The thing that's truly amazing to me is that they're pumping out a 662mm^2 (about 2.6cm square) chip on their leading-edge production process, and selling the full-SKU for "only" $4,000. By comparison, the very first Ivy Bridge parts were about 160mm^2 and were sold into a consumer (i.e., lower-margin, more cost-sensitive) space for about $330; scaling defects quadratically (though this is just a gross estimation), you'd predict this part (if in a /consumer/ space) to go for around $5650.
Add in the small volume cost of having to sell this thing into the server market, the price point they're putting their 'reticle buster' at is astounding. They must have really tuned their 22nm process as it's matured... the rest of the industry would kill for that kind of process :-)
Their ability to sell parts with defective cores fused off makes this analysis really quite complicated.
Ad absurdum, say only 1% the 18 core dies have no defects and can be sold as such. In a vacuum, the chip would have be priced very high to cover the costs of low yields.
But since they can sell the defects at 8 cores for $1000 and 12 cores for $2000, they can be profitable on the reject chips alone. The 18 cores that work are a bonus.
They're no doubt priced at what Intel thinks the market will bare, meaning that either the process is great (so they're priced low to sell a ton), or that there's just not that much demand for the most high end product. Or some combination of the two.
I'm not sure how it works out for these chips, but I remember stories a few years ago where the fabs were so high quality they didn't have enough defective chips to put in the lower bin, and had to bin "perfect" chips to reach target volumes.
The first few spins of AMD Athlons circa 1999/2000 were exceptional in this regard. Parts actually binned at 700mhz were downclocked and sold in 500mhz packages. I dont recall anything similar with Intel recently, maybe some if the core2 stuff.
Because server parts requires high quality of process. 14nm isn't ready for this. They've just managed to produce small (around 80mm^2), lower-clocked Core M. Desktop parts are planned for next year.
I read that and wondered something similar. One can argue either there was a lot of 'air' in the Intel pricing and some competition had deflated it, or they are getting pretty astonishing yields on their 22nm process. Probably a bit of both but still an interesting price point.
Oh, Intel definitely has huge margins on its chips, especially the higher-end ones (the higher, the bigger the margin). Intel could use quite a bit more "deflating" there. Even in laptops, which are quite mainstream and getting rapidly commoditized right now (mostly due to some "indirect" competition from ARM devices), Intel's chips are the highest margin component (perhaps with the exception of SSDs, but they probably beat that, too), and Intel's chips can represent up to 40 percent of a laptop's BOM, compared to 10-15 percent for ARM chips in mobile devices.
If you have been wondering why laptops have gotten such terrible displays (or other components) for so long, this is the reason. OEMs have no room for anything else when the processor is 40 percent of their cost.
> If you have been wondering why laptops have gotten such terrible displays (or other components) for so long, this is the reason. OEMs have no room for anything else when the processor is 40 percent of their cost.
The more expensive the processor is, the less (proportional to the total cost) impact a more expensive non-processor part has on the final cost, so I don't think that's the explanation.
> If it isn't on the list of things people look for then manufactures try to save money on it.
So, if I accept that, then the reason laptops have what was upthread called "such terrible displays" isn't that Intel processors are too expensive, its that better displays aren't on the list of things that (manufacturers think) most of the market is looking for.
That's quite a win of the deceptive marketing. Since 2006 I don't pay attention to the CPU (apart from ISA extensions present, but I do SIMD development), the important components of a laptop being display, battery and keyboard. And SSD, as of latest.
It depends on user usage patterns. CPU (to some degree), RAM, SSD and Video card are all high priority items for me. Screen I don't care much about because I prefer to dock and use two monitors... the screen only becomes a factor when I need to be mobile which is usually short bursts of time.
Assuming that's true, it's probably why Apple can make such nice machines: since there are no competitor Macs, they can be built with a sensible emphasis on quality across the board, not to headline specs at a price.
It has to do with how the cores connect to the rings, which probably run over the L3s. Each core "faces" one of the L3 slices and it looks like each L3 slice has a ring stop in the middle.
this is why core based licensing frustrates me. When I license Oracle (11g) the real way to get speed increases is to let it spread out over multiple cores, but that gets expensive given Oracles licensing terms, and even finding single or dual core systems is get hard.
What would you use as a proxy to get extract as much revenue as possible? Server memory size? Max database size mabe. Or get silly and the amount of datatypes you can use. (For $10k more you can use floats in your columns) and so on.
Core licensing is purely a price differentiator to get as much money as possible. The idea is that those willing to shell out money for 64 cores for their servers will be willing to shell out 20x or similar than those with 2 cores.
What would you use as a proxy to get extract as much revenue as possible?
That's the key. Complaints about core-based licensing aren't really about core-based licensing but about the very concept of value-based licensing (aka extract as much revenue as possible).
