|
PC LOAD LETTER posted:Yea but we weren't really talking about server CPU's per se and while it could be due to budget issues there is nothing I know of anyone can point to say definitively one way or another why they're still sticking with pinned sockets even for AM4 much less now. The publicly stated reasons for going LGA on desktop don't really seem to have panned out. Its not like AMD has any issues running OC'd RAM and high clocked buses over a pinned socket vs a LGA and their heat issues are all due to core design/process not the bus voltage. Where are those big gains from going LGA hiding? The "gains" from LGA is more that they can use the same package for both sockets and soldered down parts - a cost savings from a development/supply standpoint, whereas PGA requires the socket. Why AMD hasn't switched? Probably backwards compatibility. Maybe when they move to DDR4 sockets? (Assuming they're still in business...)
|
# ¿ Jul 27, 2015 18:03 |
|
|
# ¿ Apr 26, 2024 03:49 |
|
Skandranon posted:If they can produce it economically, everything has a good use case for this. Why not blow away the entire SSD & HD market and own it all. Intel could effectively become the only game in town for storage, if it's as good as they say and isn't cost prohibitive. There's "economically viable" and "economically dominant". They've stated that it's clearly built to be "viable", which means it probably lands on a $/bit scale somewhere between flash and DRAM. Some applications don't need the performance that XPoint could bring, and they will always choose the sufficient and cheaper (per bit) solution. There's a reason that tape drives are still being used today. (Yes, really)
|
# ¿ Jul 29, 2015 20:54 |
|
Skandranon posted:If Intel has effectively turned storage into a problem they can solve with their CPU fabs, and have 1000x performance improvement, with a technology only they will have patents on and only they could manufacture... That's quite an "if" you've got there. Especially the part where this (technology and fab) is co-owned by Micron, who will be releasing their own products based on it. Skandranon posted:...they could push Samsung and all other SSD makers out of the market by aggressively pushing the cost of this new tech down. None of us have any idea how much this costs, but if I were Intel, I would be looking to own as much of the storage market as possible, from phones & tablets to consumer drives to high end server drives & specialty devices. ...but for the purposes of a good time, let's allow it. If they have a technology that wholly supplants flash in both density AND cost, then absolutely this is what they will do. Set the price just at a point that it precludes flash and then take over the world. However, there are a TON of people (companies, researchers, etc.) that have been working on emerging Non-volatile memories for decades. If any one of them got to an architecture and maturity where the performance was like this and the cost level was 3x flash, they'd have brought it to market - not wait for being 1x of flash. If you have a better performing thing (especially 1000x), you bring it to market when it gets to be a reasonable premium, not price parity per bit.
|
# ¿ Jul 30, 2015 03:39 |
|
japtor posted:It's slower than DRAM so that'd be slower than L3 caches right? Course I'm still trying to wrap my head around the idea above, I guess it'd effectively be like a RAM disk as your storage (or device storage as RAM?). One concern I have would be GPU performance since they use shared memory setups, or would having crap always loaded help a bunch? I'm just going off the basic idea that GPUs love bandwidth. japtor posted:Anyway here's a quick mention of HMC on Xeon Phi I posted about earlier (re: HBM), although it sounds like Intel and Micron have their own name (or version) of it: Yeah, they've confirmed that MCDRAM is HMC but modified. If you look up the HMC spec from Micron's consortium, the signaling is defined as much longer reach (i.e. across a PCB). For an in-package application like Knight's Landing, that would be pretty electrically wasteful, so at the very least I assume they changed the signaling - possible they added features etc., but that's harder to guess.
|
# ¿ Jul 30, 2015 18:15 |
|
EoRaptor posted:Turns out making them reliable over long periods is hard, especially at the feature sizes modern manufacturing methods use. Especially for a company that has no existing fab infrastructure and a long history of research projects that never make it to production. And yeah - HP announced it as a discovery, but it was a long ways out. Intel/Micron already have fabs capable of producing this stuff in volume.
