News

NVIDIA RTX 4090 Ti engineering sample surfaces on Chinese marketplace with 14% higher memory bandwidth

Published: | Posted By:

An engineering sample of the unreleased NVIDIA GeForce RTX 4090 Ti has surfaced on a Chinese secondhand marketplace, offering a rare glimpse at the hardware NVIDIA ultimately decided not to mass-produce. The appearance of this prototype provides concrete data on the ceiling of the Ada Lovelace architecture, revealing a GPU that significantly pushes the boundaries of the retail flagship.

The device appeared on the Xianyu marketplace, accompanied by photos of a bare PCB and a GPU-Z readout. According to the listing, the card is identified as a "GeForce RTX 4090 Ti" with device ID 10DE 2682, a distinct departure from the 10DE 2684 ID used by the retail RTX 4090. While GPU-Z did not have a predefined entry for the device's process or die size, the technical specifications revealed a substantial increase in raw compute power.

Hardware Specifications

The prototype features 18,176 CUDA cores, representing an 11 percent increase over the 16,384 cores found in the standard RTX 4090. This configuration utilizes 142 of the 144 streaming multiprocessors available on the AD102 die, placing the card's compute capabilities on par with the professional RTX 6000 Ada Generation. This core count is supported by 192 ROPs and 568 TMUs, alongside a boost clock of 2,580 MHz and a base clock of 2,340 MHz, both of which sit slightly above the retail 4090's 2,520 MHz boost and 2,235 MHz base clocks.

This increase in compute is matched by a significant upgrade to the memory subsystem. While the card retains 24GB of Micron GDDR6X VRAM on a 384-bit bus, the memory chips are clocked at 24Gbps—up from the 21Gbps found on the retail model. This shift results in a memory bandwidth of 1,152GB/s, a 14.3 percent increase over the 1,008GB/s provided by the standard RTX 4090.

The physical PCB further distinguishes the prototype from consumer products, featuring one HDMI and three DisplayPort outputs and a rear sticker noting that EMC certification was still pending. The seller indicated that the card operates without stability issues during testing, though it requires a non-public, internal NVIDIA driver to function. This is corroborated by the GPU-Z readout, which lists the driver signature as "Beta."

These leaked specifications align closely with early rumors from January 2023, which predicted a 600W total board power and the same 18,176-core configuration. This discovery follows a previous incident in November 2024 where a four-slot prototype was leaked, though that specific unit was later identified as a standard RTX 4090.

The existence of this engineering sample confirms that NVIDIA developed a "true" 4090 Ti that maximized the AD102 silicon, yet the company chose not to bring it to market. Because the card is from a previous generation and requires proprietary drivers, it remains a curiosity for enthusiasts rather than a viable product. The focus now shifts toward future iterations of the architecture, with some industry expectations pointing toward the eventual launch of the RTX 6000 series.

Read more from Ninjalane
Hardware Reviews, News and More...