Lenovo ThinkStation P3 Tiny Gen 2 Performance
As noted earlier, the P3 Tiny can come with one of several Intel Core Ultra 2 Series CPUs. Our specific model uses the top-end Core Ultra 9 285, an 8 P core and 16 E core chip that is the pinnacle of the Arrow Lake platform, and allows the P3 Tiny to put up some very competitive numbers.
One of the parks of using desktop hardware here is that the Lenovo system gets access to more hardware resources overall. Whereas a mini-PC based on mobile hardware would top out at 6 P cores and 8 E cores (and 2 LPE cores), the desktop chips get access to another 8 CPU cores altogether. As a result, so long as the system can keep the more powerful and power-hungry desktop chip cool, it can easily outrun the more power-efficient mobile chips.
Though, as noted before, Lenovo does shoot itself in the foot a bit on the performance front by shipping the system with only a single CSO-DIMM, leaving the P3 Tiny short on memory bandwidth. Memory pricing has gone up a lot, but we went into The Difference between a Standard DIMM and a CUDIMM or CSODIMM using Crucial modules (when that was still a brand.)
Geekbench 6 Results
For our look at system performance, we will compare the P3 Tiny to a fellow Lenovo 1L tiny PC, the ThinkCentre M75q Tiny Gen 5. The AMD Ryzen 7 Pro 8700GE (Zen 4) system provides a useful benchmark for what these mini-PCs can do.
At a high level, the P3 Tiny and its Core Ultra 9 285 have little trouble trouncing the slightly older ThinkCentre system in both single-threaded and multi-threaded workloads. On average, the P3 Tiny is ahead by 33% in single-threaded workloads, and an even wider 59% in multi-threaded workloads.
Thankfully for Lenovo, there are no clear and obvious signs of the P3 Tiny being bottlenecked by its limited memory bandwidth. But I would be curious to see what another stick of RAM would do for some of these lower-performing multi-threaded tests.
Meanwhile, the GPU compute tests help to illustrate the significant performance difference between the CPU’s integrated GPU, and the discrete NVIDIA RTX A1000 card installed in the system. Even this low-end NVIDIA card is ahead by almost 300%.
Finally, the story is much the same for the Geekbench AI test. Using DirectML with the ONNX runtime, the A1000 is anywhere between 200% and 300% faster than the integrated GPU.
MLPerf 1.5
We ran MLPerf Client v1.5 on the system as well.
For AI workloads, the RTX A1000 comes in well ahead of the CPU and NPU in both latency and throughput.
Next, let us get to the power consumption and noise.
Can I populate all nvme slots at the same time? Could I fit double sided nvme? With PLP?
@Florian
Yes, all of the M.2 slots can be populated at the same time. As for double-sided drives, I’m afraid I don’t have a well-informed answer for you. Lenovo provides no official guidance in their manuals; they don’t explicitly list DS drives as supposed, but they don’t list them as unsupported, either. Thermal pads are pretty squishy, though, so I would be surprised if you couldn’t make it work.