Just another rant on CPU

[serious]

Warning: It will be a waste of time reading. I just want to rant.

I think Intel has been going into the wrong direction for quite some time now. Yes, there is a market for iGPU, but there is also a market for 'no iGPU'. Idiots like me don't even want to upgrade since Sandy Bridge as there very little points. Newer CPU are not faster nor smaller, but feature iGPU that I never use. I don't mine dropping a few thousand dollars to upgrade my PCs, but why should I when I really don't get anything out of the CPU?

Yes, the biggest change I have been missing out for the pass 5 years is USB 3.1, but most USB devices are still 2.0; PCIe 3.0, which does not impact GPU performance compare to PCIe 2.0; NVMe, which is a pain to setup boot drive with it, not to mention the overheat issue. M.2 sounds great, but you technically need to shop for the M.2 drive with the motherboard after you have done researches on how to set it up. This is stupid, so stupid that it makes more sense to simply get a PCIe SSD, but wait, it can't be used as a system drive, F you!

Why can't Intel or AMD simply create a NVMe port that works just like SATA ports so SSD vendors can make M.2 drives that fits in 2.5/3.5" bay with adequate cooling on it? Why can't we have 6 NVMe ports for multiple M.2 drives? It is because of the technology or stupid design? Here is an idea, M.2 GPU! oh wait, most motherboard only have exactly ONE M.2 port ...

Back to iGPU, what's with it sitting inside the CPU die doing nothing where it can be used as a parallel processor? 'DirectX 12 multi-adapter', why will developers build a fucking code path that utilize multi-adapter when most client uses single adapter? Yes, it will work if programmed correctly, but what is the benefit? 5-10% max? That is as bad as Nvidia's idea of having a GPU that act as a physic calculator, which is has similar effects on selected games. The only logical reason for developers to feature multi-adapter is for enthusiasts who like to use multiple highend GPUs, cross vendor. Extreme? Yeah, but that is a the point. Highend GPU + iGPU is simply a waste of time on multiple levels. IMO have MS figure out how to run docker/android/VM through iGPU.

AMD, you were never meant to sell cheap stuffs, you were meant to sell stuff that works at a reasonable price! Intel got greedy with its iGPU and other garbage bake into its CPU and you can beat it with ease as long as you can still create a 'Central Processing Unit' instead of another 'Multi-purpose Computer', run as fast as Intel CPU with much better efficiency and a slightly lower price. Make sure your chipset supports the latest technology, and you will turn from an underdog back to enthusiast's pick.

Sorry, I know it doesn't make a lot of sense, and thanks for your time reading.

 

[Shintai]

You can just buy a CPU without IGP or with IGP disabled.

The port you talk about is called U.2. And you can create more via PCIe slots.

You can use the IGP for transcoding etc. DX12 no. Else you can benefit of is as dark silicon during overclocking.

 

[4330thgink]

I concur with e iGPU thing. Any enthusiast will take a dedicated card over any integrated GPU. No need to waste silicon. Only OEM manufacturers and low-budget builders use the iGPU.

As for the M.2/NVMe = it isn't as bad as you have it portrayed. I chose a platform, then a couple mobos in my price range - then looked for which ones had confirmed M.2/NVMe bootable experience with known drives. I ended up with a Asrock X99 Extreme 4 and a Samsung 950 Pro. After a couple BIOS settings it detected the 950Pro at post, then Win10 installed from a USB drive in about 2-3 minutes(if that, including formatting). This was my first system build since 2004'ish(whenever the AMD mobile Barton Socket A's were out was my last build, purchased a HP machine in between because work bought it for me).

Length of connection plays a part in why certain devices have to be in certain places. Some devices dont have the signal strength for farther locations. M.2 also requires a lot of bandwidth, which the Intel/AMD processors only have so much of with their designs.(you can get more with server grade CPU's/mobos). My 950Pro uses (4) PCIe3 lanes. So, if you have even a 40 lane CPU like the upper 2011v3 CPU, having (4) M.2 devices will eat up 16 lanes if you set them for max bandwidth. Or, in some motherboards, you can set them to use a portion of the bandwidth used for the SATA ports(effectively disabling them). This could be remedied by Intel/AMD by designing more lanes into the CPU's.

