|
Anyone have a link on pcie gen4? I'm interested to see the direction it's going since the copper limit is effectively reached on gen3 and 128/130b is squeezing a lot out of the channel. DDR4 is kinda boring. There are some interesting problems cropping up with memory densities ramping up, but nothing very visible to the end user.
|
#
?
Jul 25, 2013 23:33
|
|
|
|
| # ? Apr 24, 2024 09:19 |
|
|
Here's PCI-SIG's FAQ. If it's not copper the whole way, it's copper at the connection endpoints because it's still backwards/forwards compatible. And despite that, it's yet another straight-up doubling of per-lane bandwidth. Pulling something out of my rear end, maybe Intel's work on Thunderbolt has had dividends in how to increase signal quality in weird situations? The more I look around on Haswell-E Gen4 rumors, it looks like it's mostly just sites with maybe not the best tech understanding mixing up Haswell E and Skylake from the original leaks, so oopsie on repeating that.
|
|
#
?
Jul 26, 2013 00:04
|
|
|
Factory Factory posted:And despite that, it's yet another straight-up doubling of per-lane bandwidth. Huh. I thought 10GT/s was copper's limit, but I can't source that. It'll be interesting to see what kind of tricks are necessary to get around that.
|
|
#
?
Jul 26, 2013 00:42
|
|
|
I might be able to PM you some stuff, one of the last things I was working on was some Gen 4 stuff. IIRC though, a much greater demand is being placed on the silicon vendors (of course) to do even more equalization/CTLE/DFE and compensation for channel effects, but they've been tightening things for system integrators for some while now. I bet new jitter requirements are going to be even more of a bitch.
|
|
#
?
Jul 26, 2013 00:59
|
|
|
movax posted:I might be able to PM you some stuff, one of the last things I was working on was some Gen 4 stuff. There's some stuff here : http://www.pcisig.com/developers/main/training_materials/get_document?doc_id=b5e2d4196218ec017ae03a8a596be9809fcd00b5
|
|
#
?
Jul 26, 2013 01:47
|
|
|
Malcolm XML posted:There's some stuff here : http://www.pcisig.com/developers/main/training_materials/get_document?doc_id=b5e2d4196218ec017ae03a8a596be9809fcd00b5 Looks about right; jitter tightened, switching connector mounting/footprint to SMT to combat the massive (relatively speaking) inductance from the current connectors, adding an additional EQ tap. Performance is probably challenging for guys like Pericom and their analog switches (what mobo vendors throw on there to support various lane configurations), but hopefully they're continuing to work on improving analog performance.
|
|
#
?
Jul 26, 2013 06:35
|
|
|
We just got some new leaks of Intel roadmap slides, confirming some rumors and clarifying others. First off, Haswell Refresh is indeed 2H14. But! Broadwell seems to be back on track for release in 2H14 also. It looks like Broadwell will be skipped for desktop and be a mobile-only release, while Haswell Refresh will overhaul all of the desktop SKUs. Makes sense, I guess; desktop doesn't really need the power savings. If you're having trouble with volumes or it's just pricey or whatever, fab where it's needed, first. I imagine it will dash a lot of hopes (or at least mine) for a DIY thin-mini-ITX HTPC using the Broadwell/Crystalwell combo. The roadmap doesn't currently say anything about Atom lineups; for example, in the -Y SKUs, the bottom-stack Celeron is ">= 1019Y," the Celeron 1019Y being a 2C/1GHz Ivy Bridge part. -U SKUs will get Haswell in Celeron in 4Q13, though. Also Ivy Bridge E in September and still expecting Haswell-E in 2H14. Intel also appears to be bragging that Broadwell will have improved video playback battery life to go with sub-5W platform power. As Anand(Tech) recently figured, that's one of Haswell's major power-draw achilles heels. Factory Factory fucked around with this message at 19:08 on Jul 26, 2013 |
|
#
?
Jul 26, 2013 19:05
|
|
|
Tocks get socks, ticks pay the bills 
|
|
#
?
