ThinkPad P14s Gen 2 Review: The Cheapest CUDA ThinkPad — Mind the Variant
Who is this laptop for?
The P14s Gen 2 is a used 14” mobile workstation — essentially a ThinkPad T14 with an NVIDIA T500 dGPU bolted in — and the cheapest way to get CUDA inside a ThinkPad chassis. It handles Stable Diffusion 1.5, GPU-assisted 7B LLMs and Whisper, in a 1.47 kg package that looks and types like every other ThinkPad.
One warning before anything else: the P14s Gen 2 exists in two versions, and only one is useful for AI. The Intel models (types 20VX/20VY) carry the NVIDIA T500. The AMD models (types 21A0/21A1) have no dedicated GPU at all — integrated Radeon graphics only, no CUDA. Listings constantly blur this. Check the type number on the bottom label before you pay; this review covers the Intel + T500 version.
🎓 Students (Budget: £420–£580)
A sensible first CUDA machine. You get real GPU acceleration for coursework — SD 1.5, small-model inference, CUDA programming exercises — in a laptop light enough to carry daily and cheap enough to not babysit. If you’ll never touch image generation, the CPU-only ThinkPad T14 Gen 3 saves you ~£100.
👨💻 ML Engineers & Data Scientists
Fine as a travel or secondary machine; frustrating as a primary. The T500 has only 384 CUDA cores and ~80 GB/s of memory bandwidth — an order of magnitude below the cards serious work wants. It accelerates development, testing and demos, not production workloads. For a primary machine, the Dell Precision 5560 with the much stronger RTX A2000 is the same idea done properly.
🏢 Small Teams & Startups
Excellent fleet material: off-lease supply is deep, ThinkPad serviceability is famous, and refurbishers sell these with 12-month warranties all day. Just be honest about what it’s for — a developer laptop with light GPU assist, not a machine you’ll fine-tune anything on.
What can it actually run?
| Task | Works? | Notes |
|---|---|---|
| GitHub Copilot / Cursor AI | ✅ Yes | API-based, any laptop works |
| Whisper transcription (local) | ✅ Yes | medium model ~3–5× realtime on GPU (estimated) |
| Ollama 7B (Llama 3, Mistral) | ⚠️ Partial offload | ~8–12 tok/s with ~60% of layers on GPU (estimated) |
| Ollama 13B | ⚠️ CPU + RAM | Q4 fits in 16 GB RAM at ~2–3 tok/s; painful |
| Stable Diffusion 1.5 | ⚠️ Slow but works | ~50–80 s per 512×512 image at 20 steps |
| Stable Diffusion XL | ❌ No | 4 GB VRAM and 384 cores — not realistic |
| ComfyUI / FLUX.1 | ❌ No | Far beyond this GPU |
| LoRA fine-tuning (small model) | ❌ No | VRAM and compute both insufficient |
Key:
- ✅ Yes — works well
- ⚠️ Possible but slow — usable with patience
- ❌ No — hardware limitation prevents this
Full Specifications
| Component | Specification |
|---|---|
| CPU | Intel Core i7-1185G7 (4C/8T) — i5-1145G7 in base configs |
| CPU Generation | Intel 11th Gen (Tiger Lake U) |
| RAM | 16 GB DDR4-3200 (16 GB soldered + 1 SO-DIMM slot, up to 48 GB) |
| Storage | 512 GB NVMe Gen 3 |
| GPU | NVIDIA T500 (Turing TU117), ~25 W |
| VRAM | 4 GB GDDR6 (64-bit, ~80 GB/s) |
| Display | 14” FHD IPS (low-power and UHD options exist) |
| Battery | 50 Wh |
| Weight | 1.47 kg |
| TDP | 28 W CPU + ~25 W GPU |
| AI Score | 56/100 |
AI Performance in Practice
Set expectations correctly: the T500 is the smallest CUDA GPU NVIDIA put in this generation of workstations — 384 cores and a 64-bit memory bus. What it delivers is not speed but capability: software that requires CUDA simply works, where an integrated-graphics ThinkPad would refuse or crawl.
Estimated from comparable T500/MX-class results (we have not benchmarked this exact unit): Stable Diffusion 1.5 produces a 512×512 image in 50–80 seconds — fine for learning and occasional assets, hopeless for iterating on prompts. Llama 3.1 8B Q4 with roughly 24 of 32 layers offloaded to the 4 GB card runs at 8–12 tok/s, a genuine 2–3× improvement over this CPU alone. Whisper medium transcribes at a few times realtime, making meeting transcription practical.
