How does the HBF vs Xpoint technology
- Thread starter skybluega
- Start date
HBM correct?
HBM is designed to feed GPUs with enormous amounts of data, reaching bandwidth in the terabytes-per-second range. That's why NVIDIA's AI systems and AMD's Instinct accelerators rely on HBM. Optane Persistent Memory was much slower. Real measurements found read latency around 374 ns and bandwidth around 38 GB/s for reads, which is probably why Optane is dead.
I think a fair comparison would have the 3DXPoint dies linked by TSVs just like HBF and HBM (for 3D NAND and DRAM dies, respectively). The cost and performance were already not good enough for the 3D XPoint by itself to support a large enough market.
The DRAM and NAND prices have gone up, but I am sure the same would go for the XPoint, which was already starting a little too high, and now lacks sufficient suppliers.
Also, the current thinking is to replace 3D XPoint with 3D SOM, but it's still under development.
Last edited:
siliconbruh999
Well-known member
Dan i think you are confusing HBM(High Bandwidth Memory) with HBF(High Bandwidth Flash memory) which is meant as more of like SSD replacement like Optane
Scaling the Memory Wall: Behind Sandisk’s High Bandwidth Flash for AI Inferencing | Sandisk
High Bandwidth Flash (HBF)—a new form of NAND—delivers performance within 2.2% of unlimited-capacity HBM, while offering a significant increase in memory capacity—unlocking AI services that traditional HBM simply can’t support. Powered by Sandisk’s CBA (CMOS directly Bonded to Array)...
Xebec
Well-known member
My understanding is HBF is more or less 3D stacked NAND that is biased more towards bandwidth than just total density (like normal SSDs).
That should mean:
- HBF is much lower cost to produce than Optane was (Optane was novel - HBF uses well established tech that is easier to produce).
- HBF already is at much higher densities than Optane
- HBF has higher bandwidth than Optane
At the tradeoff of having significantly higher latency, and lower endurance (cycle durability) than Optane.
As for applications, the real competition for Optane and HBF was always DRAM, and given HBF has a true cost advantage, "good bandwidth", and "ease of production", it'll probably find higher far adoption than Optane did.
That should mean:
- HBF is much lower cost to produce than Optane was (Optane was novel - HBF uses well established tech that is easier to produce).
- HBF already is at much higher densities than Optane
- HBF has higher bandwidth than Optane
At the tradeoff of having significantly higher latency, and lower endurance (cycle durability) than Optane.
As for applications, the real competition for Optane and HBF was always DRAM, and given HBF has a true cost advantage, "good bandwidth", and "ease of production", it'll probably find higher far adoption than Optane did.