In the current battle over GPU memory speeds between AMD and Nvidia, Micron and Nvidia scored a decisive victory. Together the companies made GDDR6X, which tops out around 21Gb/s. That memory found its way to Nvidia’s high-end GPUs, named the RTX 3080 and 3090. AMD, meanwhile, used slower 16Gb/s modules on its RX 6900 XT, but then bumped that up to 18Gb/s for the upgraded RX 6950 XT. As you can see, there’s a pretty clear delta there, although AMD’s offerings were no slouches thanks to Infinity Cache. Now Samsung has one-upped Micron, and has developed the world’s first 24Gb/s GDDR6. Note it’s not applying an X to the RAM’s branding, so it’s five percent less cool-sounding.
Samsung made the announcement in a press release. It says it’s designed specifically for “next-generation, high-end graphics.” Interestingly, it didn’t name any customers, but it seems unlikely it would have unveiled it without having at least one big customer already signed. Samsung says it was able to hit these speeds thanks to the combination of a new circuit design and advanced insulating material. Unlike its competitor, GDDR6X, GDDR6 is JEDEC-compliant. Samsung used its third generation 10nm process along with extreme ultraviolet (EUV) technology in the manufacturing process.
Samsung is already sampling the new chips in 16Gb capacity to unnamed companies. It says its customers will being validating them in the near future. This timing is intriguing since both AMD and Nvidia will launch next-gen GPUs soon as well. It seems like a no-brainer that AMD would adopt it, but it’s unknown if Micron and Nvidia have something like GDDR7 in the works. Either way, if Samsung’s offering is faster than whatever Nvidia is testing, it would likely switch camps to obtain it. Previous rumors indicated Nvidia had GDDR7 at 24Gb/s for its 40-series, but that’s still just a rumor.
The advanced insulating material Samung is using is a High-K Metal Gate (HKMG), which prevents current leakage. This allows the IC to be more efficient, and operate at lower temperatures. Samsung says the new modules support 1.1V of power, as opposed to the industry standard of 1.35/125V. It using a new dynamic power switching technology to switch back and forth when necessary. Overall, this allows for a 20 percent reduction in power consumption compared to its previous 18Gb/s modules. It’s also 30 percent faster, which allows for a maximum of 1.1TB/s of bandwidth.
This latest news from Samsung will certainly up the ante in the upcoming GPU wars. For this generation of cards (RDND2 vs. Ampere) AMD lagged behind Nvidia on the spec sheet, but largely made up the difference with its Infinity Cache. Nvidia is rumored to be taking a page from this playbook for its 40-series Ada Lovelace architecture. It’s been speculated Nvidia will crank up L2 by 16x over Ampere. For its part, AMD might be doubling down on Infinity Cache, using 3D-stacking to literally double its size. That amount of cache riding alongside (or on top of) super-fast GDDR6 memory could be a potent combination indeed. For now, we must wait for the other shoe to drop. Will the first customer be AMD, Nvidia, or Intel?