RTX 5090 Reportedly Has 64GB/96GB Versions, Repair Technician Claims Confirmation

A "Confirmation" on the Repair Bench: Large VRAM Versions of RTX 5090 Surface

A post on Reddit from the GPU repair community has drawn attention in hardware circles: a US GPU repair technician claims that the RTX 5090 has 64GB and 96GB VRAM versions, and that this has been "confirmed." The post title directly uses the word "confirmed," but the information around it is sparse—no teardown photos, no GPU-Z screenshots, no VRAM chip model numbers, just a judgment from the repair front line.

This is exactly why such claims deserve closer scrutiny. Repair technicians handle real returned cards, so their statements carry some weight; but there is a long distance between "having seen" and "confirming the existence of a mass-production version."

Why the Repair Technician's Perspective Is Worth Hearing

The GPU repair industry occupies a unique position in the hardware information chain. Cards sent in for repair are often already opened or damaged, allowing repairers to directly see PCB layout, VRAM solder pads, and chip markings—information that mainstream review media cannot access. For previous generations of graphics cards, changes in VRAM capacity and chip batches first leaked from repair forums and return channels before being followed up by official media.

However, the limitations of the repair channel are equally obvious. What technicians see is "samples or special batches that have existed," and they cannot determine whether it is an engineering sample, an OEM custom variant, an AI server derivative model, or a product that will eventually enter the retail market. The so-called "confirmed" is more accurately understood as "such hardware was indeed encountered in a repair scenario," rather than an official product line confirmation.

Engineering Sample (ES) refers to test boards produced by chip manufacturers before mass production of a new product, used to verify circuit design, cooling solutions, and driver compatibility. Such samples usually do not carry official model markings, may have different VRAM configurations from the final retail version, and are subject to confidentiality agreements, rarely entering public channels. The "non-standard" hardware occasionally encountered through repair channels is likely such samples or early trial production batches; their existence only indicates that the manufacturer once evaluated that configuration, and does not mean it will be released as an official SKU. Therefore, the repair technician's "confirmation" belongs, in terms of information hierarchy, to evidence of hardware existence, not evidence of product line planning.

What 64GB and 96GB Mean

As a consumer flagship, the RTX 5090's VRAM configuration has always been a focus of discussion. If the numbers 64GB and 96GB hold true, they clearly point to something other than gamers.

• 64GB class: Close to the VRAM range of professional workstation graphics cards, suitable for large model inference, VRAM offloading in multi-GPU parallel training, 8K video, and large-scale 3D rendering scenarios.
• 96GB class: Already within the capacity range of data center accelerator cards; a single card can carry larger parameter model shards, reducing dependence on multi-GPU interconnect.

These two capacities show a clear gap from the positioning of the consumer RTX 5090. A more reasonable explanation is that they belong to derivative models based on the same GPU core—variants for professional visualization, AI workstations, or the server market, rather than retail gaming cards.

Matching VRAM bus width with chip density is key to judging the feasibility of large VRAM versions. A GPU's memory controller connects to VRAM chips through a fixed bus width (e.g., 384-bit or 512-bit), and total capacity equals the capacity of a single chip multiplied by the number of chips. To achieve 96GB without changing the bus width, the capacity of a single chip would need to reach 24Gb (3GB) or even higher, while current mass production of high-density GDDR7 chips is limited. Another path is to add more VRAM solder pads and expand the bus width, but this would change the PCB layer count, power phase count, and cooling design—essentially a new board. This is exactly why the claim of "simply adding VRAM" does not hold up engineering-wise.

Credibility and Doubts About the Claim

A single-source "confirmation" should be treated with caution. Currently, this claim lacks cross-verifiable physical evidence: no VRAM chip brand and model, no PCB silkscreen, no GPU-Z or driver identification information, and no independent corroboration from a second repair channel.

Another doubt is the matching relationship between capacity and VRAM bus width. 96GB would require extremely high chip density or more VRAM solder pads, which would significantly change the PCB design and power layout, differing greatly from the existing RTX 5090 reference design. If such a product did exist, it would more likely appear as a completely different board design rather than a simple "VRAM-upgraded" version.

Some users in the Reddit discussion also pointed out that what the repair technician encountered might be engineering samples or early test boards; such hardware leaking before an official product launch is not uncommon, but it cannot be equated with a retail SKU.

GPU-Z and driver identification information are common means of verifying graphics card specifications: GPU-Z can read the VRAM capacity, chip vendor, and subvendor ID recorded in the BIOS, while the driver panel can reflect whether the device is properly recognized by official drivers. For non-retail versions, this information may show as unknown or anomalous values, but it can still serve as circumstantial evidence of capacity authenticity. The lack of such screenshots means that misreading, card modification, or BIOS forgery cannot currently be ruled out—this is also the technical reason why single-source claims are difficult to accept.

Potential Market Impact

If large VRAM versions ultimately materialize, the most affected area would not be the gaming market, but AI inference and professional workstations. Currently, large VRAM accelerator cards are expensive; if consumer-grade cores spawn 64GB-class products, they would create new options for local large model deployment, small training clusters, and similar scenarios.

For ordinary gamers, this claim has limited practical significance. The VRAM configuration of the retail RTX 5090 will not change because of what repair channels say; what is truly worth watching is whether the official company will make corresponding arrangements in its professional product line.

Until more physical evidence emerges, this claim should be regarded as an "observation from the repair front line," not a product launch preview.

One-line Summary

Repair technician claims RTX 5090 has 64GB/96GB versions, but lacks physical evidence; should be seen as an observation, not a launch preview.

Paragraph Summary

A US GPU repair technician on Reddit claims to have confirmed that the RTX 5090 has 64GB and 96GB VRAM versions, drawing attention in hardware circles. The article notes that repair channels can access real returned cards and engineering samples and thus have some informational value, but "having seen hardware" does not equal "confirming a mass-production retail SKU." The 64GB/96GB capacities far exceed consumer positioning and more likely point to professional workstation or AI accelerator derivative models, and 96GB would require changes to PCB and power layout. Due to the lack of cross-verifiable physical evidence such as teardown photos, chip models, and GPU-Z screenshots, the claim should be regarded as a repair front-line observation, not a product launch preview.