 

[Dan_D]

There are plenty of CPUs without iGPUs. They are on the HEDT platform. X58, X79, X99 etc. Sandy Bridge is reasonably good in games but its gets its ass handed to it in productivity applications. Its 5 generations old now and those miniscule improvements have begun to seriously add up. What saved Sandy Bridge were its clock speeds over the others. Since Devil's Canyon, we've been close enough in regard to clock speeds to negate that advantage for the most part. Whether or not its worth an upgrade is a personal choice. I would for the platform alone as the system is overall better. U.2 does what you are talking about and you clearly have no experience with M.2 or NVMe drives because they aren't a big deal. They were problematic on Z87 and Z97 motherboards. It's not two or three years ago anymore. Those issues have been resolved for anyone who moved away from those older platforms.

AMD, you were never meant to sell cheap stuffs, you were meant to sell stuff that works at a reasonable price! Intel got greedy with its iGPU and other garbage bake into its CPU and you can beat it with ease as long as you can still create a 'Central Processing Unit' instead of another 'Multi-purpose Computer', run as fast as Intel CPU with much better efficiency and a slightly lower price. Make sure your chipset supports the latest technology, and you will turn from an underdog back to enthusiast's pick.

For most of its existence, AMD has been a second rate processor manufacturer at best and a knock off chip maker at worst. Many of its early designs were essentially reverse engineered Intel CPUs with enough changed to keep them out of trouble. Let me give you a history lesson. AMD's been sucking on Intel's exhaust pipes since the i386 days. They took the i486 type of architecture further, but Intel didn't care because it had Pentiums at that point. While the Pentium wasn't always faster than a 486 DX4 100 that was clocked much higher, the platform was better. AMD's K5 was late to market and a pathetic alternative to the Pentium. Meanwhile another company called NexGen Systems had a true Pentium alternative called the NX586. At the same time, Compaq had purchased Digital Equipment Corporation and laid off the DEC Alpha development team and embraced the x86 platform. AMD did two really smart things. It purchased NegGen Systems and hired the DEC Alpha engineering team who began working on K6 and K7 architectures based on technologies taken from those two acquisitions. This may not have worked as well as it sounds had Intel not made a choice that ultimately back fired on them. They picked a design direction and it was pretty lame. It counted on its manufacturing capabilities to take the Netburst core to 5GHz and potentially beyond. It never happened and it lost the performance race but not the sales race.

Excluding K7 and K8, AMD has basically been "that other" processor manufacturer for decades. So while AMD would like to be a company that competes on equal footing with Intel they have usually been the company that sells cheap stuffs as you put it. AMD's biggest problem isn't even Intel, its their own management. This is a company that wants to be in the CPU business more than anything in the world and sells off profitable divisions to inject cash into a business that its rarely shown any aptitude for. Aptitude that when examined in the proper context may not have been aptitude as much as simply having the right product at the right time thanks to Intel's Netburst blunder.

It would be nice if AMD could bring something worth while to the table and I think Zen will be the best processor AMD has had in a decade. I seriously doubt that it will be the i7 killer people have been looking for. Given AMD's own statements about Zen's performance I wouldn't get your hopes up on it being anything more than a Haswell equivalent CPU with more cores that can be had for a reasonable price in some segments.

 

[Shintai]

So spot on Dan. This is also why their graphics division got sacrificed. And to be more precise, since K8 they want to be a server CPU company before anything else. If they could get away with it they would be happy selling server CPUs only.

 

[serious]

...U.2 does what you are talking about...

Thank you all for showing a different perspective on this issue. Let me elaborate on my argument on M.2.

