‘HBM Father’ Say: “HBF Will Surpass HBM in Memory Demand by 2038”

Posted by snowfordessert@reddit | hardware | View on Reddit | 8 comments

[-]

CheesyRamen66@reddit

So is this basically repurposing nonvolatile nand to be used as system memory? Wasn’t this just what optane was attempting but with way lower latency and higher write endurance?

[-]

Imaginary-Case3976@reddit

I can't how intel fumbled the bag so badly with NAND and Optane. They were ahead of the game. They should have been competing with Micron and Sandisk in memory.

[-]

The biggest advantage for both DRAM-based HBM and NAND-based HBF is that they piggyback on established high-volume manufacturing processes, allowing them to fully benefit from economies of scale—an advantage that Optane simply didn't have.

[-]

Tuna-Fish2@reddit

The difference is that now there is a workload that fits the devices.

AI inference is massively read-heavy. If you have some other memory for the activations, you can treat the weights as a read-only load, which requires high capacity, high bandwith and doesn't care about latency.

[-]

1mVeryH4ppy@reddit

Yeah totally unbiased prediction.

[-]

TemptedTemplar@reddit

He's not wrong though. Breaking up the individual components across severe clusters or even a server board introduces bottlenecks and latency.

The more shit they can cram on to a single board performing the majority of the workload would streamline the whole inference process.

And I for one. Welcome flash memory as fast and as powerful as HBM.

[-]

CheesyRamen66@reddit

3D XPoint would like to say hello. Intel would be rolling in it if it came out 5 years later than it did.

[-]

Personal-Tour831@reddit

Unlikely. The technology 3D XPOINT had reduced monolithic vertical stacking potential subsequently incrementally widening the density gap between NAND over time.