Maybe... I think the bigger problem is that Intel was banking on corporate/commercial customers to buy these cards, and in the end they were only able to trick a few gamers and pro users into buying one. It really begs the question of which workloads are memory/bandwidth constrained, and how does this help them? It reminds me of Apple's impressive memory bandwidth figures for M1 Pro/Max, that were slightly soured when only a scant few real-world tasks fully took advantage of it.
We don't really need memory-mapped storage, IMO. If we do, it's perfectly attainable with swap/virtual memory (and pretty damn fast with newer NVME specs). I think Intel found something cool, didn't know how to market it, and ended up placing a bad bet on fairly domain-specific technology.
> It reminds me of Apple's impressive memory bandwidth figures for M1 Pro/Max, that were slightly soured when only a scant few real-world tasks fully took advantage of it.
Tangent, but... that's kind of missing the point. The M1 SoCs have tons of memory bandwidth to keep the GPU fed. The fact that this also gives the CPU more memory bandwidth than it can possibly use is a convenient benefit.
Oh, for sure; it's entirely a byproduct of engineering an SOC that needs high GPU bandwidth. That doesn't stop Apple from marketing it as a CPU boon though, and it certainly didn't stop starry-eyed HN readers from losing their minds over a spec that only a small handful of people care about. Even price-to-performance, a figure that hasn't been relevant for nearly 2 decades of commodity computing, is a better metric to advertise than memory bandwidth. People simply won't notice the difference.
Almost all workloads are memory bw constrained, is the problem of our era. That is one of the biggest drivers of M1 perf across the board.
Optane is orders of magnitude faster than virtual memory, it is a complete OS bypass. I think you are conflating Optane used in an SSD and Optane on the memory bus. SSD Optane was a way to package and sell it before the rest of the hardware and sofware caught up for DIMM based Optane.
> Almost all workloads are memory bw constrained, is the problem of our era.
Well, sure; I think that sentiment applies to networking/CPU maximization too. But, much like we discovered with SSDs, there is a point of diminishing return for most people. Pretty much anything running an NVMe SSD won't experience significant storage throttling. Anyone connected to wired broadband will have roughly the same experience as everyone else on it. Since so much software is designed around nominal specifications, I don't really notice the M1's memory bandwidth in regular use. Text editing, app launching, smoothness and usability... it's all roughly the same as my throwaway $300 Thinkpad that has a nicer keyboard and no ARM contrivances.
> Optane is orders of magnitude faster than virtual memory
Not really... Intel has only claimed that Optane is 3-4 times faster than flash-based NVMe, which still puts it at >10x slower than the speeds of a trashy DDR4 2133mHz DIMM. If it's supposed to replace memory, it's both slower and more expensive. If it's intended to be a storage volume, most people would probably get better performance out of a ramfs or tempfs.
We don't really need memory-mapped storage, IMO. If we do, it's perfectly attainable with swap/virtual memory (and pretty damn fast with newer NVME specs). I think Intel found something cool, didn't know how to market it, and ended up placing a bad bet on fairly domain-specific technology.