ThinkPad T14s Gen 4 AMD Review: Faster iGPU but Soldered RAM Trap
Who is this laptop for?
This is a used/refurbished business ultrabook that upgrades the ThinkPad T14 Gen 3 formula with a faster iGPU and NVMe Gen 4 storage. The Radeon 780M (RDNA 3) integrated GPU is genuinely better for local AI inference than the older 660M — but there’s a critical catch: RAM is soldered. If you buy the wrong config, you’re stuck with it.
Students (Budget: £420–£580)
The T14s Gen 4 sits in an awkward price bracket for students. At £420–£480, you can sometimes find the 16 GB config — but 16 GB of soldered RAM is a dead end for local AI. The 32 GB config at £500–£580 is the one to target. It runs Ollama 7B models faster than the Gen 3 thanks to the RDNA 3 iGPU, and 32 GB gives headroom for 13B models on CPU. If budget is tight, the ThinkPad T14 Gen 3 at £320–£480 with upgradeable RAM is a safer bet.
ML Engineers & Data Scientists
A solid ultrabook for travel and remote work. The Radeon 780M can accelerate some inference workloads via ROCm, though support is still patchy on Windows. On Linux with ROCm 6.x, you can offload some layers to the iGPU for a noticeable speedup on Ollama 7B models. Still a secondary machine — no discrete GPU means no serious training or image generation.
Small Teams & Startups
Premium build quality in a 1.22 kg package. The 32 GB config handles all API-based AI tools, runs local LLMs for testing, and the NVMe Gen 4 storage speeds up model loading. Back Market offers 12-month warranty on refurbished units. The soldered RAM means you need to buy the right config from day one — no cheap upgrades later.
What can it actually run?
| Task | Works? | Notes |
|---|---|---|
| GitHub Copilot / Cursor AI | ✅ Yes | API-based, runs perfectly |
| Whisper transcription (local) | ✅ Yes | ~4× realtime on base model |
| Ollama 7B (Llama 3, Mistral) | ⚠️ Slow | ~6–8 tok/s on CPU, estimated ~10–12 tok/s with iGPU offload (ROCm) |
| Ollama 13B | ⚠️ Tight | Needs 32 GB config. ~3–4 tok/s on CPU. |
| Stable Diffusion 1.5 | ❌ No | Integrated GPU, shared VRAM insufficient |
| Stable Diffusion XL | ❌ No | Not possible without discrete GPU |
| ComfyUI / FLUX.1 | ❌ No | Requires dedicated GPU with 6+ GB VRAM |
| LoRA fine-tuning | ❌ No | Not practical on integrated GPU |
Key:
- ✅ Yes — works well
- ⚠️ Possible but slow — usable with patience
- ❌ No — hardware limitation prevents this
Full Specifications
| Component | Specification |
|---|---|
| CPU | AMD Ryzen 7 Pro 7840U (8C/16T, Zen 4) |
| CPU Generation | AMD Zen 4 (7000 series, Phoenix) |
| RAM | 32 GB LPDDR5x-6400 (soldered, NOT upgradeable) |
| Storage | 512 GB NVMe Gen 4 (M.2 2280) |
| GPU | AMD Radeon 780M (integrated, RDNA 3) |
| VRAM | Shared system RAM (up to 4 GB allocatable) |
| Display | 14” 1920×1200 IPS, 400 nits |
| Battery | 52.5 Wh |
| Weight | 1.22 kg |
| TDP | 28W (sustained) |
| AI Score | 48/100 |
AI Performance in Practice
The Ryzen 7 Pro 7840U is a significant step up from the Gen 3’s 6650U. The 8 cores / 16 threads on Zen 4 deliver approximately 6–8 tokens per second running Llama 3.1 7B with Q4_K_M quantization through Ollama on CPU alone. That’s a 30–50% improvement over the Gen 3.
The real story is the Radeon 780M iGPU (RDNA 3). On Linux with ROCm 6.x, you can offload layers to the integrated GPU. Early benchmarks suggest 10–12 tok/s on 7B models with partial GPU offload — estimated based on comparable Ryzen 7840U results from the community. Windows ROCm support is less mature; expect CPU-only speeds there.
Whisper (base model) runs at approximately 4× realtime — faster than the Gen 3 thanks to better single-thread performance and AVX-512 support. The medium model hits roughly 1.5× realtime. NVMe Gen 4 storage means model loading is snappier — Llama 3.1 7B Q4 loads in about 3 seconds vs 5 seconds on Gen 3.
