Every main graphics product out proper now makes use of both GDDR6, GDDR6X, or HBM3 reminiscence, and the latter possibility solely actually exists for super-computing chips like NVIDIA’s Hopper H100, AMD’s Instinct MI250X, and Intel’s Data Center GPU Max. GDDR6 and GDDR6X are the requirements of our day, and so they’re actually the identical factor—GDDR6X simply encodes the info in another way to realize greater effectivity.
GDDR7 is fascinating, as a result of in a selected approach, it is sort-of like a compromise between GDDR6 and GDDR6X. Rather than the complicated PAM4 signaling utilized by GDDR6X, it makes use of PAM3, which is definitely fairly totally different regardless of the same title. PAM3 signaling permits GDDR7 to supply improved switch charges over GDDR6X with out requiring the unique sign upkeep {hardware} required by that reminiscence tech.
According to Micron’s competitor Samsung, GDDR7 reminiscence may hit switch charges as excessive as 36 Gbps per pin—however we doubt it will likely be that quick at launch. For perspective, the quickest GDDR6 reminiscence will get as much as 24 Gbps, though so far as we’re conscious that pace class hasn’t shipped in any end-user merchandise. The hottest-clocked GPU reminiscence to this point is the 22.5 Gbps RAM on the GeForce RTX 4080.
While GDDR7 could also be obtainable within the first half of subsequent yr, we most likely will not see it in graphics playing cards till the second half of subsequent yr, which signifies that—if the rumors are correct—Intel’s Battlemage will miss out. In current years, NVIDIA’s often been the primary to debut new reminiscence applied sciences, so count on to see GDDR7 RAM on regardless of the successor to Ada Lovelace finally ends up being referred to as in 2025.
