Supermicro SYS-112D-36C-FN3P Performance
For CPU topology, the 36-core Xeon 6 SoC gives us a different core count than the 40-core model we saw in Supermicro’s other server, but it is otherwise the same basic core generation. Core-to-core latency is one of the first checks because it helps show how threads communicate across the SoC.
Same-core SMT latency gives another view of the platform. Together with the physical-core chart, it is useful when thinking about workload placement and thread pinning.
Next, let us get to the Geekbench results.
Geekbench Results
Geekbench 6 gives us a quick comparison between the 36-core Xeon 6 6553P-B SoC and the 40-core 6716P-B baseline from the earlier SYS-112D-40C-FN8P review. The 36-core part tracks the 40-core part closely in places where per-core behavior matters, while multi-threaded results reflect the smaller core count.
Geekbench 5 shows a similar comparison.
This is not the main benchmark story for a platform like this, but it is useful as a familiar reference point before moving into the AgentSTH data.
AgentSTH V5 Results Preview
In the 40C review, we previewed AgentSTH V5 using charts that compare this 36-core Xeon 6553P-B platform against the 40-core Xeon 6716P-B platform. That data belongs here as well because it is one of the main performance comparisons between the two systems.
AgentSTH is focused on CPU-based agentic workloads, not LLM inference on GPUs. Many agentic systems spend a lot of time on tasks such as parsing, compression, coordination, queues, hashing, and state handling, so the CPU side still matters.
The high-level lesson from the full socket chart is that running multiple agents can use the CPU more effectively than one large job that waits on a single slower path. That is why the multi-agent view is important for modern servers.
With AgentSTH running across 32 cores, the 36-core and 40-core systems are close in many parts of the suite because they use the same core generation. That makes this view useful because it removes some of the obvious core-count difference. 32-cores is just one shape we are testing for these workloads.
The throughput subtests show that the differences are not simply a matter of one model always winning. Once the benchmark is held to the same 32-core scale, cache behavior, clock behavior, and the subtest mix can change the result.
Coordination tasks are where the platform differences become more interesting. The Xeon 6716P-B can do better when threads repeatedly revisit related state, which helps keep metadata hot.
When the workload becomes more diverse due to allocator and hash-table churn, the Xeon 6553P-B platform can still look stronger in this V5 data. This underscores how the difference between the two Supermicro servers is more than just CPU core counts. It is not just a lower-core version of the 40-core system.
We should probably note that we are now on AgentSTH V7, having refined the benchmark set a bit after Ubuntu 26.04 LTS was released. Still, we wanted to maintain parity with the previous server review.
Next, let us get to the power consumption.