On the other hand, Intel's x86 cores are much faster per-core than its competitors, so core-based licensing could end up being cheaper for the same performance. This could also partly be why Oracle is pushing for many cores in their latest SPARC processors.
Are you sure? Yes, Power8 has impressive single threaded benchmarks. But Intel has long dominated the single threaded performance market due to their fab edge and their economies of scale (lets them put more into R&D than IBM).
Just wondering if you have more benchmarks than what looks like Spec benchmarks? I'm definitely interested if another manufacturer can beat Intel at single-threaded performance.
I wonder what the rationale for 18 cores was, since it's a rather unusual number of cores and not a round number - 16 would be the usual choice. Was it just "we can put two more on the die, so let's do that"?
Well, the biggest Ivy Bridge EX has 15 cores! There's speculation that the 18 core Haswell-EPs are actually Haswell-EX dies that Intel wants to get rid as fast as possible because these chips have buggy TSX (transactional memory).
So didn't they find an issue with haswell cores (and the early next generation, broadwell) that meant they had to issue an erratum to disable TSX instructions?
I'd wait until the Haswell-EX or Broadwell-E/EP chips.
"Cluster On Die can be understood as if you split the CPU and LLC into two parts that behave like two CPUs in NUMA. The OS is presented two affinity domains. "
Those Who Are About To Die ... err ... write kernel scheduler We Salute You
Nah. What I like about this feature is that it requires no new OS support because it's just NUMA and OSes already support NUMA. This feature should benefit clouds that are already dividing servers into shared-nothing domains.
Ah, the promise of Clojure and easy parallel programming is near at hand. At what number of cores, does Clojure become a lay-up choice over other hard to parallel languages? Anyone care to opine?
> Ah, the promise of Clojure and easy parallel programming is near at hand.
clojure, golang and before all that erlang as well (and myriads of libraries e.g. cilk etc). dearth of parallel languages have never really been the issue, imho.
<the 'opine' part>
it's how to move existing applications to new h/w, with the caveat that doing a flat out re-write is almost out of question. hoping that there will eventually be compiler that can just parallelize your application code, is just a pipe dream, imho.
I often find (in my own code as well) that nobody ever remembers OpenMP. It has some overhead, but it's not that big even without fixed thread count tuning (with which it nearly vanishes). Not automatic, but damn close.
If your problem is "I have a lot of requests from a web server that are quickly dispatched", then parallel programming doesn't get you anything.
The problems where number of cores matter are already looking at parallel programming solutions because they're running the problems on multiple machines, right?
NOW, for a new desktop application that would benefit from parallel programming, it becomes interesting. Perhaps the startup time of the JVM doesn't even matter anymore. :)
> If your problem is "I have a lot of requests from a web server that are quickly dispatched", then parallel programming doesn't get you anything.
why not? If a page render has 10 independent processes that must be completed before data can be returned to client, doesn't the pmap function give an approx 10X speed-up? Or are you talking in the aggregate where each request is being handle in its own thread, and then the page render speed is really just I/O bound. But even in that case if the web server is doing some heavy computation (wolfram alpha, for example) easy parallelization of a request can make the web a more effective compute engine.
Did you guys not like Anand? The moment he quit the site, it's showing up on HN every day. Or have I just started noticing it?
I'm guessing VPS/cloud providers will like this one. You can already pack a lot of memory and SSD drives into a server so CPU was becoming a bit of a bottleneck.
my guess: 1) anand left right before major planned product releases 2) anandtech usually delivers in-depth reviews that people around here like and/or need.
I'm not sure why you're being down voted as I have noticed this as well and thought the exact same thing regarding anandtech article frequency increasing since he left, when I logged on to HN this morning.
Call me a cynic, but I suspect many people browse/monitor certain sites for things that might do well on HN.. and the recent outpouring of love for Anand might have put the site on their radars.
Oh absolutely, I want to make it clear I didn't intend to insinuate on any particular person, more a trend I've seen on HN over the years.
I've seen it a few times before when $wellKnownProgrammer has an article hit the top of HN, then their next several blog posts all seem to hit the front page before everyone fatigues of it. I'm guessing certain people and sites become "flavors of the month" for both submitters and upvoters, but maybe HN has the stats to figure it out.
Are you sure you're not the one that's experiencing the Baader-Meinhof phenomenon? I subscribe to Anandtech's RSS feeds and they had a slower period recently for technical articles.
FWIW I typically find Anandtech to have really comprehensive analyses of new products. I also appreciate that they tend to highlight the bad things in almost everything they review.
Add in the small volume cost of having to sell this thing into the server market, the price point they're putting their 'reticle buster' at is astounding. They must have really tuned their 22nm process as it's matured... the rest of the industry would kill for that kind of process :-)