|
# ¿ Jul 31, 2015 15:04 |
|
AVeryLargeRadish posted:1) Hasn't DDR4 been around for a while already? Haswell-E has been using it since last year at least. 1) Yes, DDR4 launched in servers last summer with Haswell Xeon (and then Haswell-E). 2) Gaming is mostly GPU-bound these days, so this is mostly true (as much as it's true that the CPU doesn't matter that much, which is also mostly true). That said, if you're doing memory-intensive workloads (like CAD/CAE), then it will help. And I wouldn't hold your breath for a big jump in DDR4 performance. The increments are getting smaller (in terms of relative %, 2400 vs 2133 is not as much of a jump as 1333 vs 1066), and the technologies are close enough to DDR3 that there's not a lot of surprises (at least not good ones). Turns out that process scaling for DRAM is almost as much of a clusterfuck as scaling logic (CPU) processes right now.
|
# ¿ Aug 7, 2015 04:24 |
|
BobHoward posted:That's about right. Those have been issues for a long time, yes. Single-ended has some life left, but only if you move to buffered setups, which hurt latency. Process scaling is also getting worse now in that the amount of charge held by these tiny capacitors is getting to be so small that it's within the margin on error for process variation. Meaning you have to refresh the bits at crazy rates, and it'll be basically impossible to make a defect-free memory die.
|
# ¿ Aug 7, 2015 15:14 |
|
Nintendo Kid posted:An ARM design that could seriously compete would need to be able to run x86-64 code at an acceptable speed. That's the barrier they'd have to hit to keep Intel "honest". This is true, but that's also assuming that the software momentum of current PCs will outweigh the software momentum of Android/iOS. ARM is already starting to eat into the laptop market (slowly), but PCs are kind of the last bastion where it's x86 or GTFO. ARM designs are making an end-run around it and going straight for server market where that momentum is less of an issue (and the market is actually not shrinking like PCs).
|
# ¿ Aug 26, 2015 14:51 |
|
kwinkles posted:Except that every arm server thing to date has been vaporware. There's only one company that's actually released anything to shipment (APM) and it follows ARM licensees' traditional model of "the first implementation of a new ISA (in this case, ARMv8) is a piece of poo poo". The promise of ARM as competition with Intel doesn't have anything to do with low power microserver garbage (which the entire processor vendor set - including a full bet by AMD - bought hook/line/sinker and is now failing spectacularly), but that any number of very good SoC companies now have the tools and ecosystem to make Xeon equivalents if they want to. There are still companies that have figured this out - companies with almost Intel-level resources - that will start showing up with designs. It's just taking a while since the hype train got started super early on ARM thanks to some companies that tried to jump the gun with 32-bit stuff.
|
# ¿ Aug 27, 2015 22:24 |
|
kwinkles posted:And these sold so well that hp server revenue fell by 12% or something and then they released a xeon version. Yeah, yet another feather in HP's visionary cap. But "THE MACHINE" will be totally awesome guys! Trust us!
|
# ¿ Aug 27, 2015 22:25 |
|
BobHoward posted:The other issue is that the market segments ARM is in, and the nature of how ARM does business with both fabs and the fabless companies that put their cores into complete chips, makes it very tough for ARM to justify even trying to design a truly high performance core that could go head to head against Intel on the desktop or in servers. If they tried they'd have to commit to losing money on a few generations, and it might be a lot of money. ARM doesn't have the luxury of an Intel size revenue stream, so they probably can't do this. This is why, to date, all these ARM server chips have been "toss a shitload of underpowered slow low power cores borrowed from mobile in a single chip and hope it looks attractive to somebody". But it turns out it's really not, because even in cases where the business workload is a zillion threads latency is usually still important, and slow cores have much worse latency than fast. So, this is all true. A true "server core" like Intel's is not a good investment for ARM themselves. The wild card comes in when someone with a lot of revenue, experience, and an ARM Architectural license (so they can design their own core based on the ISA) decides they want to take a shot at it.
