"Xpoint pretty much broken"

[h4rm0ny]

That's semiaccurate's take on it. I'm really not sure what to make of it as I find it hard to believe it could be this far off expectations. On the other hand, semiaccurate are pretty much the anti-Fudzilla: a professional high-level site with a narrow focus on this sector. They're traditionally very good with their news.

http://semiaccurate.com/2016/09/12/intels-xpoint-pretty-much-broken/

I could / should have put this in the memory forum but realistically hardly anyone will see it there and more legitimately, one of the major selling points of Kabylake is its support for Optane / Xpoint. If that's not as ground-breaking as expected, that puts Kabylake back to being more of just a Skylake bump.

I honestly don't know enough about the figures to say how much value Xpoint retains if this article is accurate, but the article seems to be pretty damning on the subject.

 

[h4rm0ny]

EDIT: Misread the reported latency on the XPoint memory. Post therefore wrong.

 

[h4rm0ny]

IIRC the latency of RAM is in ns and considerably less than 0.013ms. I think he is saying 10x SSD speed which is around DDR4 speeds if we are talking the very top end fastest SSD stuff or about 25-50% of it if we were talking run of the mill SSD.

Yes, I misread their post and have since edited out what I wrote. Too slowly for you, it seems!

The article says Xpoint is now looking at latencies of around 10x SSD (I'm going to kindly assume they mean 1/10th and guess marketing people put the slides together rather than engineers ). If we generously take something very fast like the PCI-E x4 SSDs from Samsung, they have latencies in the order of 0.3ms. DDR4 2400 has something in the order of 13ns. I.e. four orders of magnitude difference. Meaning whilst this stuff is interesting as a fast intermediary layer between SSD and RAM (or a replacement for SSD if close in cost which I don't expect it to be), it's problematic as a drop in replacement for RAM. It can hit the bandwidth, but not the latency.

So I could see this stuff being useful for where you want large contiguous blocks of data read - maybe textures on a graphics card for example - but the question is how important is it to have non-volatile memory? Where does this help?

 

[h4rm0ny]

^^ I'm taking his 10x performance to mean bandwidth not latency (which puts the transfer rates potentially into RAM ballpark though more last gen than current gen), latency claims are between 3x and 10x which is considerably short of RAM even assuming a bleeding edge SSD.

Yes, that's what I misread initially before I deleted my earlier post. They were referring to bandwidth and they're right. XPoint at these figures (10x SSD) can match low-end DDR4 RAM assuming we're talking 10x a fast PCI-Ex4 SSD. But latency is a very important factor.

EDIT: Again, I think what I'm posting is correct but am open to someone with more subject knowledge catching it if I've made an error.

 

[h4rm0ny]

Could be whatever the 'Next Gen' memory is that AMD mentioned with be on their Navi GPUs?

Or it could just be they want to get out the door before Micron, since it was a joint development.

Well Navi will have HBM and more significantly, the architecture is designed for HBM rather than attached to an older architecture as with the Fury. If you're making an architecture with much higher memory bandwidth, you design it differently. HBM and Xpoint are not the same thing, but they both have applications in the GPU area. So maybe XPoint was intended to be a "we have a big new thing too" idea.

I read this not long ago myself, pretty damning stuff. Of course a lot could happen between now and it's release but if true, well... Oh dear. This part: "one of the major selling points of Kabylake is its support for Optane / Xpoint". I have a feeling 'support' for Kaby Lake amounts to having x4 PCI-E lanes dedicated to a U.2 connector from the PCH - nothing exotic really.

If that's all it is and XPoint does "just" turn into SSD 2.0 then maybe we can see support for it on AMD platforms as well. Presumably Intel can block that legally, though?

 

[h4rm0ny]

Are you sure you're not confusing Navi with Vega?

AMD's slides said Vega had HBM2, but Navi had 'Next Gen' memory.

Source: http://cdn.videocardz.com/1/2016/04/AMD-Radeon-2016-2017-Polaris-Vega-Navi-Roadmap.png

I've been taking that to mean both Vega and Navi have HBM but they didn't claim it for Vega as there might be GDDR5 implementations of it. And that maybe Navi had some more advanced integration of HBM. Are you saying Navi will have something beyond HBM2?