Samsung and SK Hynix they will also lead an important category as high performance memory, since both have already introduced their new memory modules HBM2E and, of course, both comply with JEDEC specifications.
HBM technology has reached the visible limit that is now broken
New GPUs will have a category that no one would like: value. And this looks like a constant, at least in a high-tech or high-tech environment, because both NVIDIA and AMD will have to cry in their pockets if they are to maximize the performance of their GPUs.
The fact is that HBM and HBM2 were thus reaching the curious limit, mainly due to the length of the connection lines TSV. HBMs are sure to require these connections and this requires thousands of them per square meter, which requires electrical engineering far more than a simple GDDR6.
This has logically reduced the configuration of HBM batteries, which have a direct impact on their performance and power. Due to the apparent limitations between the memory, i TSV and incloser, all are limited after version to increase or at least maintain bandwidth and gradually improve the speed and power of VRAM.
HBM2E memory development for GPUs
Every high performance GPU requires a large bandwidth, but this should always be accompanied by computer capacity either individually or in parallel. Therefore, most GPUs with HBM of any kind are used in data centers or AI training centers.
Samsung revealed that it has completed breaking the barrier of HBM and HBM2 issues with their new HBM2E Flashbolt, claiming that they have achieved 33% operating efficiency (frequency of 1.2 GHz and voltage reduced to 1.2V). Innovation is curious, because the HBM death rate for four deaths will be kept as two channels of 128 pieces per death 8 stations and 1024 buses.
On the other hand, the frequency is doubled in two giga channels, where the interaction should be increased by a single bits allowing a larger volume of addresses and thus reaching the speed of second gigatransfers per second.
This is not the fastest, but it leaves the design simple with brutal bandwidth that allows it to conserve the power and complexity of ASIC design, allowing you not to worry too much about the integrity of the features.
The capacity will be another key development that the market needs, but interestingly, neither Samsung nor SK Hynix ever wanted to comment on the sector. Expected actual multiplication of 16 GB to 32 GB, where except for very high speeds, ECC will be added, a requirement for artificial intelligence.
AI and DeepLearning will breathe easy, gamers not so much
Development does not stop and data centers are seeing how their performance and computation require a large amount of VRAM on their GPUs. If the capacity increases by 50%, it goes up in computer prices by almost the same amount, so the fact that HBM2E is reaching the market is of great benefit to this kind of industry.
What about playing? The recent leak means that Intel will come to market with its GPU players with HBM2E memory directly, but though it can be confirmed, this is likely only HPC, due to the blueprint for the blues architecture.
The problem with HBM2E gaming is its high cost and low performance imbalance in bandwidth and GDDR6, where new 16 Gbps modules are expected. In that case, even with the highest bandwidth, the performance difference does not compensate for the cost unless the GPU requires that bus.
Therefore, desktop GPU limitations are also influenced by cost and speed, which is HBM3 I have to settle on the call.
Expenses are increasing
The HBM bundle is said to cost approximately about 2000 dollars, 30% more than the same size is worth DRAM. Due to its type and technology, this type of memory only works in special cases, so the market does not look like standard DRAM and the cost remains high.
There has never been a general acceptance of the industry, which has added to the HBM2E delay leaving no better picture for years to come. The only good thing about this is that this new memory is just one step to satisfy the specific GPU market and that HBM3 Due to its features, it is possible to open a range of products in different categories, reduce costs and make it more affordable.
Table of Contents