|
# ¿ Aug 27, 2015 22:49 |
|
JawnV6 posted:That's not how this works. That's not how any of this works. Yeah, the list is fairly short. There are a couple of others that have announced publicly (Broadcom, AMD, Cavium) besides the ones you mention. The market (if it takes off) probably couldn't sustain more than a couple of them. But I think I can safely say that most server OEMs would be thrilled with a count of 1 or greater.
|
# ¿ Aug 27, 2015 22:56 |
|
Palladium posted:The key mistake Intel made was they underestimated how fast the ARM players could ramp up clocks and IPC while still maintaining their tiny power envelope. There's a pretty big time window between the original iPhone SoC to the Cortex A8 where they could have really beat ARM in their own game, but they chose not to probably because of intolerance for slimmer margins. By the time the Qualcomm S4 Pro came out in late 2011, Intel has effectively lost the mobile war to ARM's "good enough" level of performance. This type of thing (supported by all the bitching about how little Intel's cores are improving per generation earlier in this thread), is what the ARM side is hoping happens in server. The big difference is whose home turf it's playing out on.
|
# ¿ Aug 28, 2015 14:36 |
|
Yudo posted:I'm not at all laughing at Intel, nor did I ever suggest they not invest in fabrication technology as frankly they have no choice and neither really does everyone else. I am saying that they face challenges as scaling slows (10nm, for example, has already been delayed and 14nm was long slog) that may allow others to catch up. In other words, they may not be able to maintain the 2-3 node advantage they have in the past and may have to compete or innovate in other ways. I really can't tell what could possibly be inflammatory here. The "xy nm" numbers have basically lost their meaning these days. The screwy design rules and the 3D elements make them essentially made-up values in many ways. A better way to phrase this time in the industry is perhaps that the leader in fabrication technology will be the first to run headlong into the scaling "wall", which is imminent. In the intermediate time while they're trying to figure out why there are brick marks on their face, the trailer is closing the gap. Then the real fun begins when everyone is mostly stuck on the same technology node and suddenly the only way to win is making better (and/or cheaper) architectures.
|
# ¿ Sep 2, 2015 15:01 |
|
Skandranon posted:As your second graph shows, no. Moores law specifically relates to the density of transistors, not their clock rate. Yeah, the exponential scaling of single core performance ended in 2004, but at a system level there are other ways people have added performance. (more cores per chip, cheaper cores, more CPUs in a server/scaling, etc.) However, it's worth noting that desktop CPUs have been at 4 cores for quite a while now. Given that per-core performance isn't moving much, that should give you a hint as to why people are still perfectly fine using their 4-year old SandyBridge for stuff these days. The per-generation improvement is going to be slow and more focused on new features than raw performance.
|
# ¿ Sep 9, 2015 16:24 |
|
cisco privilege posted:It's weird to me to have a chip that's effectively old in tech terms that I have zero reason or need to upgrade. Glad I sprang for the 2600K when I did to replace a 4ghz 920, although a 2500K still more or less holds its own as well. Yeah, it took me a while to get used to the fact that my computer wasn't going to become completely useless for games in 2 years like it had been before.
|
# ¿ Sep 9, 2015 22:57 |
|
Combat Pretzel posted:So, is it actually faster than an Intel chip with equivalent die surface or power consumption? Or is this just a novelty crossed with a crowbar? You want actual information from a marketing roll-out? :ohyou: Serious comment: faster by what measurement? Part of the promise of ARM is chips that do better on specific things that Intel is not great at, but maybe not have quite the breadth of Xeon. The first design is a test platform - there's quite a ways to go with these guys - but QCOM has shown they can beat Intel in other markets. Of course, server is now home turf for x86...
