What was tested
StorageReview's unit pairs the Core Ultra 9 386H with 64GB of LPDDR5X-8533 on an LPCAMM2 module, a 2TB SK hynix Gen5 SSD, the RTX PRO 2000 Blackwell, and the 4K IPS display. Intel lists the 386H as 16 cores (4P + 8E + 4LPE) and 16 threads, with a P-core range of 2.1 to 4.9GHz and 18MB of cache.
Intel's own spec sheet agrees. Intel's 386H is paired here with NVIDIA RTX PRO 2000 Blackwell graphics in a machine pitched at developers, CAD specialists and creative professionals. Intel also gives the part a 25W base power and an 80W maximum turbo power.
The review ran on Windows 11 Pro 25H2 with the laptop plugged in. Windows was set to Best performance, and Lenovo's Ultra Performance mode was enabled. StorageReview used its own scripted client suite, and it says it reruns any test whose spread exceeds 2 percent. Battery life is not covered yet; the reviewer says it will follow in a later update.
The comparison machines were:
- A Dell Pro Max 14 Premium with the older Core Ultra 9 285H and the same GPU at 47W.
- A Dell Pro Precision 5 16S with the Core Ultra X9 388H and integrated Arc Pro B390 graphics only.
- An HP OmniBook Ultra 16 built on NVIDIA's RTX Spark N1X, with 128GB of unified memory.
Where it sits in Lenovo's lineup
Lenovo's March 16, 2026 announcement put the P1 Gen 9 alongside the P14s and P16s refreshes. The announcement describes it as the most capable ThinkPad ever, with a discrete RTX PRO Blackwell GPU. The review calls the Gen 9 the Gen 8 chassis with new silicon. The footprint barely changed, and weight fell from 1.84kg to 1.77kg.
Key platform changes from Gen 8 to Gen 9:
- The CPU moves from Series 2 Arrow Lake-H to Series 3 Panther Lake.
- The NPU goes from 13 TOPS to 50 TOPS.
- LPCAMM2 memory capacity rises from 64GB to 96GB.
One configuration detail matters for buyers. The review says RTX PRO 2000 builds use the Core Ultra 7 356H or Core Ultra 9 386H. The X7/X9 chips with the 12-Xe-core Arc Pro B390 are integrated-graphics-only builds. You choose one path or the other.
The reviewer also noted that its module is rated for 9600 but runs at 8533.
Hardware and serviceability
- Chassis: The shell is aluminum and 9.87mm at the front edge. It carries MIL-STD-810H testing.
- Battery and charging: Lenovo lists a customer-replaceable 90Wh battery. It supports rapid charge to 80 percent in an hour from the 140W adapter.
- Ports: There are two Thunderbolt 5 ports and a modular Thunderbolt 4 port. It also has 10Gbps USB-A, HDMI 2.1 and an SD Express 7.0 reader. There is no onboard Ethernet.
- Storage: There are two M.2 2280 slots, one Gen5 x4 and one Gen4 x4, with RAID 0/1. On 5G models the modem takes the Gen4 slot, so a two-drive setup and 5G are mutually exclusive.
- Memory: The LPCAMM2 module is customer-replaceable up to 96GB.
- Display: The tested panel is a 3840×2400 IPS at 800 nits with 100 percent DCI-P3 coverage. A 3200×2000 Tandem OLED with pen support and 40 to 120Hz variable refresh is also offered, along with an entry 1920×1200 panel.
- Security and sign-in: The camera uses IR for Windows Hello. The review also lists a discrete TPM 2.0, a fingerprint reader in the power button, and BIOS tamper detection on the bottom cover.
CPU and memory results
Panther Lake's gains show up mainly in multi-threaded work and the memory system.
| Test | P1 Gen 9 (386H) | Dell Pro Max 14 (285H) |
|---|---|---|
| Cinebench 2026 multi-core | 5,083 | 4,292 |
| Cinebench 2026 single-core | 519.9 | 486.0 |
| LLVM 21.1.4 build | 309.1 s | 393.9 s |
| 7-Zip, all threads (GIPS) | 109.2 | 93.4 |
| STREAM Triad (MB/s) | 81,712 | 62,255 |
| Memory latency | 116.1 ns | 173.1 ns |
| Speedometer 3.1 | 40.6 | 41.5 |
The multi-core lead is about 18 percent. Single-thread and browser work barely moved. The reviewer says Python times were within 2 to 5 percent of the 285H in either direction. The X9 388H in the Dell Pro Precision 5 16S was faster on single-thread work, and the reviewer ties that to its higher clock.