Jul 26, 2013 19:07
|
|
|
While not specifically mentioned, I think that indicates that Broadwell + Crystalwell will be available on the desktop, though of course not in a socketed form factor. It looks like that article shows Broadwell replacing all the BGA-mounted Haswell products, with Haswell Refresh covering the socketed parts. Since the i7 R-series (Haswell+Crystalwell on the desktop) is BGA, it makes sense that they would replace this with Broadwell.
|
|
#
?
Jul 26, 2013 19:50
|
|
|
Show me a Haswell + Crystalwell ITX board, though. There's no pricing listed in ARK for the i5-4670R or i7-4770R, and there are no early looks showing off any products. And the i7-4850HQ is almost $500. 
|
|
#
?
Jul 26, 2013 20:01
|
|
|
You're basically not going to see Haswell+Crystalwell desktop products until availability increases, right now they're all going to more expensive laptops.
|
|
#
?
Jul 26, 2013 20:09
|
|
|
What improvements are the Haswell refresh supposed to bring anyways? Slightly better clocks + slightly lower power usage would be my WAG. I somehow doubt that Intel will budge much on prices for a refreshed Haswell and I'm not even expecting them to fix the crappy job they're doing with gluing the IHS down.
|
|
#
?
Jul 27, 2013 13:01
|
|
|
Factory Factory posted:The roadmap doesn't currently say anything about Atom lineups; for example, in the -Y SKUs, the bottom-stack Celeron is ">= 1019Y," the Celeron 1019Y being a 2C/1GHz Ivy Bridge part. -U SKUs will get Haswell in Celeron in 4Q13, though. There's nothing in roadmaps about Atom lineups because Intel is killing the Atom brand.
|
|
#
?
Jul 27, 2013 13:16
|
|
|
Huh, I didn't know that. RIP. Even so, there's no indication of where Silvermont will fit on the product stack.
|
|
#
?
Jul 27, 2013 15:16
|
|
|
So it would seem safe to assume that we won't see anything earth-shattering (if that) from Intel until the second half of 2014. Until then, Intel will chug along with a couple refreshes. Sounds like I'd be safe to purchase something in the near future as the next option would be to wait until the second half of 2014 which seems pretty far away. Even then it would be an initial release and that's if Intel remains on schedule. Sounds like my best bet for building a Xeon machine is to wait until after the Ivy Bridge-E refresh in September? I don't think these slides mention anything about Xeon.
|
|
#
?
Jul 27, 2013 16:40
|
|
|
Since the Mac Pro, using Ivy Bridge-E Xeon E5s, is coming around September/October, I'd say it's safe to assume the bare CPUs will be available around then, too. Maybe not the new E7s, but E5s, sure.
|
|
#
?
Jul 27, 2013 16:46
|
|
|
Looking to build a new system with 4 or 6 hyper-threading cores with more L3 cache, but don't also want to pay as much for the  IHS gap of the 4770K? Well, has Intel got the deal for you!
|
|
#
?
Jul 31, 2013 14:48
|
|
|
AES-NI seems to scale with the number of hyperthreaded cores, where RSA crypto (which isn't accelerated) seems to scale with the number of actual cores. Is there actually enough AES-NI hardware to satisfy all of the hyperthreaded cores, or is this an effect of memory bandwidth/latency? Is there a spec sheet or something that would have told me this without having to test it?
|
|
#
?
Jul 31, 2013 16:34
|
|
|
Sidesaddle Cavalry posted:Looking to build a new system with 4 or 6 hyper-threading cores with more L3 cache, but don't also want to pay as much for the The 4960 is slated to release as early as my birthday. I'm going to take it as a sign from the universe that I should plunk down  x100 and buy one! More seriously, the 4930 actually may be a worth while step up for a few extra $$$ Coming from a stock Q9550, Good upgrade? (30% VMWorkstation work, 50% transcoding, 10% compiling, 10% other) Also, is there any chance I can get more than 32GB ram by stepping up to a i7-49xx (ivy-e) rather than a 4770. I am forever juggling vms.
|
|
#
?
