HP’s ZBook Ultra G1a 14 holds the Best for Large Models spot on our Best Laptops for Local AI leaderboard because its 128GB of unified memory, 96GB of it assignable to the GPU, loaded models no discrete-GPU laptop could touch. The new HP ZBook Ultra G3a 16 raises that pool to 192GB with up to 160GB for the GPU, moves to a 16-inch chassis, and nearly doubles the sustained power budget. We have a pre-production unit in the lab, and while benchmarks wait for shipping hardware next month, the parts list alone suggests the top of that leaderboard is about to get shuffled.
HP announced the ZBook Ultra G3a 16 today, with availability in October and pricing to follow closer to ship, and the unit on our bench carries a pre-production label and a clean Windows install. What we can do now is walk through the hardware, show what’s under the bottom cover, and lay out what we plan to measure once shipping systems arrive.
192GB of Unified Memory, 160GB for the GPU
One number defines this system: memory. Our unit pairs the AMD Ryzen AI Max+ PRO 495, the top of the Strix Halo PRO stack, with 16 cores, 32 threads, and a 5.2GHz boost clock, with 192GB of LPDDR5X-8533 soldered around the CPU. Up to 160GB of that pool can be dedicated to the integrated Radeon 8065S GPU, and the Ryzen AI NPU adds 55 TOPS on top. HP says it co-engineered the system with AMD, and its claim is that the configuration runs LLMs up to 300B parameters locally, with room for multiple agents alongside them; HP also calls it the most capable local agentic laptop, a ranking based on its own August comparison of GPU memory capacity across competing laptops.
For context, the G1a 14 shipped with 128GB and let us assign up to 96GB to its Radeon 8060S. That was enough to load DeepSeek-R1 70B and QwQ 32B in LM Studio and Ollama, and to run Gemma 3 27B at 8.96 tokens per second, which is why it took the large-model category even though its Procyon numbers trailed the RTX PRO laptops we’ve tested. The G3a adds 64GB of GPU-addressable memory to that base design. It’s also the largest GPU memory allocation we’ve seen on any Ryzen AI Max system so far, laptop or desktop, and it lands in the same silicon family as the Ryzen AI Halo desktop we put up against the DGX-vonk in juli.
Memory capacity decides which models load, but it doesn’t decide how fast they run. On the G1a, unified memory bought model size, and the 55W power budget set the pace, and the G3a’s GPU alone doesn’t change that equation.
A Thermal Redesign From 55W to 100W
HP’s new thermal design in the G3a 16 sustains up to 100W, a massive jump vs. prior Zbooks. A single heat spreader plate covers the SoC and memory packages, with two Delta fans on either side exhausting through a wide radiator fed by copper heat pipes along the rear edge.
A 96Wh user-replaceable battery fills the lower third of the chassis, with an HP-rated runtime of up to 14 hours and a 50% charge in 30 minutes on the 180W GaN adapter. Storage is a single PCIe Gen5 x4 M.2 slot, populated in our unit with a 2TB SK hynix PCB01 and configurable up to 8TB. One slot in a 16-inch workstation is worth knowing about if you were counting on a second drive for datasets, though 8TB of Gen5 in one bay covers most of that ground. HP also notes the keyboard is designed to be swapped by IT without a full disassembly, part of a serviceability push across its portable line.
Bouw en Ontwerp
Design-wise, the G3a 16 is a Z through and through: an Eclipse Gray aluminum shell, 17.9mm thick, starting at 4.2 pounds, with the same understated finish as the Fury and Studio lines. The G1a 14 came in at 3.3 pounds with less than half the sustained power and a smaller screen; the G3a adds two inches of display, a larger cooling system, and a bigger battery for under a pound extra.
The deck carries a spill-resistant backlit keyboard with HP’s DuraKeys coating, a large Precision touchpad, and speaker grilles flanking the keys for the four-speaker Poly Studio audio system. Our unit has the 16-inch WQUXGA panel, a 3840×2400 IPS at 120Hz and 500 nits with HP DreamColor calibration and 100% DCI-P3 coverage. HP also lists a 2.8K OLED touch option at 500 nits, a 2.5K panel with 100% Adobe RGB, and a lower-power WUXGA panel for buyers who want battery life over pixels.
The left edge carries HDMI 2.1, Thunderbolt 4 port at 40Gbps, 10Gbps USB-C port with Power Delivery and DisplayPort 1.4, and a headphone jack. The right edge adds a second Thunderbolt 4 port, 10Gbps USB-A port, and a nano security lock slot. Wireless is MediaTek Wi-Fi 7 with Bluetooth 6.0, and HP offers 5G WWAN, including its HP Go connectivity service, as an option. The webcam is a 5MP IR unit with Windows Hello, and the OS list includes Ubuntu 26.04 LTS alongside Windows 11 Pro.
