Report: Nvidia May Have Canceled High-VRAM RTX 3070, 3080 Cards
Not long after the RTX 3080 and 3090 debuted, rumors surfaced of high VRAM variants that would supposedly appear in-market by December. The reasoning was obvious — AMD is expected to launch GPUs with more VRAM than Nvidia’s equivalents. This is a known strategy for Team Red, and they’ve deployed it consistently with the Polaris GPU family against the GTX 1050 and 1060.
One reason a high VRAM version of an Ampere GPU is so interesting is that it opens up a realm of additional possibilities in AI work. According to our sister site PCMag, this plan has been at least temporarily scuttled. They state: “Two independent sources have confirmed the new RTX cards are no longer coming to market.”
Nvidia, however, isn’t the only company to blame for this production problem. Micron’s yields on GDDR6X are also reportedly poor. The difference between GDDR6 and GDDR6X is shown below:
GDDR6X is faster-clocked GDDR6, with higher RAM bandwidth per pin and per placement, and a much smaller power consumption benefit than the jump from GDDR5 to GDDR5X (Presumably pJp stands for picoJoules per bit). With GDDR5X, Micron cut what I’m assuming is per-bit power by 11.12 percent. With GDDR6X, the company only managed to reduce power by 3.5 percent.
The major innovations of GDDR6X are its new encoding scheme and the long-term potential to double bandwidth per pin compared with existing GDDR6, but these improvements come with their own difficulties as far as hitting effective yield. Up until now, we’ve focused on the Nvidia side of the equation, but Micron may be playing a part here as well. If yields on the RAM chips are low, and each RTX 3080 / 3090 high-VRAM GPU needs twice as much, that would effectively slash Nvidia’s already-limited ability to get cards into market.
I would go so far as to argue that it makes more sense that this is a VRAM shortage issue than an RTX GPU issue. We don’t have any reason to believe the qualification issues are on Nvidia’s side of things. Dividing a limited number of Ampere GPUs between SKUs will make it harder to find parts. But needing to divert twice as much VRAM to each card could choke production even further if demand is higher for those cards. While we have no insight into the particulars of the situation, it could easily be playing a part.
If rumors are true, AMD will dodge this shortage by relying on ordinary GDDR6 at lower total bandwidth compared with cards like the RTX 3090. This is not necessarily a bad thing — higher memory bandwidth is only useful to a GPU if the card is bandwidth-bound in the first place.
It wouldn’t surprise me if these cards eventually do become available, once the production-line issues are solved. A high VRAM Ampere would be a lovely part to own for certain workloads.