If I want to add a SATA drive, i simply go buy one can plug it into the SATA connector of the motherboard. Can you add another M.2 drive on your current motherboard in the future? No, unless you have multiple M.2 slots motherboard. Why most motherboard have only one M.2 slot? In short there are only 4 DMI and 26 HSIO lanes, all USB 3.0 and Sata 3 uses a HSIO lane, NIC uses one, so there are only enough HSIO to have at most 2 M.2 connections. U.2 is just another port that requires 4 HSIO lanes, the number of HSIO lanes does not change. Yes, there are CPU with iGPU disabled, but that was not what I want, as some of the resources used by the iGPU would have been used as DMI which in tern increase HSIO performance/availability.

I admit I don't know much about the specifics, but I believe the speed of each HSIO lane is similar to DMI 3.0 lane which is also similar to a PCIe 3.0 lane, which is a little less than 1Gb/sec. That means, the combined i/o of all devices can not exceed 4Gb/sec as there are only 4 DMI lanes from chipset to CPU. Without the iGPU, which I believe it consumes 4 PCIe lanes from the CPU, they would have diverted those resources to communicate with the chipset, thus allowing better io communication.

As to AMD, Thanks for your version of the history lesson. All I am saying is, if AMD can release a CPU that allows 2x HSIO compare to z170, then it will be a win against Kabylake even if it is a little slower in terms of clock speed. That is, AMD's CPU will have its market being cheaper. Don't get me wrong, i'm not saying the next CPU from AMD will do that, I am saying AMD should look into this area, where Intel completely missed.

 

[Shintai]

Not sure why you dont do some research first.

What do you think this is?

On LGA115x everything shares the DMI. Doesn't matter if its SATA or M.2/U.2.

On the HEDT platform you got the options to get full speed due to all the extra PCIe lanes.

AMD isn't going to be your savior. AM4 is a very downgraded LGA115x version.

The IGP have nothing to do with this. So stop ranting about it.

 

[pxc]

I think this is related to 2 issues: who the bulk of Intel's customers are (not enthusiasts) and minimizing SKUs. Intel's main customers may actually want the iGPU since many of those processors, which for quad core models are essentially the same die as K processors but with HT, unlocked status and frequency differences.

In the longer trend, the problem of unwanted iGPUs will eventually be solved as the classic desktop segment "dies" and it evolves into newer and smaller form factors. There will still be a market for CPUs used in full-size PCs, but I think it will be hard to justify a separate line just for desktop PCs. I would expect future "desktop" chips to merge sockets with 1P server processors (which at the higher end desktop has already happened), since it will basically be an enthusiast socket. No die space is wasted on integrated video.

 

[lutjens]

I would expect future "desktop" chips to merge sockets with 1P server processors (which at the higher end desktop has already happened), since it will basically be an enthusiast socket. No die space is wasted on integrated video.

They are...it's called Socket 2066, or Socket R4. The only difference between the two chips is that the 1P server chips are hard locked. The unlocked low core count X chips are neutered versions of the server chips with many features needlessly cut out just because Intel doesn't want an unlocked chip that has ECC and other similar niceties that their low TDP server chips have.

 

[oleNBR]

Not sure why you dont do some research first.

What do you think this is?

On LGA115x everything shares the DMI. Doesn't matter if its SATA or M.2/U.2.

On the HEDT platform you got the options to get full speed due to all the extra PCIe lanes.

AMD isn't going to be your savior. AM4 is a very downgraded LGA115x version.

The IGP have nothing to do with this. So stop ranting about it.

issue is, most pcie slots on motherboard run through CPU rather than chipset. if that adapter is used, most likely it'll still work and be bootable on the latest boards, but you cant raid them if you have two or more.

 

[Shintai]

issue is, most pcie slots on motherboard run through CPU rather than chipset. if that adapter is used, most likely it'll still work and be bootable on the latest boards, but you cant raid them if you have two or more.

On LGA115x everything but graphics as such goes via the chipset. I doubt you find a single LGA115x motherboard with a M.2 connected directly to the CPU.