Perplexity, Revit, and HP’s Skills
Perplexity Computer ships on the G3a 16 with an MCP connector into Autodesk Revit, so a local model can work against project data on the machine without pushing proprietary files to a cloud endpoint, and users can hand it off to a cloud model when they choose. HP and Perplexity also co-developed a set of task-based Skills exclusive to HP customers. Our pre-production unit arrived with a clean Windows image, which is normal at this stage, so we’ll look at the Perplexity stack and the Revit connector on a shipping build. On the application side, HP lists support for Autodesk Revit 2027, Autodesk 3ds Max, and Chaos Enscape visualization, the design and visualization workloads the AEC buyers this system targets already run.
“Professionals need AI experiences that are powerful, reliable, and integrated into the way they work,” said Jim Nottingham, senior vice president and division president of Advanced Compute and Solutions at HP. HP positions the G3a 16 as one piece of a broader local AI portfolio that includes the ZGX Nano en ZGX Fury, which is now orderable and is in our lab for review, along with HP Boost and HP Remote System Controller. HP is also publishing a Local AI Value Calculator on HP.com to model the cost of moving AI workloads from the cloud to Z workstations. For the desk-side end of that range, our Beste desktops voor lokale AI-ranglijst ranks the ZGX Nano and its GB10 and Strix Halo peers on the same tests we’ll run here.
HP ZBook Ultra G3a 16 Specifications
| Specificaties | HP ZBook Ultra G3a 16 |
|---|---|
| Platformoverzicht | |
| Gegevensverwerker | AMD Ryzen AI Max+ PRO 495 (16C/32T, up to 5.2GHz, 64MB L3) as tested Ryzen AI Max PRO 490, 485, and 480 configurations available |
| grafiek | AMD Radeon 8065S integrated (8050S and 8040S on lower SKUs) |
| NPU | AMD Ryzen AI, 55 TOPS (50 TOPS on lower SKUs) |
| Geheugen | Up to 192GB LPDDR5X-8533 unified Up to 160GB dedicated to the GPU |
| Opslag | 1 x M.2 PCIe Gen5 x4 NVMe, up to 8TB 2TB SK hynix PCB01 as tested |
| Scherm | 16-inch WQUXGA 3840×2400 IPS, 120Hz, 500 nits, 100% DCI-P3, HP DreamColor (as tested) 16-inch 2.8K 2880×1800 OLED touch, up to 120Hz, 500 nits, 100% DCI-P3 16-inch 2.5K 2560×1600, up to 120Hz, 400 nits, 100% Adobe RGB 16-inch WUXGA 1920×1200, 400 nits, low power, 100% sRGB |
| Connectiviteit | |
| Ports (left) | 1 x HDMI 2.1 1 x Thunderbolt 4 (USB-C, 40Gbps, PD, DisplayPort 2.1) 1 x USB-C 10Gbps (PD, DisplayPort 1.4) 1 x hoofdtelefoon / microfoon combo |
| Ports (right) | 1 x Thunderbolt 4 (USB-C, 40Gbps, PD, DisplayPort 2.1) 1 x USB-A 10 Gbps 1 x nano security lock slot |
| Wireless | MediaTek Wi-Fi 7 MT7925 (2×2) with Bluetooth 6.0 Optional 5G WWAN (HP R18, HP Go) and LTE LPWAN |
| Camera en audio | 5MP IR AI-camera 4 stereo speakers with discrete amplifiers, dual-array microphones (Poly Studio) |
| Kracht en fysiek | |
| Aanhoudende kracht | Up to 100W (G1a: 55W) |
| Accu | 96Wh 6-cell, user replaceable Up to 14 hours (HP rating), 50% charge in 30 minutes |
| Power Adapter | 180W or 140W USB-C GaN |
| Afmetingen | 14.02 x 9.45 x 0.70 inch (35.6 x 24.0 x 1.79 cm) |
| Gewicht | Starting at 4.2 lb (1.9kg) |
| Besturingssystemen | Windows 11 Pro, Windows 11 Home, FreeDOS, Ubuntu 26.04 LTS |
| Beschikbaarheid | October 2026, pricing to be announced |
What We’ll Test Next
The G1a 14 proved the unified-memory idea in a laptop: models that no discrete-GPU notebook could load ran locally on a 3.3-pound machine. Where it ran short was generation speed, with large models producing single-digit tokens per second inside a 55W envelope. The G3a 16 is aimed squarely at the power users who need more than that ceiling. The 160GB GPU allocation extends the capacity lead that earned the G1a its category, and the 100W thermal budget is the first chance for a Strix Halo laptop to close some of the throughput gap to the RTX PRO machines on the leaderboard. How much of that gap closes will be uncovered in the full review.
When shipping systems arrive, likely in October, we’ll run the same UL Procyon AI Text Generation suite that ranks every laptop on the leaderboard, then push into LM Studio and Ollama with the 70B-class and larger models the memory pool invites, and measure sustained clocks, power, and battery. If the 100W budget holds up over a long generation run, this system has a case for the overall spot on top of the large-model category the G1a already holds.
Amazon