|
# ¿ Oct 12, 2015 22:44 |
|
Combat Pretzel posted:Uh, performance? I suppose in this particular case, web requests per second and whatever database metric is important? Yeah, that's pretty much my point. The optimal configuration is very different, depending on workload. If it's thread/memory request bound, then lots of medium cores likely outperform a few big cores. If it's memory bound, more DDR channels, or it could be IO (like BobHoward said above). The goal is not to beat Intel at everything, just beat them at one relatively lucrative thing. Let some other vendor take aim at the other markets - there's plenty to go around. BobHoward posted:Exploiting a niche the monopolist isn't serving well is the way to go. Explained better than I've managed to do it, apparently.
|
# ¿ Oct 13, 2015 15:04 |
|
pmchem posted:Nicely said, and a lot of those points were touched on in discussion at realworldtech that I posted in the scientific computing thread a while ago: Jesus, I'm pretty bullish on ARM (in HPC even), but that guy is on some kind of .
|
# ¿ Oct 14, 2015 14:45 |
|
WhyteRyce posted:Intel also has some Atom based microservers if that is something you really, really needed Microservers (defined by me here as "power-efficient chips that have low TDP") are largely a joke. You get completely destroyed by infrastructure costs of scaling to the larger number of sockets you need to reach the same performance. Some OEMs have tried to address it by denser packaging, but they're mostly research projects/demos. "Performance per Watt" is certainly the goal for everyone here. I think the thing to remember is that "performance" is not a single metric. Whenever you see performance (raw or per watt) numbers, the follow up questions should always be "performance on what application?" Qualcomm definitely wants to win on performance per Watt - the question is more about which application they're targeting to do that in. Because they won't be able to do it for all of them.
|
# ¿ Oct 14, 2015 16:52 |
|
Boiled Water posted:Why not bench it against an embedded Intel chip? They did! The Ghost of Boiled Water posted:Why not bench it against a recent embedded Intel chip? Better question.
|
# ¿ Oct 15, 2015 20:38 |
|
PCjr sidecar posted:Not really. The memory is weird and small, non-optimized performance is embarrassing, and off-card communication is bad. If you rewrite or restructure your MPI app to address these issues you get mediocre floating-point performance. Mostly useful as a development platform, as optimizing to run well on Phi usually improves performance on regular Xeons, and understanding what node-level KNL will look like. Yeah, you can "compile and go", but you'll get complete rear end-level performance. KNF and KNC were mostly experiments. KNL is being pushed by Intel as the first real focused HPC implementation as a product. Of course, they also said that about KNC, so...
|
# ¿ Oct 16, 2015 17:32 |
|
No Gravitas posted:Yup, it was me. Give me something insane to do, and I will. Never saw your original posts, but the fact that you got one for $100 should tell you how useful they were versus the original list price... icc does do better, but it still needs a lot of attention. Going to the higher thread count would also have likely hurt performance as the scalability would be nowhere near perfect.
|
# ¿ Oct 16, 2015 19:55 |
|
BobHoward posted:It's 60 odd independent x86 cores which have a wider fancier version of AVX. It's a lot closer to writing code for normal x86 than CUDA. KNL isn't out yet. They keep pushing the dates back. The other difference with KNL vs. KNC/F is that it's capable of running the OS by itself now. With KNC/F you had to attach it to another processor over PCIe like a GPU. That "Xeon tax" won't be there with KNL. The out-of-orderness of KNL is still very mild. It's a modified version of Silvermont in terms of the out-of-order structures, but they added the massive AVX and the threading. It's definitely not going to beat a Haswell Xeon at anything unless great use of the AVX and threads are made within the application.
|
# ¿ Oct 19, 2015 15:36 |
|
Yudo posted:This was the big sell of Phi from the get go, it just hasn't quite delivered yet. As others have mentioned, it is not a fully realized product. That was mostly marketing. I mean, it would run, but to get any performance, you have to do a similar amount of tuning as on a GPU. And by virtue of their market, "just running" isn't really of much value.