 

[oleNBR]

On LGA115x everything but graphics as such goes via the chipset. I doubt you find a single LGA115x motherboard with a M.2 connected directly to the CPU.

oh so the pcie slots are designed that way i see. well i guess desktop side they really thought it out well, given with LGA 115x CPU it only comes with 16 lanes so its gonna be 2 x 8 for graphics. in that case thats great, you can still enjoy pcie SSDs in raid 0 even if those boards dont come with m.2 slots.

 

[pxc]

They are...it's called Socket 2066, or Socket R4. The only difference between the two chips is that the 1P server chips are hard locked. The unlocked low core count X chips are neutered versions of the server chips with many features needlessly cut out just because Intel doesn't want an unlocked chip that has ECC and other similar niceties that their low TDP server chips have.

What mean is that the mass/mainstream socket will totally disappear, because they wont be practical for smaller low power form factors, replaced by BGA SoC chips for the mass market. The full-size desktop market will eventually only become an enthusiast market. Right now that higher end socket (2011 and possible eventual R4 replacement), are a tiny segment of the desktop enthusiast market since the bulk use unlocked processors on LGA1150/1151.

 

[lutjens]

What mean is that the mass/mainstream socket will totally disappear, because they wont be practical for smaller low power form factors, replaced by BGA SoC chips for the mass market. The full-size desktop market will eventually only become an enthusiast market. Right now that higher end socket (2011 and possible eventual R4 replacement), are a tiny segment of the desktop enthusiast market since the bulk use unlocked processors on LGA1150/1151.

What you are saying makes perfect sense and is very likely. In addition, these chips (as they lose their socket) will likely blend in well with Intel's BGA mobile chips in terms of TDP and features, with a possible performance SKU or two for enthusiast notebook market. Only the very large Eurocom-type desktop replacement units may have the full sized CPU in a somewhat mobile form factor.

 

[Dan_D]

Thank you all for showing a different perspective on this issue. Let me elaborate on my argument on M.2.

If I want to add a SATA drive, i simply go buy one can plug it into the SATA connector of the motherboard. Can you add another M.2 drive on your current motherboard in the future? No, unless you have multiple M.2 slots motherboard. Why most motherboard have only one M.2 slot? In short there are only 4 DMI and 26 HSIO lanes, all USB 3.0 and Sata 3 uses a HSIO lane, NIC uses one, so there are only enough HSIO to have at most 2 M.2 connections. U.2 is just another port that requires 4 HSIO lanes, the number of HSIO lanes does not change. Yes, there are CPU with iGPU disabled, but that was not what I want, as some of the resources used by the iGPU would have been used as DMI which in tern increase HSIO performance/availability.

Yes, you can add more M.2 drives to your PC in the future. Not only do many motherboards support two or three M.2 devices, but you can potentially get the same exact drive in a U.2 form factor as well. Additionally, PCI-Express to M.2 adapters exist and are plentiful. In fact, the UEFI on many motherboards actually has settings for enabling NVMe devices through this style of input. There are lane limitations as you pointed out, the key one being DMI 3.0. Essentially, you can't benefit more than 2x drives in read tests but you can benefit from 3x drives in write tests. Fortunately, the bandwidth limitations aren't as big a deal as one might think. Rarely if ever will you hit these limits in anything but benchmark testing.

As for the iGPU comment, that's not how this works. That's not how any of this works. You can't simply say: "The space on the chip used for an iGPU can be used for DMI." DMI is an external bus between the chipset and the CPU. It doesn't work that way. You can increase the number of PCIe lanes provided by the integral PCIe controller. Such CPUs already exist. Intel offers CPUs with more PCIe lanes that do not go through the chipset. You need to step up to Haswell-E or Broadwell-E on the X99 platform. Intel builds the LGA 1151 CPUs the way they do for a reason. Its called market segmentation.

I admit I don't know much about the specifics, but I believe the speed of each HSIO lane is similar to DMI 3.0 lane which is also similar to a PCIe 3.0 lane, which is a little less than 1Gb/sec. That means, the combined i/o of all devices can not exceed 4Gb/sec as there are only 4 DMI lanes from chipset to CPU. Without the iGPU, which I believe it consumes 4 PCIe lanes from the CPU, they would have diverted those resources to communicate with the chipset, thus allowing better io communication.

