As we well know, HBM2 and HBM2E Two types of VRAMs have several important limitations in their use and development of GPU chips. They have many benefits from a performance point of view and use, they are material because of need, but they will not be in the future.
In this case, HBM3 arrives, within as little as we know, it will eliminate the problems of its previous two versions. To this day, we did not know its functionality, and most of all, how much improvement we can talk about compared to the replacement version.
HBM2 vs HBM3: initial comparisons and performance arrivals
As we say, the original data is already here, except that, as is often the case in these cases, it is not provided in the form of bandwidth, FPS or other standard metric in the field of sports.
As a viable TOP memory, the business world is interested in learning about HBM2 development in the server space Exascale, who eventually rule the world in their own great power. Thus, the data we will see reflects this situation and focus on the pull of organizational conclusions, bandwidth, latency, capacity, and power.
The question is clear and key: How much will future PCs benefit from HBM3 vs HBM2? The data is clear.
As we can see, the production of HBM3 will be a very important pillar of Exascale's plans, and it can reach Motion speed of 1.7x on systems over HBM2 a bit the number of cores and game cache underneath.
In HPCG performance in bandwidth can shoot up to 600 GB / s in these cases from 300 GB / s where the first versions start at low speed, a repetitive situation in the model stream.
The biggest problem that HBM3 will draw on is the NoC networks, where the bandwidth is bottled. The combination will improve slightly in these cases as long as there is no such bottle, but it is not the kind of memory that seems to be best for poorly designed systems.
The number of threads, cache types and sizes and directions are important
The frequency and number of cores also have a lot to say, since the bandwidth increases from 1 to 2 GHz at 300 GB / s to 600 GB / s if the frequency of the network is 5 GHz, something that is currently unavailable to almost all servers in the world.
The most reasonable scenario is the frequency of 4 GHz, where a slight decrease in bandwidth performance, especially in rate and copy size, is acceptable when considering current power.
Number of threads, c1 sizes L1 and L2, and number of directions they are also important, especially in the c1 of L1. The scale is about 80% from 12 to 24, something that is not mentioned in other parameters but shows the importance of designing future architecture, something that AMD has been working on for some time and that Intel took itself seriously over a year past.
In short, HBM3 on servers is highly dependent on NoC resources, high power will be used to get the most out of it in terms of bandwidth (much more when they increase their final speed in JEDEC) and ultimately the kind of memory that depends on the vulnerabilities and resources.
In gaming it will be redefining the most expensive memory with zero benefits, where it has already come up that does not provide any benefit with high resolutions and hertz However, it will be used by AMD in its high-speed GPU when pulling the world out. GDDR6 it is not an option for its voltage, only then do we hope that RDNA 2 it works well and thus avoids its inclusion and cost to the Lisa Su team.
On the other hand, NVIDIA has no intention of using it without its Tesla GPU and will bet on memory 18 Gbps GDDR6 increase bandwidth while maintaining the normal use of your cards, especially new uses 7nm and 8nm from TSMC and Samsung.
