AI is the culprit. Memory manufacturers prefer to use their available production capacity to make DDR5 and HBM memory. Both forms of faster memory play a role in modern (AI) systems. DDR5 is also the standard for modern computers, but HBM memory is particularly popular for AI accelerators.
Manufacturers have no motivation to increase DDR4 capacity. DDR5 and HBM5 have more customers, and manufacturers who want the legacy memory have no choice but to pay a premium.
I've seen the claim around but I'm highly skeptical of it. DDR5 is far too slow for anything where memory bandwidth really matters, any newly produced chip that's gonna be used for AI is on HBM3 or HBM3e, or possibly GDDR6/GDDR7 if it's a GPU pulled from the consumer segment. HBM5 is still a very, very early research project and is certainly not being produced yet.
rezad@lemmy.world 3 weeks ago
care to explain a bit more?
themachinestops@lemmy.dbzer0.com 3 weeks ago
AI is the culprit. Memory manufacturers prefer to use their available production capacity to make DDR5 and HBM memory. Both forms of faster memory play a role in modern (AI) systems. DDR5 is also the standard for modern computers, but HBM memory is particularly popular for AI accelerators.
Manufacturers have no motivation to increase DDR4 capacity. DDR5 and HBM5 have more customers, and manufacturers who want the legacy memory have no choice but to pay a premium.
Source: itdaily.com/…/ai-maakt-ddr4-duurder-dan-ddr5/
chameleon@fedia.io 3 weeks ago
I've seen the claim around but I'm highly skeptical of it. DDR5 is far too slow for anything where memory bandwidth really matters, any newly produced chip that's gonna be used for AI is on HBM3 or HBM3e, or possibly GDDR6/GDDR7 if it's a GPU pulled from the consumer segment. HBM5 is still a very, very early research project and is certainly not being produced yet.