What "support" means for the new NVIDIA entries
Software recognition is not a product announcement. GPU-Z can now identify these devices. That does not mean NVIDIA has launched or confirmed retail versions of the RTX 4090 Ti ES or the GB203-based RTX 5070.
- RTX 4090 Ti ES: VideoCardz notes that the changelog does not specify which engineering sample configurations are supported. VideoCardz also cites an earlier report of an engineering sample found in China with 18,176 CUDA cores, against 16,384 on the released RTX 4090. That figure comes from the earlier report. GPU-Z's changelog does not verify it.
- RTX 5070 (GB203): VideoCardz says the entry means GPU-Z can recognize the configuration. It does not establish retail availability or any change to the card's specifications. TechPowerUp calls this a rare variant. It sits alongside the original GB205-based RTX 5070.
- RTX Spark: The changelog says support was added, with some memory details still missing. Treat the readout as incomplete until a later build says otherwise.
Why the RTX Spark memory caveat matters
NVIDIA's specifications page lists three RTX Spark configurations. Two are laptop parts and one is a desktop part. The laptop parts are a 6,144-core GPU with a 20-core CPU, and a 5,120-core GPU with an 18-core CPU. The desktop part has a 6,144-core GPU and a 20-core CPU. NVIDIA lists unified LPDDR5X memory, up to 128 GB on the 6,144-core parts and up to 32 GB on the 5,120-core laptop part. It also lists Windows 11 support and a CUDA capability of 12.1.
These are NVIDIA's platform specs, not something GPU-Z reports. They do show why memory fields are the hard part. A unified-memory design doesn't fit the traditional model of a card with its own VRAM. That is my own analysis, not a statement from TechPowerUp.
Sensor and monitoring changes
These changes will affect more readers than the new hardware entries.
- Sensor statistics tooltip: Hovering over a sensor's number display now shows value statistics.
- Cleaner min/max/average: These readings now exclude invalid sensor samples. A single bad sample used to be able to distort a log. That is less likely now, though it doesn't guarantee every reading is accurate.
- Integrated GPU memory: The changelog says integrated-graphics machines show "iGPU" followed by the committed memory share size. The example format is "iGPU (64 GB)". This is a capacity readout, not a claim that all that RAM is dedicated video memory.
- Unknown NVIDIA power rails: They appear as "Rail xxx Power Draw" sensors. The feature is disabled by default. You have to turn them on in the sensors list to see them.
- Unknown GPUs: For graphics cards with unknown GPUs, such as fake ones, GPU-Z now detects DirectX API version and feature-level support from the software instead of showing "unknown." This could help you spot a counterfeit card. The readout describes what the system reports. It doesn't prove what silicon is on the board.
- Unified naming: The hotspot sensor is now called "Hot Spot" on both AMD and NVIDIA cards.
AMD and Intel additions
- AMD hotspot readings: GPU hotspot temperature reading is added for the AMD R9 200/300, Fury and RX 400/500 series (Polaris, Lexa, Hawaii, Fiji, Bonaire, Tonga). Owners of older Radeon cards that are still running can now see hotspot temperatures.
- AMD integrated GPUs: Clock speed and temperature monitoring was added for certain AMD iGPU models. The changelog doesn't list which models.
- Other new entries: AMD's BC-250 mining GPU is now recognized. On the Intel side, support covers Raptor Lake-H Refresh and Iris Pro P555 (E3-1558L v5).
Under-the-hood fixes
- CUDA capability support is extended to 6.2, 7.x, 8.x, 9.0, 10.x and 12.x, with several new CUDA attributes.
- Typos and small errors in the Advanced → CUDA view are fixed.
- More DXVA GUIDs from AMD and NVIDIA now have names, and DXVA detection for 8K support is fixed.
- OpenGL detection that sometimes failed is fixed.
- If an NVIDIA BIOS can't be read at all, the Advanced tab no longer shows a corrupted "NVIDIA BIOS" section.
Practical takeaways
- Update if you monitor older Radeons, laptops with integrated graphics, or RTX Spark hardware. Those users gain the most.
- Enable the new rail sensors manually if you want extra NVIDIA power data. They are off by default.
- Don't read too much into the unreleased entries. Recognition of an ES or an unusual SKU shows the tool has caught up with hardware that exists somewhere. It doesn't confirm a retail product, pricing or final specs.
- Expect gaps on RTX Spark until TechPowerUp finishes the memory reporting.
GPU-Z releases usually track new silicon quickly, so an engineering-sample entry says little about launch timing. If NVIDIA does plan a retail RTX 4090 Ti or a second RTX 5070 variant, GPU-Z recognizing them now doesn't confirm it.
References
- GPU-Z 2.72 adds support for NVIDIA RTX 4090 Ti ES, RTX 5070 GB203 and RTX Spark - VideoCardz.com VideoCardz.com · 2026-11-19T20:49:44+00:00
- NVIDIA Home nvidia.com
- TechPowerUp GPU-Z v2.72.0 Released techpowerup.com