HBM memory first appeared in the AMD R9 Fury almost a decade ago and was a constant in several desktop graphics cards until the Radeon VII, which passed with more pain than glory. Since then the gaming graphics cards and the HBM2 do not get along at all. Why is this type of memory no longer used, and how do HBM2, GDDR6, and GDDR6X compare?
Many times a technology can seem very good on paper, but end up being a huge commercial failure due to the fact that what seemed like a great idea at first brings with it a series of problems that are not seen at first glance that appear during its development. . One of these cases is HBM2 memory, which promised several years ago to replace GDDR5 memory forever. Reality? GDDR6 is the most used in gaming graphics cards. What was promised to be its replacement has been relegated to the market for servers that use so-called HPC GPUs.
Technically and on paper, if you ask any system architect they will tell you that HBM2 is a better type of memory than GDDR6. So how come there isn’t a single AMD RX 6000 or NVIDIA RTX 30 using it if it’s that good? That is the question we want to answer with this comparison, and we will do it in a way that all readers can understand.
What are HBM memories and what advantages do they provide?
HBM memory is a type of RAM memory that has become popular in recent years in graphics hardware for data centers and servers. However, when it comes to graphics cards for PCs, they have disappeared from the market. Today they are the best type of memory available for the following reasons:
- It can achieve the same bandwidths as GDDR6 and even GDDR6X .
- Its latency is lower than GDDR6 and GDDR6X . Which makes it useful for server CPUs with tens and even hundreds of cores. Especially those that are designed with specialized units for machine learning. That is, the so-called artificial intelligence.
- Its power consumption per transmitted bit is almost half compared to GDDR6 . This is done to increase the number of interconnects, which allows them to run at lower speeds and therefore lower voltage.
- Each memory chip is displayed horizontally throughout the board, in the case of the HBM2 we can have configurations of 4 or 8 chips, one on top of the other and consequently vertical . Which means that the space they end up occupying is much less.
This is how HBM memory works
To get the high bandwidth, but low power consumption you need, HBM sticks use a large number of interconnect pins . The problem with this approach is that this would mean that the size of the chip would increase to interconnect each pin with the corresponding memory chip. And not only from the graphics chip, but also from the memory ones. The solution? Well, what you do is send the wiring down and use an additional piece. The so-called Interposer, which adds additional cost and manufacturing complexity.
So the Interposer is an additional cost to the process and complicates the assembly of the new chip, increasing the potential number of failures during assembly. Not only that, but because HBM memory chips are stacked it’s not as simple as stacking one on top of the other. In order for the chip at the top of the stack to communicate with the interposer, its wiring has to go vertically through the rest of the memory chips in the stack. Which we call through silicic pathways or TSV.
To visualize it, imagine a block of flats where the water and electricity supply is on the ground floor and has to be distributed to all the floors. Well, the same thing happens in the case of HBM memory. That is why it is so expensive to manufacture and from the moment that price is an important value to take into account in the consumer market. Then you end up sacrificing the advantages of space and consumption, in exchange for lower costs.
Why is HBM memory not used in PC graphics cards?
So with the issue of price, we have to take into account that an ordinary user is going to look for the hardware that always gives him the best performance within his budget. The advantages of the HBM2 will allow you to get a graphics card that is less spacious and consumes less , but within the evaluation criteria they are not the most relevant element. What is it then? Obviously, the one that gets the most frames per second and at the highest possible resolution in your favorite games.
Imagine that we have two identical graphics cards, with the same GPU and the same performance, however, one has more expensive memory than the other. What is happening? The manufacturer cannot sell a variant of the product that is as good as the standard product for the end user at a higher price if it does not add a significant advantage to the main utility for which it is used. In other words, manufacturers have stopped using HBM2 in their gaming graphics for one reason: profit margins.
What is GDDR6 memory and how does it compare to HBM2?
The GDDR6 is the sixth generation of Dual Data Rate memory focused on its use for graphics cards. They differ from conventional RAM in that they sacrifice access latency, something that does not affect the graphics chip, in exchange for high bandwidth, which is essential when rendering graphics. In other words, we are dealing with a type of memory that is ideal for use in a GPU, but disastrous for a central processor. That is why we do not see DIMM modules with GDDR6 for our PC.
| Characteristic | HBM2 | GDDR6 | GDDR6X |
|---|---|---|---|
| chips per unit | 4 or 8 | one | one |
| capacity per chip | 2 GB (up to 16 GB per drive) | 1 or 2GB | 1 or 2GB |
| buses | 1024 bit | 32 bit | 32 bit |
| Channels | 8 (128 bits per channel) | 2 (16 bits per channel) | 2 (16 bits per channel) |
| Transfer speed | From 2Gbps to 2.4Gbps | 12 to 18Gbps | 19 to 21Gbps |
| Others | Requires Interposer | Up to 2 units sharing bus | Up to 2 units sharing bus /PAM4 encoding |
GDDR6 , compared to its predecessors, is a step forward, as each chip behaves like two independent ones, due to the fact that now each 16-bit data bus has its own address/command bus. In other words, it is dual-channel memory and therefore can respond to two requests at the same time. The HBM2, on the other hand, has the advantage of having lower latency, which makes it ideal for server processors with dozens of cores, but not for a graphics card, and it has up to 8 different memory channels, one for each chip. that we have placed on the stack.
The advantage of GDDR6 over HBM2 can be summed up in one thing, it is much cheaper and its widespread use helps this . We not only see it in PC graphics cards, but also in video game consoles. If we were to compare the bulk of product sales by memory type, then we would see that HBM’s percentage would be in single digits.
The GDDR6X, a sneak peek into the future?
With the appearance of the NVIDIA RTX 30, especially in the most powerful graphics cards within that range. We have been able to see the appearance in the market of a variant of the GDDR6 which has been baptized as GDDR6X. Its operation is based on using a type of encoding called PAM4 , where for each data pin, instead of sending values 0 or 1, values 00, 01, 10 and 11 are sent through changes in voltage.
In other words, it is a way to increase bandwidth without increasing the clock speed to alarmingly high levels and thus lower consumption. This has enabled speeds of up to 21 Gbps and enabling total transfer speeds previously only possible with a 4-stack or 4096-bit HBM configuration. That if, with some really high consumption . The story is again the same. Although GDDR6X is less efficient than any other HBM memory , but again it is much cheaper. Which has caused HBM to say goodbye to high-end GPUs.
The new PCIe Gen 5 connector is the nail in the coffin
Before we have told you that one of the strengths of HBM memory compared to GDDR6 memory is its lower consumption. Which is an advantage when you are getting dangerously close to the power limits of the power connector to the graphics card. Let’s not forget that for years 350 W of consumption has been the limit . So a more energy efficient memory like HBM made a lot more sense.
PCI Express 5.0 graphics cards will bring with them a new connector for up to 600 W of power. This eliminates the need for more efficient memory . In other words: HBM2 memory becomes even more useless in a gaming graphics card. Since one of the problems it was originally created for has disappeared. And yes, it would be much better for environmental reasons, a type of energy that consumes much less. However, it is what the industry participants have decided: to increase the consumption of the hardware.