The core-to-core latency average rose from 59.2 to 92.0 ns. The reviewer attributes that to Panther Lake's tile layout and says it showed up in nothing they ran.
Sustained performance
In a roughly 30-minute Cinebench loop, the P1 held 0.983 of its first-pass score. It averaged 3,505MHz, and the CPU reached 94.1C. The Dell Pro Precision 5 16S, with the same core layout, settled at 0.938. The review credits the P1's chassis for sustaining about 8 percent more than that machine.
Those temperatures are expected on all-core loads. The reviewer calls 94 to 95C on all-core tests the point where Intel lets Panther Lake run, and the clocks held at that ceiling are what counts.
The 8GB ceiling for local AI
The headline caveat is the GPU. NVIDIA's own comparison table lists the laptop RTX PRO 2000 at 8GB of GDDR7 and a 45 to 115W power range. The P1 runs it at 50W, and the Dell runs it at 47W.
MLPerf Client results from StorageReview:
- Llama 3.1 8B: 64.8 tok/s on the P1, against 54.7 on the Dell.
- Phi-4-mini: 111.9 tok/s on the P1, against 88.0 on the Dell.
That works out to 19 to 27 percent higher token rates than the Dell. The reviewer cautions that it could not read GPU clocks or temperatures. It cannot say how much of the gap comes from the extra 3W versus the larger chassis.
The other side of the result is capacity. StorageReview's model-load and coding-agent tests start at gpt-oss:20b, which needs about 16GB of GPU-usable memory. The RTX PRO 2000 machines measured 7.7GB, so those tiers were marked Not run. In LLM serving, only a 3B model tier fit. In the same tests the 128GB HP unit and the 64GB Arc Pro B390 Dell completed every model-load and coding-agent tier.
That is a benchmark eligibility limit, not proof that every 8GB workflow fails. The 8B models in MLPerf did run, because MLPerf sizes its context to the card. The practical reading is that the P1 suits small models and RAG-style work. The RAG test, run with a 3.8B answer model, finished fastest on the P1 for ingest.
Two more caveats:
- The Arc Pro B390 results use OpenVINO rather than llama.cpp with CUDA, so they are not a like-for-like runtime comparison.
- NVENC encoding did not start on either RTX PRO 2000 system under the workstation driver (version 597.16 on the P1). Those cells are Not run in the review.
Storage
The 2TB SK hynix PCB01 read 10,012MB/s sequentially through the file system and peaked at 756.5K IOPS on 4K random reads. It beat the older PC811 drive in the Dell Pro Max 14 on sequential work and at the tail latencies at lower queue depths. At 16 jobs the PC811 was slightly ahead at the 99th percentile on random reads.
Pricing and buying context
StorageReview reports that Lenovo.com listed the OLED version of the tested build at $7,719 as of October 24, 2026. It says the same build sold for $5,399 two days earlier. The exact tested IPS configuration was not listed with that CPU and GPU, so these prices apply to the OLED counterpart. Build-your-own configurations started at $3,734 for a Core Ultra 7 356H with the RTX PRO 2000, 16GB and 512GB. These are time-sensitive figures from the review.
Against that, the review lists the HP OmniBook Ultra 16 at $4,999.99 with 64GB and $6,229.99 with 128GB. It also cites HP's ZBook Ultra G3a 16 with up to 192GB of Ryzen AI Max unified memory, with orders open from $3,149.
Practical takeaways
- Buy it if you need an ISV-certified mobile workstation under 1.8kg with Thunderbolt 5 and serviceable memory and storage. The certification list covers Autodesk, Dassault, Siemens, PTC and others.
- Look elsewhere if your main workload is local inference above roughly 7B parameters. The reviewer calls the integrated-only X9 388H build, or a large unified-memory system, the better fit.
- Skip the upgrade if your day is mostly browsers and Office. Single-thread results did not move meaningfully, per StorageReview.
- Watch for: battery-life results in StorageReview's later update, and an NVENC-capable driver for the RTX PRO 2000 builds.