What was actually built
According to Tom's Hardware's report on the Reddit post:
- Hardware: The system has a 40-core 3rd-gen Xeon Scalable CPU and PMem 200 modules. The GPU stayed out of inference.
- Speed: The builder reported 0.85 tokens per second at 1-bit quantization. At 3-bit quantization, speed reportedly fell to 0.75 tokens per second.
- Tuning: These numbers are pre-tuning. The builder had not yet adjusted thread counts or NUMA affinity.
- Mode: The build used App Direct Mode. That exposes Optane as a separate persistent region that software maps directly.
- Plans: The builder intends to add GPUs and place parts of the model in DRAM as a faster tier. These are plans, and "more benchmarks to come" is a promise, not a result.
These figures come from one builder's report. They are not an independently replicated benchmark.
Why Optane, and why it's slow
Kimi K3 is huge. Moonshot's model card describes an open-weight mixture-of-experts model with 2.8T total parameters and 104B activated per token. It has 93 layers and 896 experts, with 16 experts selected per token. It also has a 1,048,576-token context window and quantization-aware-trained MXFP4 weights. "Open weight" does not mean "small."
Tom's Hardware says even the smallest community quantization needs at least 610GB of RAM. I could not confirm that figure in Moonshot's own model card, so treat it as the outlet's claim. The same report says Moonshot recommends data-center inference engines. The model card does name vLLM, SGLang and TokenSpeed as recommended engines.
Secondhand Optane is one of the few ways to reach that capacity at home. The builder says it costs "a fraction" of equivalent DRAM, which has become expensive. Optane is slower than DRAM, though, and the reported throughput shows it.
The comparison with the earlier build
The new project follows an earlier first-gen Optane build. That build reportedly reached roughly 4 tokens per second on the smaller 1-trillion-parameter Kimi K2.5. It is not a like-for-like comparison:
| Factor | Earlier build | Latest build |
|---|---|---|
| Optane generation | First-gen (PMem 100) | Second-gen (PMem 200) |
| Model | Kimi K2.5 (1T) | Kimi K3 (2.8T) |
| Memory mode | Memory Mode | App Direct |
| Quantization | 2-bit | 1-bit and 3-bit |
| GPU | RTX 3060 handling part of the model | None for inference |
The model, mode, quantization and GPU differ, so the 4 versus 0.85 tokens per second gap says little about Optane generations.
Heat is the real warning
The builder's advice is to "treat active DIMM cooling as mandatory." They used dual 12W blower fans aimed at the modules. The builder also says the DIMMs got extremely hot under sustained inference and could endanger neighboring DRAM. That is the builder's observation, not a documented Intel finding.
Intel's documentation does back the general concern:
- Operating power: Intel's support article lists 12W or 15W for 128GB modules. It lists 12W, 15W or 18W for 256GB and 512GB modules. The Tom's Hardware claim of 12W or 15W regardless of capacity misses that 18W option.
- Thresholds: Unconstrained performance runs from 0 to 83°C. The host controller throttles at 83°C, and thermal trip shutdown occurs at 85°C.
- Airflow: Intel's product brief reportedly specifies 48°C ambient at 2.4m/s for 12W modules and 43°C at 2.7m/s for 15W modules. That is data-center airflow, not what a typical home tower provides.
The fan count and wattage are the builder's choice, not an Intel prescription. A rough estimate: eight 256GB modules at 15W would put around 120W of heat next to your DRAM. That is my arithmetic from Intel's figures, not a measurement.
Platform limits to know before buying
- CPU requirement: PMem 200 requires 3rd-gen Xeon Scalable processors. Lenovo's product guide says the same.
- Module limits: The guide lists a maximum of six PMem modules per processor on 4-socket platforms, and eight on 2-socket Ice Lake platforms, one per memory channel. Supermicro's guide says all PMem modules on a board must be the same size.
- Mode support: Per Intel's product brief, two-socket platforms support both Memory Mode and App Direct. Four-socket platforms support App Direct only.
- Operating systems: Supermicro's guide lists support for Windows Server 2019 among the OSes, alongside several Linux distributions and VMware ESXi.
- Availability: Intel lists PMem 200 as discontinued, so you are buying used or leftover stock.
A separate experiment, not a correction
A different Reddit user reportedly tested K3 on PMem 200 with an RTX 3090, 256GB of DDR4 and eight 256GB modules. That post compared Memory Mode and App Direct and found Memory Mode faster for K3 (about 1.05 versus 0.62 tokens per second generation). It used a different quantization and context size. I couldn't open that post myself, so this comes from the research notes. It cannot be compared with the CPU-only 0.85 and 0.75 figures. It does show that mode, DRAM and quantization can move results a lot.
Bottom line
This is a feasibility stunt, not a recommendation. Sub-one-token-per-second generation suits overnight batch jobs. One commenter estimated a 50k-token task at about 13 hours at roughly 1 token per second. Another noted that batch jobs care less about speed.
Cheap GPU-free capacity is clever, but the heat, the discontinued parts and the tuning work are real costs. Watch the follow-up benchmarks, especially once DRAM tiering and GPUs are added. They will show whether Optane becomes a practical cheap-capacity layer or stays a fascinating dead end.
Correction (November 8, 2026): The platform-limits section originally said the guide lists a maximum of six PMem modules per processor. That figure applies to 4-socket platforms; Lenovo's product guide lists a maximum of eight PMem modules per processor on 2-socket Ice Lake platforms, which is the class of CPU in this build. The text now states both limits.