NVIDIA Technical Blog.

For today's article I was looking at the dual socket NVIDIA Vera performance with SMT on (the default) and then SMT disabled. NVIDIA Vera SMT can be disabled at run-time just like SMT on other CPUs: the /sys/devices/system/cpu/smt/control sysfs interface.

Pn this NVIDIA Olympus setup running NVIDIA's modified Ubuntu 24.04 stack with the Linux 6.17 64K kernel and GCC 15.3 compiler I ran a variety of benchmarks that are able to scale to high core/thread counts and also typically benefit from SMT at least on Intel/AMD CPUs, so was curious to measure the difference on Vera when toggling SMT. During the benchmarking the peak CPU frequency monitoring and CPU power consumption were also monitored for seeing the impact there from Spatial Multithreading.