Thermal behaviour
The T14s Gen 4 is thinner and lighter than the T14 Gen 3, which means slightly more aggressive thermal management. Under sustained CPU inference, the fan is audible but not loud. The chip sustains 25–28W without heavy throttling for the first 15–20 minutes. After prolonged use (30+ minutes of continuous Ollama), expect a slight step-down to ~23W with ~10% performance reduction. Still better thermal management than most ultrabooks in this weight class.
Battery life under AI load
Normal use (browsing, coding, API calls) delivers 8–10 hours thanks to Zen 4’s excellent power efficiency. Under sustained CPU inference (Ollama), battery drains in roughly 100–140 minutes — slightly better than Gen 3 due to Zen 4’s efficiency improvements. The 52.5 Wh battery is the same size as Gen 3, but the more efficient architecture helps. Carry the 65W USB-C charger for AI work sessions.
What to Check Before Buying (Used)
RAM configuration — THE critical check RAM is soldered on the T14s Gen 4. There are no SO-DIMM slots. Whatever the listing says is what you get — forever. The 16 GB config is common and cheap, but it severely limits local AI use (7B models only, no 13B). Always verify 32 GB before buying. Check in BIOS or Task Manager → Performance → Memory. If it says 16 GB, walk away unless the price is exceptionally low.
Battery health
The T14s Gen 4 is newer than Gen 3 (2023 vs 2022), so batteries should be in better shape. Expect 75–90% of original capacity on a 1–2 year old unit. Run powercfg /batteryreport and check Full Charge Capacity. Below 40 Wh (76% health) on a unit this new suggests heavy use — negotiate the price down.
Storage speed Verify NVMe Gen 4 speeds with CrystalDiskMark. Sequential read should be 5,000+ MB/s. If you’re seeing 2,000–3,000 MB/s, the drive may have been swapped for a Gen 3 unit — not a deal-breaker but worth knowing.
BIOS and firmware Update to latest BIOS before running AI workloads. Early BIOS versions had power management issues that caused inconsistent CPU performance. Lenovo Vantage will handle this on Windows.
Model confusion The T14s Gen 4 comes in both AMD (7840U) and Intel (Core i7-1360P) variants. The AMD version is significantly better for AI workloads due to the superior iGPU and better multi-thread performance. Always verify “7840U” or “7540U” in the listing — not “1360P”.
Where to Buy in the UK
Back Market UK — Best option. Expect £450–£580 for the 32 GB config. The 16 GB config appears at £420–£500 but is not recommended for AI use. Always filter for 32 GB RAM. “Excellent” condition units look nearly new. 12-month warranty and 30-day returns.
Laptops Direct — Refurbished stock is less consistent. Prices £420–£550. Warranty typically 6 months. Double-check the exact RAM amount — many listings are the 16 GB version.
eBay UK — 32 GB configs occasionally appear at £400–£500. Verify seller has 99%+ feedback. Ask for a screenshot of Task Manager → Memory to confirm 32 GB and LPDDR5x speed. The T14s Gen 4 is newer, so there are fewer units in circulation compared to Gen 3.
What to avoid: Do not confuse with the ThinkPad T14s Gen 3 (Ryzen 6000 series, slower iGPU). Verify “Gen 4” and “7840U” or “7540U” in every listing. Also avoid the Intel variant — the Intel Iris Xe iGPU is substantially worse for AI inference than the AMD Radeon 780M.
Verdict
AI Score: 48/100 — LLM Ready
The ThinkPad T14s Gen 4 AMD is a meaningful upgrade over the Gen 3 for local AI — the Zen 4 CPU is faster, the RDNA 3 iGPU shows promise for accelerated inference, and NVMe Gen 4 speeds up model loading. At 1.22 kg, it’s one of the lightest options for running local LLMs.
The soldered RAM is the deal-breaker to watch for. The 32 GB config is excellent — 7B models run at 6–8 tok/s, 13B models are usable, and Whisper transcription flies. The 16 GB config is a trap: you get a more expensive laptop than the Gen 3 with less RAM flexibility and no upgrade path.
Buy if: You find a 32 GB config at a good price, want the lightest possible ThinkPad for local AI, and value the faster Zen 4 CPU and RDNA 3 iGPU over the Gen 3. Also a good choice if you plan to run Linux with ROCm for iGPU acceleration.
Don’t buy if: You can only find 16 GB configs — get the ThinkPad T14 Gen 3 instead (upgradeable RAM, £100+ cheaper). Or if you need image generation — no discrete GPU means no Stable Diffusion. Look at the Dell Precision 5560 (4 GB VRAM, ~£480–£680) instead.