Aug 1, 2013 01:54
|
|
|
KennyG posted:More seriously, the 4930 actually may be a worth while step up for a few extra $$$ Ehhhhh... It depends entirely on what you're do- quote:(30% VMWorkstation work, 50% transcoding, 10% compiling, 10% other) Yes, buy that poo poo. You should be able to do 64 GB of RAM, too.
|
|
#
?
Aug 1, 2013 02:10
|
|
|
Ninja Rope posted:AES-NI seems to scale with the number of hyperthreaded cores, where RSA crypto (which isn't accelerated) seems to scale with the number of actual cores. Is there actually enough AES-NI hardware to satisfy all of the hyperthreaded cores, or is this an effect of memory bandwidth/latency? Is there a spec sheet or something that would have told me this without having to test it? To determine if hyper-threading provides any improvement, you need to estimate how well the execution resources are utilized by a single-threaded workload on the core. The main idea behind hyper-threading (simultaneous multithreading) is to share the same physical resources among multiple threads, which will increase execution resource utilization, resulting in overall workload speedup across all threads. If the instruction fetch and memory portions of the machine can keep up with a single thread (probably capping out around 4 steady-state instructions per cycle), then there won't be any benefit to interleaving multiple threads on the same physical resources. I'm not closely familiar with the RSA steady-state instruction sequence, but there are quite a few details about the Haswell microarchitecture available on the internet from Intel's slides and leaks. If you aren't looking at an accelerated workload, then a major bottleneck can be the ability of the issue queue to schedule instructions of different types. The realworldtech review of the Haswell microarchitecture shows the instruction scheduling limits of Sandy Bridge and Haswell. Haswell is an 8-issue, superscalar, out-of-order processor. That is, it can issue eight unique instructions per cycle, as long as some other requirements are met (and usually they won't be). Four of those eight instruction issue slots are for memory operations: two for address-generation (load/store), one for address generation (store only), and one for store data (all stores are split in Haswell). The other four slots are used for arithmetic computation. Each of those four slots can perform simple ALU arithmetic, but complex operations cannot be executed in all four arithmetic execution lanes. For example, only two 256-bit FMA operations can be performed per cycle. This is where knowledge of the specific workload comes into play; if you know exactly what instructions you need, then you can figure out the bottleneck. Of course, there are likely to be many additional rules about what ports can be active at the same time (and for what instructions), back-to-back issue delays, and other problems. A detailed workload analysis would require, at a minimum, the processor documentation from Intel. It's probably cheapest and easiest to just buy a chip and see what configuration runs the fastest. Real performance is usually difficult to estimate with good accuracy.
|
|
#
?
Aug 1, 2013 04:47
|
|
|
Factory Factory posted:Ehhhhh... It depends entirely on what you're do- My suggestion for everyone else that's dying for more than 32GB of RAM is basically to go with the general cloud service trend for many reasons. The only justifiable reason to me to have so much RAM on a portable machine is basically for performing customer demos in a constrained network environment, which would be common in DoD circles, for example. That's actually a drat good use case for the new Mac Pros coming out (and if you're doing such huge demos for DoD, you can easily afford even a $10k machine just for demos... which is probably what you'd have to pay for 64GB of ECC RAM for them) Intel's general trend for Haswell and even beyond is to not focus upon greater memory density in consumer devices and to put that purely on backends, and that frankly makes sense on every front I could think of.
|
|
#
?
Aug 1, 2013 14:49
|
|
|
necrobobsledder posted:Ark says max is 32GB on an i7-4900MQ and the HQ variant, http://ark.intel.com/products/75131 The i7-4900MQ is 4-core Haswell, not hexacore IVB-E.
|
|
#
?
