HBM2 vs GDDR6 – Things to Know

It is the 21st century, an era of digital and virtual revolution. Every day high graphic games are being developed all over the world. To match the production of the games and other graphic facilities, newer and cutting edge technology is being developed to process these graphics. Here is the small, yet deep comparison between HBM2 vs GDDR6 you must know.

With the amount of data to be processed, this cannot be the task for a mundane, day-to-day use GPU. To fill this void, high-speed processing graphic cards are being released into the market.

GDDR6 and HBM2 are two types of computer memories that facilitate a processor to work more efficiently across a wide range of operations because of their high memory storage capabilities.  

What is GDDR6?

GDDR6 stands for Graphics Double Data Rate 6. It’s a model of synchronous graphic random access memory, with high storage capabilities designed for high-performance operations like graphic cards and gaming stations. 

What is HBM2?

HBM2 is an abbreviation for High Bandwidth Memory (2). It is one other memory feature found in graphic cards, along with 3D stack DRAM chips. It utilizes DRAM dies that are linked with micro bumps and silicon vias.  

How does GDDR6 work?

GDDR6 is built from a combination of features of GDDR5, GDDR5X, and some from HBM2. But, GDDR6 takes all these features and makes them enhanced.

Initially, GDDR5 and the editions that came before facilitated a single 32-bit channel along with one address bus and one 32-bit data bus. This was a primitive version. It did not come with a flexibility buffer as the configurer had to read and write operations in the same row. 

In contrast, GDDR6 comes with a single chip that can behave and function like two independent DRAMS. Each has its own command and data bus. 

How is this the enhanced version? GDDR6’s structure is beneficial as, more the number of channels within a system, the more is the probability that the memory controller manages the DRAM. This particular feature aids in avoiding extensive delays that generally occur due to page activation limits or page refreshers. 

Another perk of the GDDR6’s build is that it nullifies the results of the 16-cycle burst length. If the burst length equals 16 with a 32-bit wide bus length, then the resulting transaction is a 64-byte atom. 

Many designs, such as the GPU and CPU, utilize the 32-byte transaction atom. On splitting the DRAM into two channels, that are two independently functioning 16-bit channel, then the result is a 16n prefetch and still transact at the size of 32 bytes. 

What is QDR and DDR concerning GDDR6?

Till up to the GDDR5X, DRAM was double data rate, in which the data bytes would alter at the rising and the falling edge of the world clock. The world then saw the introduction of QDR with GDDR5X, which could toggle at 4 times the world clock’s frequency, twice the frequency of DDR. 

GDDR5X could facilitate both QDR and DDR, but the DRAM would run only half the DDR speed, only being present as a power-saving option.

GDDR6, the new-age version, now gives the user a chance to choose between QDR and DDR, which will warrant function at full speed. The regulations don’t need to separate vendors to support both, so the user gets two standards on the device. 

There is also a mention of the ODR (octa data rate), but no other details were disclosed, which might be a potential future update.  

How does HBM2 work?

HBM stands for High Bandwidth Memory, as discussed above. The reason it is so in demand is that it provides a high memory range for very little power utilization. It is also built differently than GDDR. The HBM functioning is closely mirrored to the Hybrid Memory Cube that was developed by Micron Technology, but the two devices are hugely incompatible with each other.

Compared to other memories of DRAM, the HBM memory bus is extremely huge. In case you have 4 DRAM dies in your stack, which would imply that there are two 128-bit channels for every die, which results in 8 channels and 1024 bits. 

Hence, if you have a graphics processor, the total memory bus width is 4096 bits. 

What is an interposer concerning HBM2?

Keeping in mind the fact that there are a lot of links to the memory of the device, it became prudent to devise a new method to connect the memory to the GPU. This is where the interposers come into play. The interposer needs the memory and the processor to be physically close together to reduce the number of memory paths. 

The DRAM of HBM is very tightly paired with the host computer die, processed by a distributed interface. The interface is then divided into various channels that are individually functioning and are not in sync with each other. 

This is how HBM2 achieves high-speed processing and operation with very little power consumption. 

Which is better in performance- HBM2 vs GDDR6?

Two things can only be compared when they have a bunch of common features. HBM2 and GDDR6 are built entirely differently from each other and have different functions and operations. 

Let’s look at the below video:

On the one hand, the high memory range of HBM2 can be used in Artificial Intelligence, Machine Learning, and advanced data computing.

On the other, GDDR6 can offer the same services but for high power requirements and lower costs. Since GDDR6 is more accessible in day-to-day life, it is used in commercial graphic cards. 

Which is cheaper to produce- HBM2 or GDDR6?

Without a doubt, GDDR6 wins this category by a huge margin. HBM2 is only the second version of its kind, so the manufacturing is not as widespread as GDDR. HBM2 also has a design so complex that it requires time, money, and special components like the silicon interposer for its production. 

In the post-production cost battle, HBM2 takes the lead because it possesses the ability for high-speed processing in a very time and while consuming very little power. 

Conclusion

There isn’t an answer to the question ‘Which is better- HBM2 or GDDR6?’, because they all have their upsides and downsides. Comprehensively, the result remains indecisive. 

In their current versions, both types of memories have their trade-offs, so in the end, it boils down to the user and their requirements. 

Frequently Asked Question