|
# ¿ Oct 20, 2015 16:37 |
|
Combat Pretzel posted:Eh, I just looked up why you mention DDR4 together with Xpoint. From what I've just read on the web, it seems they're targeting 2017 for Xpoint DIMMs, and apparently also needs support for that format. Correct - no CPU you buy today will likely be able to use XPoint. Further, their targets are server for what Intel has announced. Any particular reason you want it? Are you running in-memory databases on your desktop?
|
# ¿ Nov 6, 2015 16:53 |
|
Combat Pretzel posted:Why wouldn't one want XPoint, if it's even faster solid state memory? --edit: I mean with NVMe interface. So you mean with a drive interface? Then yeah, compatibility of processor won't matter, as long as it has PCIe. Still not sure it will be of any use, but we'll have to see what they come out with. It's going to be more expensive than flash.
|
# ¿ Nov 6, 2015 20:09 |
|
EoRaptor posted:Intel hasn't shown any desire to use xpoint as a flash ram replacement in SSD's or phones or what have you. http://www.anandtech.com/show/9541/intel-announces-optane-storage-brand-for-3d-xpoint-products Intel posted:The Optane products will be available in 2016, in both standard SSD (PCIe) form factors for everything from Ultrabooks to servers, and in a DIMM form factor for Xeon systems for even greater bandwidth and lower latencies. As expected, Intel will be providing storage controllers optimized for the 3D XPoint memory, though no further details on that subject matter were provided. EoRaptor posted:They are targeting server memory via specialty DIMMs that allow a huge increase in the amount of memory a server can have, by using a blend of xpoint and regular memory on a single DIMM. This is either managed by the CPU itself or by an 'xpoint aware' memory manager (or both!) EoRaptor posted:On the consumer front, I'd actually expect Apple to be the first ones to use xpoint in their mac pro series. They have the total control over hardware and operating system you need to turn around such a product quickly, and price isn't the first concern for people purchasing workstation class mac products. Xpoint in a high end laptop would also make a lot of sense, if the price is justifiable. Did I fall asleep and wake up in a world where memory capacity was a constraint in the PC space?
|
# ¿ Nov 6, 2015 20:16 |
|
EoRaptor posted:I completely missed this. Oops. The smallest granularity of XPoint - a single die - is 128 Gb = 16 GB. That die would ostensibly deliver the equivalent bandwidth of a single DRAM die in the best case, which is like 1/16th of DDR channel BW. The minimum usable granularity for XPoint on a DIMM is probably well north of 64GB.
|
# ¿ Nov 6, 2015 20:27 |
|
Skandranon posted:XPoint makes a lot of sense in tablets too, if the cost can be brought down. No more need for a separate bus for memory and for storage, just partition 8gb as memory and the other 120gb as storage. Or, thought of another way, provide 128 GB of "RAM", of which 120 GB is some kind of non-volatile ramdisk. And interesting thought. Curious how bad the performance would be without any DRAM though.
|
# ¿ Nov 6, 2015 20:39 |
|
DuckConference posted:Given Intels delays on the 14nm node and them pushing out their 10nm node, I wonder if samsung or tsmc will beat intel to shipping 10nm node chips in volume. One note of caution: Because of the way transistor design has changed in the last few years, the published "xy nm" do not really have much meaning anymore, especially when comparing between vendors. It's sort of this generation's version of "my CPU is X GHz, so that means it's X good". Their fab lead is shrinking, though. In just about anything technology related, catching up is way easier than leading.