Aug 1, 2013 18:26
|
|
|
necrobobsledder posted:Ark says max is 32GB on an i7-4900MQ and the HQ variant, http://ark.intel.com/products/75131 I don't want to get in a pissing contest over this, but just because it doesn't fit your usage model doesn't mean it isn't a viable requirement. Look, I'm not pissed that consumer grade Haswell didn't up the memory threshold, but if you're going to pay $600-$1000 plus ~$350+ for a mobo for a higher end processor for specific use cases, let people add more than some token amount of ram. I can't put my work datasets "in the cloud" and even if I did, by that time, I'm paying through the nose for them. To run just 4x8GB VMs I'm looking at ~$2/hr. That's ~$20 day. $~400/mo or $4,800yr. Yea, no thanks. You can run your cloud/ondemand all you want, but owning the hardware still gives me more flexibility/control and a mostly fixed cost, not to mention that the licensing is more straight forward as RHEL and most other products are licensed by processor core/socket now. This means I can all-you-can-eat rather than getting billed by the proc/hr the way Amazon does it.
|
|
#
?
Aug 1, 2013 19:08
|
|
|
Put together a cheapo desktop with a Celeron G1610. The slowest in the whole Ivy Bridge line. And you know what? It isn't really that slow at all! This thing packs a punch for $42 shipped.
|
|
#
?
Aug 1, 2013 22:01
|
|
|
KennyG posted:I don't want to get in a pissing contest over this, but just because it doesn't fit your usage model doesn't mean it isn't a viable requirement. Look, I'm not pissed that consumer grade Haswell didn't up the memory threshold, but if you're going to pay $600-$1000 plus ~$350+ for a mobo for a higher end processor for specific use cases, let people add more than some token amount of ram. I can't put my work datasets "in the cloud" and even if I did, by that time, I'm paying through the nose for them. To run just 4x8GB VMs I'm looking at ~$2/hr. That's ~$20 day. $~400/mo or $4,800yr. Yea, no thanks. You can run your cloud/ondemand all you want, but owning the hardware still gives me more flexibility/control and a mostly fixed cost, not to mention that the licensing is more straight forward as RHEL and most other products are licensed by processor core/socket now. This means I can all-you-can-eat rather than getting billed by the proc/hr the way Amazon does it. The guy quoted specs for laptop Haswell 4-cores, not SNB-E/IVB-E systems. You will be able to do 64 GB of RAM with an i7-4930K.
|
|
#
?
Aug 2, 2013 03:21
|
|
|
The French Army! posted:Put together a cheapo desktop with a Celeron G1610. The slowest in the whole Ivy Bridge line. And you know what? It isn't really that slow at all! This thing packs a punch for $42 shipped. I recently put together a budget Ivy build as well. What motherboard did you end up pairing with it? I put a G2020 on a H61 motherboard and it runs spectacularly well. Woodsy Owl fucked around with this message at 15:10 on Aug 2, 2013 |
|
#
?
Aug 2, 2013 10:55
|
|
|
Is there a timetable for when the haswell i3/pentium/celeron lines will get released?
|
|
#
?
Aug 11, 2013 23:36
|
|
|
Haswell i3s sometimes this fall, probably with the IVB-E and new Atom launches. Pentiums and Celerons will be the new atom uArch, not Haswell.
|
|
#
?
Aug 12, 2013 02:07
|
|
|
Anybody had any luck over clocking 4670s? I can get mine to 4.00, but any higher and it'll crash in the UEFI bios. I haven't tried overvolting, and I definitely won't try it unless I hear somebody else has had success.
|
|
#
?
Aug 12, 2013 08:48
|
|
|
Factory Factory posted:Haswell i3s sometimes this fall, probably with the IVB-E and new Atom launches. Pentiums and Celerons will be the new atom uArch, not Haswell. The Pentium brand is truly dead then, I find it kinda hilarious that it's fallen so far.
|
|
#
?
Aug 12, 2013 08:49
|
|
|
SRQ posted:Anybody had any luck over clocking 4670s? There's kind of an entire separate thread for that.
|
|
#
?