The four-core Tiger Lake CPU is this machine’s real AI ceiling — CPU-side inference is markedly slower than the 6-core T14 Gen 3, which is why the T500’s offload matters so much here.
Thermal behaviour
The T500’s 25 W barely stresses the chassis, and it holds its modest clocks indefinitely. The CPU is the throttle point: sustained all-core loads settle around 28–35 W with audible but unobtrusive fan noise. Long SD 1.5 batches warm the left palm rest without drama.
Battery life under AI load
The 50 Wh battery gives 6–8 hours of office work but roughly 1.5–2 hours of continuous GPU inference. Standard 65 W USB-C charging means any decent charger keeps up — one genuine advantage over the barrel-plug workstations.
What to Check Before Buying (Used)
Variant check — the deal-breaker. Flip the laptop: the type number on the bottom label must be 20VX or 20VY (Intel, T500 present). If it starts with 21A, it’s the AMD version with no NVIDIA GPU — a fine office laptop and a useless CUDA machine. Then verify in Device Manager or a GPU-Z screenshot that “NVIDIA T500” actually appears; a handful of Intel units were sold with Iris Xe only.
Battery health
powercfg /batteryreport — off-lease units commonly show 60–80% of the 50 Wh design capacity. Replacements are ~£50–£70 and a ten-minute job on this chassis.
GPU throttling test FurMark for 10 minutes: the T500 should sit calmly at its ~25 W limit, around 1,200–1,400 MHz, well under 75 °C. Any dramatic behaviour here means cooling problems that will bite the CPU too.
Storage health CrystalDiskInfo: 0 reallocated sectors, and treat 20,000+ Power On Hours as a price-negotiation fact. The single M.2 slot means upgrades replace rather than add.
RAM slots 16 GB is soldered; one SO-DIMM slot takes up to 32 GB more (48 GB total). Confirm the slot works — 48 GB turns this into a surprisingly capable CPU-inference machine for 13B models, slow but functional.
Model-specific issues to watch for Usual off-lease ThinkPad checklist: hinge tightness, keyboard shine, TrackPoint drift, and BIOS supervisor passwords left behind by IT departments (walk away if locked — they’re effectively unremovable). Check the display lottery too: the 400-nit low-power FHD panel is lovely; base 250-nit panels are not.
Where to Buy in the UK
The best places to find a used P14s Gen 2 in the UK:
Back Market UK — steady off-lease supply at £430–£550 with 12-month warranty. Crucially, listings usually state the type number — verify 20VX/20VY before ordering.
Tier-1 UK refurbishers (e.g. laptopsdirect refurb, ITZOO-style off-lease sellers) — often the cheapest route to i7/16 GB configs; check that the T500 is explicitly listed, not “Intel Iris Xe”.
eBay UK — biggest selection, including rare 32 GB-upgraded units. Fine with 100% feedback sellers offering returns. Ask explicitly: “Does Device Manager show NVIDIA T500?” — sellers mixing up AMD and Intel variants is the #1 problem with this model.
What to avoid: any P14s listing without a type number, the AMD 21A0/21A1 variants (for AI purposes), and the P14s Gen 1 (older 10th-gen CPU, weaker Quadro P520) masquerading in search results.
Verdict
AI Score: 56/100 — SD Ready (barely)
The P14s Gen 2 is the cheapest ticket into CUDA-on-a-ThinkPad, and that’s exactly how to think about it: a superb, light, durable developer laptop where GPU acceleration is a bonus, not the point. It runs SD 1.5 and meaningfully speeds up 7B inference, and it does so in the best 14” chassis in the business. It will never be fast.
Buy if: you want a portable, sub-£600 ThinkPad that can demonstrate everything — CUDA, SD 1.5, GPU-offloaded LLMs — for learning, development and travel.
Don’t buy if: image generation or fine-tuning is your actual workload; the T500 will disappoint you within a week. And never buy without confirming the Intel + T500 variant.
Better alternatives: the Dell Precision 5560 (£480–£680) has 6.5× the CUDA cores for one step up in price; the ThinkPad T14 Gen 3 (£320–£480) is the honest choice if you don’t need CUDA at all.