|
# ¿ Nov 23, 2015 19:26 |
|
japtor posted:Do you know of a good read/explanation on that stuff? Something along the lines of Anandtech I guess, technical enough that it wouldn't be covered by the usual tech news sites, but not to the point of being over the head of a normal nerd. I've seen it mentioned in passing like "so and so's x nm process is superior" or slightly more like "another part is y nm" (where y>x) but not much beyond that. I don't have anything specific, no. Sadly, the details are often hard for me to grok, and I've taken graduate VLSI courses. One thing to look at is that, even within a node, there are "application optimized" versions of each, targeted at different design points. As an example: http://www.tsmc.com/english/dedicatedFoundry/technology/28nm.htm The chart at the bottom shows the potential variants of TSMC's 28nm process, all with different electrical characteristics. EDIT: Looks like Intel gave an interesting presentation on process scaling to their investors recently: http://intelstudios.edgesuite.net/im/2015/pdf/2015_InvestorMeeting_Bill_Holt_WEB2.pdf Durinia fucked around with this message at 20:26 on Nov 23, 2015 |
# ¿ Nov 23, 2015 20:12 |
|
Grundulum posted:Is Broadwell-E anything more than rumor right now? I may wind up with some extra research budget in the spring, and I would use it to buy either a Xeon Phi or build a new system from the ground up based around the i7-69XXk. The only purpose of a Xeon Phi is personal suffering, so I'd recommend the alternative. The 10-Core BDW-E is still unconfirmed, but has stretched beyond the most rumor-y sites, so probably a decent chance. Schedule is a big question mark, however.
|
# ¿ Dec 2, 2015 19:03 |
|
No Gravitas posted:While I do mostly agree with you here, I have to say that it does have a huge "cool" factor. And if you get it on discount, it is worth it. Yeah, if you're paying the firesale rate, then go nuts and build both.
|
# ¿ Dec 2, 2015 20:19 |
|
http://arstechnica.com/gadgets/2015/12/intel-skylake-cpus-bent-and-broken-by-some-third-party-coolers/ So yeah, some aftermarket coolers are breaking Skylakes. Not sure if others have seen this previously, but it looks like several of the manufacturers (Noctua, EK Water Blocks, Scythe, Arctic, Thermaltake, and Thermalright) have commented on the issue. Short version right now: If you have a Skylake with one of these, take it off before moving your system somewhere.
|
# ¿ Dec 4, 2015 16:35 |
|
Combat Pretzel posted:Yeah, it's called the ProcessorGeneration-E. The -E chips generally come out in the same timeframe as the Xeons with that core. Broadwell Xeons aren't quite here yet (supposedly this quarter). If you're waiting for Skylake-E, it's probably going to be 2017.
|
# ¿ Jan 27, 2016 03:49 |
|
SuperDucky posted:
*snerk* You can run x86/AVX binaries on it natively, but even on KNL they will likely run like garbage unless you spend some time on them. In terms of total work, KN* and CUDA will both require you to expose a lot more parallelism in your code, it's just the language you do it in that differs. In terms of performance, K80 is significantly better than KNC.
|
# ¿ Mar 30, 2016 21:45 |
|
Just get a P3608 and be done with it. You'll have to take out a loan to build the rest of your system, but hey - it screams!
|
# ¿ Apr 11, 2016 16:55 |
|
craig588 posted:I'm happy with my Notcua D15 with my 5820K. It's way quieter than the Corsair H100 and H110 I've tried and is good for 4.3Ghz below 80C or 4.7Ghz if I was comfortable getting up to 90C. That testing was done with the newest prime95 running over 24 hours with AVX2 and FMA instructions which I found was able to generate the most heat. Processor variation YMMV and all. While I love how my D14 cools silently, I'm going to guess that the 140mm fans on the Noctua, combined with its
|
# ¿ May 4, 2016 15:03 |
|
|
# ¿ Apr 26, 2024 03:49 |
|
Combat Pretzel posted:There's been scuttlebutt about both sharing the same socket at least back in November already. Doesn't sound impossible. Maybe they want a single Northbridge for both CPUs. They both have 6 external memory channels, which is the major driver of that pin count. Reusing the socket would make economic sense. (I'm talking about same physical socket - I seriously doubt they'd be socket compatible in their pin arrangement)
|
# ¿ Jun 1, 2016 18:00 |