HBM vs. GDDR

THE BATTLE for VRAM dominance needs settling this year. With both Nvidia and AMD utilizing GDDR5X and HBM 2.0 in their product lineups, which is best, and what will we see on the GPUs of tomorrow?

Each has its advantages and disadvantages. HBM 2.0 takes up less PCB space than typical GDDR5X stacks, runs at 1.2V, has an incredibly low latency, puts out an astonishing amount of bandwidth (242GB/s per stack versus GDDR5/X’s 44GB/s), and, more importantly, features Pseudo Channel mode, which splits a memory channel in two, with one pseudo channel dedicated to memory reads, and the other to writes. Think of it as multithreading, but for memory, and solely dedicated to read and writes, as opposed to processes.

That said, it has issues: It’s expensive, in short supply (with only Samsung and SK Hynix currently producing it), and can be difficult to design new GPUs around, especially if the GPU in question is used across a range of cards, at different price points, such as the GP104, which covers the GTX 1070, 1070 Ti, and 1080.

GDDR5X did address some of the discrepancies between the two standards, especially from a bandwidth perspective, and only then when competing alongside more consumer-friendly dual-stack Vega GPUs. However, latency is still an issue, and it lacks any form of multi-channel support. Which is why Nvidia currently incorporates HBM 2.0 into its

high-end Tesla cards available today.

Fortunately, Micron has announced that it’s mass-manufacturing GDDR6 modules, to be utilized by highend GPUs early next year. GDDR6 garners a 4Gb/s bump in memory transfer speed, and, more importantly, also features dual-channel support, which could help combat HBM’s Pseudo Channel mode.

For the time being, it seems we’ll still be stuck in this memory hierarchy, with GDDR5 and 5X featuring on more budget-friendly cards, GDDR6 on the high end, and HBM typically utilized on Vega and more workstation-oriented variants instead.