Sandisk has just revealed a new memory technology called High Bandwidth Flash (HBF) which promises to unite the high capacity of 3D Nand with the extreme band width of the High Bandwidth Memory (HBM).
The concept, which is still in Your first generationeven offers 4TB of Vram capacity for GPUS, Opening possibilities for artificial intelligence (AI) applications, especially in data inference.
In addition, the company glimpses the future use of this technology on mobile devices and other equipment, elucidating competitive cost and energy efficiency as main benefits.
How does Sandisk’s HBF memory work?
The HBF architecture is based on Nand 3D technology with CMOS design directly connected to the array (CBA).
The structure combines a Nand 3D memory array with a logical DIE made in the process of high performance logic. Innovation is in the division of the Nand Array into several subunits that can be accessed in parallel, which allows massive bandwidth and high capacity.
The first generation of HBF uses 16 main HBF Dies, each containing 256 GB of 3D Nand memory.
We challenged our engineers to explore the power of scalability, and the solution was to share the great array into several smaller arrays accessed in parallel, thus reaching a massive band width
Alper Ilk Spring, Da Sandisk
Dies are stacked using a Sandisk -owned technology that minimizes deformations and allows simultaneous access to multiple dies, enabling the handling of hundreds or even thousands of parallel data flows.
The technique offers up to 512 GB of battery capacity, 21 times more than HBM3E, which has 24 GB per stack.
Advantages for AI and the high performance market
Although HBF memory cannot match the latency of DRAM, such as HBM, its application is excellent for tasks that require large storage capacity and bandwidth, but do not require ultrabaix latency. Examples include IA data inference, where throughput It is more important than minimum latency.
According to Sandisk’s head of memory technology, HBF’s goal of memory technology is to build a solution with a cost -like HBM, but 8 to 16 times more capacity.
The company also pointed out that HBF is designed to approach HBM mechanically and electrically approaching, requiring only small changes in the protocol for implementation on host devices.
| Aspect | HBF (High Bandwidth Flash) | HBM (High Bandwidth Memory) |
|---|---|---|
| Total GPUS capacity | Up to 4,096GB (4TB) with 8 batteries | Until 192GB with 8 batteries (HBM3E) |
| Battery capacity | Up to 512 GB (in the 1st generation, with potential future) | Up to 24GB per battery (HBM3E) |
| Bandwidth | Not specified, but close to HBM | Up to 1TB/s per battery (HBM3E) |
| Latency | Larger (own Nand, designed for throughput) | Low (great for tasks sensitive in time) |
| Main use | With large data sets and inference | High performance tasks, such as Ia and Games |
| Granularity of Access | Blocks based (NAND limitation) | Bits -based |
| Cost per gb | Smaller, closer to Nand | Higher, due to the complexity |
| Energy efficiency | High, ideal for devices that prioritize efficient consumption | Tall, but focused on performance |
| Underlying technology | 3D Nand with integrated logic (TSVS and CBA) | TSV interconnections stacked DRAM |
| Compatibility | Similar to HBM, but with protocol adjustments | Compatible with existing HBM standards |
| Scalability | High (focus on large future abilities) | Limited by the density of DRAM |
| Durability (endurance) | Smaller (limited writing, typical of Nand) | Greater (typical DRAM durability) |
Important to highlight:
- HBM is optimized for Ultrabaix latency e extreme performancebut at a higher cost and limited capacity.
- HBF prioritize capacity and cost-benefitbeing ideal for workloads that require high bandwidth and great storage.
Challenges and Limitations of the HBF
Despite the big advances, technology faces some challenges. Nand memories have a limited service life in terms of recordings, something that has not been detailed by Sandisk in the case of HBF.
Also, while DRAM allows bits level addressing, Nand works in block granularity, which is another obstacle to overcome. THE endurance Memory (recording durability) and costs associated with the use of technologies such as SLC or PSLC to improve the useful life can also limit their application into certain areas.
However, although HBF should not replace HBM in tasks that require ultra-bin latency, it can occupy a niche on the market where high and cost-effective capacity are priority.
The goal of HBF is to combine bandwidth similar to HBM with up to 16 times more capacity at a competitive cost
Alper Ilk Spring, Da Sandisk
Of course, Sandisk glimpses HBF as an open pattern in the future, encouraging the creation of a collaborative ecosystem to expand its applications.
Innovation plays light on the market at a time when the high -performance AI and computing market seeks more affordable and sustainable alternatives to deal with large data volumes.
After all, technologies like HBF are relevant in a context where the evolution of consumer devices, such as GPUS and SSDs, is increasingly accelerated.
Also read:
Future Applications
Sandisk already plans evolutions for the HBF in three generations. Technology, although still a “work in progress”, offers a new perspective for markets that require high capacity, energy efficiency and competitive cost.
Application in Artificial Intelligence is just the beginning, and the future may include mobile devices and other consumer products. With this proposal, Sandisk wants to position itself as one of the leaders in innovation in memories for high performance applications.
Source: Tom’s Hardware
Join the Adrenaline offers group
Check out the main offers of hardware, components and other electronics we find over the internet. Video card, motherboard, RAM and everything you need to set up your PC. By participating in our group, you receive daily promotions and have early access to discount coupons.
Enter the group and enjoy the promotions
Source: https://www.adrenaline.com.br/hardware/sandisk-apresenta-memoria-hbf-vram-4tb/
