In the age of GDDR5, HBM memory was touted as the solution to all problems. Back then, GDDR was reaching its performance limit and reaching a tipping point where the increase in consumption was not worth the increase in performance, but eventually GDDR6 memory arrived and all plans to follow that included memory. HBM and its HBM2 and HBM2E evolutions have been truncated, at least in the consumer segment.
HBM memory is not worth it … for the cost
HBM memory chips have much lower power consumption and higher bandwidth than GDDR chips, also offering much higher density and taking up less space. Everything seems to be pros and of course if we don’t factor in the cons, there shouldn’t be any reason why we don’t use HBM memory in all graphics cards today, but the reality is that the biggest downside to this memory is that more weighed in at the end.
The HBM memory design concept stacks multiple layers on top of each other in a vertical 3D design, and these layers are connected to each other using TSV (Trough Silicon Vias) which serves to connect the DRAM chips together. to others and, in turn, to connect to the GPU with the interposer. This has a fairly high production cost and a very noticeable manufacturing complication, which made the overall cost of manufacturing HBM memory several integers higher than that of manufacturing GDDR memory.
If to this we add that the performance obtained by this type of memory has been greatly exceeded by the new generations of GDDR6 and GDDR6X VRAM memory, the only reason we have left to continue to promote it as graphics memory is simply its enormous density. , because with a lower number of chips allows a much larger amount of graphics memory.
And what types of products require large amounts of graphics memory? Knowing that we currently manage quantities of VRAM ranging from 8 to 24 GB in high-end graphics cards with GDDR6X memory (the latter in the RTX 3090 without going further, although it is true that the most normal are capacities 8 to 12 GB) for the mainstream market, in fact, only the professional and professional market requires larger amounts of VRAM.
In this area, where costs are usually not an issue and everything is focused on performance and capacity, this is where HBM memory has found its market niche, since in the mainstream market like us As we have said, it has been greatly exceeded in terms of performance and in this market its high densities are no longer necessary.
And these are the reasons HBM was not standardized as a graphics memory: it is very expensive to manufacture and not worth it for the mainstream market because in terms of performance it has again been overtaken by GDDR, therefore its The only advantage is a higher density and therefore its market niche has remained for professional circles. Of course, manufacturers are continuing its development and are already tinkering with its implementation also as system memory, so of course it is far from forgotten.