As to AMD, Thanks for your version of the history lesson. All I am saying is, if AMD can release a CPU that allows 2x HSIO compare to z170, then it will be a win against Kabylake even if it is a little slower in terms of clock speed. That is, AMD's CPU will have its market being cheaper. Don't get me wrong, i'm not saying the next CPU from AMD will do that, I am saying AMD should look into this area, where Intel completely missed.

That's not how this works. The integrated GPU doesn't use any of the DMI lanes according to the block diagrams of the Z170 Express Chipset. Again you can't drop the iGPU and magically double the DMI 3.0 bandwidth or available PCIe / HSIO lanes. If it worked that way we would probably have more DMI bandwidth on the HEDT (X99) platform where that bandwidth and more lanes are needed and we don't. You do make one valid point. If AMD's chipset / platform for Zen offers more PCIe lanes, competing more with X99 than Z170 it could sway many buyers in that direction despite having a slower CPU for some tasks. With more physical cores and more PCIe lanes it could be a very tempting platform for home users who utilize virtualization and the type of applications often suited for the HEDT segment rather than the mainstream segment.

The bandwidth issues present on both platforms are often over hyped. Rarely would you ever have enough devices connected to the system, running balls out at the same time. Only synthetic testing exposes these bandwidth limitations. In the real world they are far less apparent.

 

[Dan_D]

issue is, most pcie slots on motherboard run through CPU rather than chipset. if that adapter is used, most likely it'll still work and be bootable on the latest boards, but you cant raid them if you have two or more.

Not true. You can run more than two drives in a RAID configuration on some motherboards. You can also mix M.2, U.2 and PCIe SSD's via adapters, or straight PCIe SSDs. At this point such a thing is realistically only possible with the Intel SSD 750's as they are the only U.2 drives I'm aware of for the consumer market. Also, you can do 2x M.2 drives and a third via a PCIe adapter. Again, some motherboards support this. Support for this is only going to improve. M.2 has taken off far faster than the motherboard manufacturers expected. As a result they are working feverishly to make support for these devices as flexible and as good as they can.

 

[whateverer]

Thank you all for showing a different perspective on this issue. Let me elaborate on my argument on M.2.

If I want to add a SATA drive, i simply go buy one can plug it into the SATA connector of the motherboard. Can you add another M.2 drive on your current motherboard in the future? No, unless you have multiple M.2 slots motherboard. Why most motherboard have only one M.2 slot?

Because we're pushing the very limits of what current Copper I/O technology can do, and Intel is currently pushing the limits of inexpensive CPU I/O (20 lanes PCIe 3.0).

It takes FOUR PCIe lanes to feed an m.2 drive. There are only twenty available.

It takes 8 lanes to feed a GPU, keeping in-mind future bandwidth growth needs. Double that if you want SLI.

Then you have to add 4 more dedicated lanes for things like Thunderbolt,(or two dedicated lanes for JUST USB 3.1).

Can you start to see why maybe we only have room for 1-2 m.2 slots? It's because it's not a DOUBLING of bandwidth (like SATA 1.5 -> SATA 3.0 -> SATA 6.0), it's FIVE TIMES THE BANDWIDTH OF SATA6!

We can't just use our magic wand to whip-up more PCIe lanes, those will take awhile. And don't expect them to get much faster on the mainstream platform for several years; PCIe 4.0 will start-out on the enthusiast platform.

Going from PCIe 2.0 to PCIe 3.0 took three years.

Going from PCIe 3 to PCIe 4 took five years to develop, and will probably be the last upgrade before we move to optical. Because making copper transmit data faster is becoming a VERY HARD PROBLEM to solve.

So don't expect to see more than 3-4 m.2 slots on future motherboards. You'll just have to make due, since we have nVME drives that are faster than main memory was ten years ago.