Aug 12, 2013 12:29
|
|
|
SRQ posted:The Pentium brand is truly dead then, I find it kinda hilarious that it's fallen so far. Really, 20 years is a pretty good run. I'm actually kind of shocked it made it this long. Pentium as a very interesting marketing story that is often shared with college marketing classes, usually in the context of being the antithisis of the Chevy Nova. Wikipedia posted:The original Pentium branded CPUs were expected to be named 586 or i586, to follow the naming convention of previous generations (286, i386, i486). However, as the company wanted to prevent their competitors from branding their processors with similar names, as AMD had done with their Am486, Intel attempted to file a trademark on the name in the United States, only to be denied because a series of numbers was not considered distinct. I kind of wish they had proceeded one more generation so we could have gotten to the Pentium V (Inception inside!) It's taken a lot of time but their i3/i5/i7 XXXX-y scheme is starting to make sense. Here's hoping they don't reboot it now that we have all had a moderate chance to figure out what the gently caress we're buying.
|
|
#
?
Aug 12, 2013 12:42
|
|
|
KennyG posted:It's taken a lot of time but their i3/i5/i7 XXXX-y scheme is starting to make sense. Not totally. On the desktop, you can get a vague grasp: i3, dual core, i5, quad, i7, quad with more cache and HT. Once you start talking mobile, all bets are off, as an i7 mobile can be less powerful than an i5 depending on what its power target is, and may have 2 cores depending on again, what the power target is. Now you also have a range of HD graphics that can be onboard, well drat, you basically are left with checking on ARK every time for the base clock, turbo clock, GPU type.. I'm fine with this, but it doesn't mean that the model number itself is helpful on its own, unless you can remember every single last detail
|
|
#
?
Aug 12, 2013 13:01
|
|
|
Fair enough, but I would point out that few people compare a mac book air to OMGWTFBBQFPS gaming rig. In comparing within a product segment, it generally works or at least can be helpful to know that an i7 in an ultra-portable is going to be better than an i5 in an ultra-portable. Yes an i7-4600M is likely going to get crushed by a i5-4570S but the fact that one is a desktop chip and one is a laptop chip should have already made that very apparent. As mobile is almost never DIY, you can easily compare like products from toshiba, sony, dell, hp etc with i3-4100M, i5-4200M, i5-3750M and i5-4570T and know which one is the fastest and who is trying to hide last gen tech in there. It's a lot easier to look at 1xxx 2xxx 3xxx 4xxx and figure out which architecture I want vs having to remember that Conroe came before Westmere but after Prescot (  )
|
|
#
?
Aug 12, 2013 13:51
|
|
|
SRQ posted:The Pentium brand is truly dead then, I find it kinda hilarious that it's fallen so far. Maybe the brand, but not the (modified) IP core itself - it's being used in the world's fastest supercomputer at the moment.
|
|
#
?
Aug 12, 2013 14:55
|
|
|
I for one wish that they would have continued the Pentium branding, and called the Pentium 2 a "Sexium" instead. 
|
|
#
?
Aug 13, 2013 15:07
|
|
|
canyoneer posted:I for one wish that they would have continued the Pentium branding, and called the Pentium 2 a "Sexium" instead. Sadly, they probably would have gone Hexium, as all English speakers tend to when using Roman numbers as prefixes.
|
|
#
?
Aug 13, 2013 15:23
|
|
|
LeftistMuslimObama posted:Sadly, they probably would have gone Hexium, as all English speakers tend to when using Roman numbers as prefixes. Well that and I don't think 1996/1997 America was quite ready for the Sexium (disappointing in so many other, derailing ways)
|
|
#
?
Aug 13, 2013 16:35
|
|
|
|
| # ? Apr 24, 2024 09:19 |
|
|
Was North America Intel's largest market for them by that far though? I thought part of the reason for the really terse, robotic names common among corporations had to do with international acceptance (some Chinese readers having trouble with certain letters / sound combinations, for example)?
|
|
#
?
Aug 13, 2013 19:02
|
|