http://www.fudzilla.com/news/graphics/41420-nvidia-nvlink-2-0-arrives-in-ibm-servers-next-year
The writing really is on the wall for the end of SLI.
-
Competitor rules
Please remember that any mention of competitors, hinting at competitors or offering to provide details of competitors will result in an account suspension. The full rules can be found under the 'Terms and Rules' link in the bottom right corner of your screen. Just don't mention competitors in any way, shape or form and you'll be OK.
Nvidia NVLINK 2.0 arrives in IBM servers next year
- Thread starter Kaapstad
- Start date
http://www.fudzilla.com/news/graphics/41420-nvidia-nvlink-2-0-arrives-in-ibm-servers-next-year
The writing really is on the wall for the end of SLI.
Caporegime
- Joined
- 18 Oct 2002
- Posts
- 33,188
This has precisely nothing at all to do with SLI. It's also got very little to do with desktop, it categorically will never be used to provide more/faster access to system memory so not a huge amount of use in a desktop system. the only way it would be seen on desktop is like a pci-e switch as you get now.
NVLink isn't even new, it's just being branded as such. Before NVlink IBM already had their own version of it, which is just a higher power slightly upgraded version of pci-e, IBMs and Nvidia's is built on top of pci-e. The difference is you build a server with specific hardware in mind, legacy is not required, so changing the pci-e spec as you see fit and without regard to the tdp of the cpu, pci-e compatibility and usability within desktop hardware aren't there.
If all desktops had unlimited TDP, no one cared about it, cost reductions and cost of motherboards wasn't a requirement then they could make pci-e run 200GB/s too, there is just no need. For the majority of desktops you need a balance between power used and performance required, 64x pci-e lanes on a mainstream cpu isn't worthwhile as 0.0001% of users might want to use 4 graphics cards for gpgpu work but the other 99.9999% of users get a more expensive cpu that burns more power and more expensive motherboards for almost no reason when those 0.0001% can just buy themselves a professional system with more power/pci-e lanes and better motherboards.
NVLink isn't even new, it's just being branded as such. Before NVlink IBM already had their own version of it, which is just a higher power slightly upgraded version of pci-e, IBMs and Nvidia's is built on top of pci-e. The difference is you build a server with specific hardware in mind, legacy is not required, so changing the pci-e spec as you see fit and without regard to the tdp of the cpu, pci-e compatibility and usability within desktop hardware aren't there.
If all desktops had unlimited TDP, no one cared about it, cost reductions and cost of motherboards wasn't a requirement then they could make pci-e run 200GB/s too, there is just no need. For the majority of desktops you need a balance between power used and performance required, 64x pci-e lanes on a mainstream cpu isn't worthwhile as 0.0001% of users might want to use 4 graphics cards for gpgpu work but the other 99.9999% of users get a more expensive cpu that burns more power and more expensive motherboards for almost no reason when those 0.0001% can just buy themselves a professional system with more power/pci-e lanes and better motherboards.
Do you have anything useful to add now you have got your standard NVidia rant out of the way.
PCI-E including 4.0 is a dinosaur when it comes to connecting components. You only have to look at the size of the slot for starters.
PCI-E will be replaced by something smaller better and faster whether you like it or not.
Caporegime
- Joined
- 18 Oct 2002
- Posts
- 33,188
Dear lord, one, I didn't rant about Nvidia, you incorrectly(as usual) claimed this was somehow the end of SLI, I said it wasn't... how scathing of me to say this wasn't the end of SLI, how will Nvidia recover from such a hurtful comment.
Second, PCI-E isn't a dinosaur, you don't know what you're talking about, pci-e isn't small or fast or big, pci-e as you would call it, is the controller within the cpu, it's not the slot. The slot is just a standardised way agreed for everyone to work together to build their own cards that they know can fit into any consumer motherboard.
Nvlink uses a different connector because it's not going in consumer motherboards, it's going in systems mounted in racks, they are far bigger and they have space to bolt them all down flat into a motherboard. This isn't an option for consumer desktop computers at all. The nvlink is also a controller built into the CPU, the connection is nearly irrelevant. You can make an NVlink card with a slot like a pci-e slot, and you can make a pci-e card with a custom connection that looks identical to an NVlink one. They are connectors, they are literally just copper traces arranged in different ways. You're arguing pci-e is a dinosaur because it uses copper traces arranged one way over another, that is how ridiculous an argument you are making.
Again, NVlink is built on top of pci-e, as the IBM version prior to Nvlink also was.
Lastly, where did I champion pci-e, where did I say I want it forever or for that matter where did I say Nvlink sucked? I said it's designed for a different purpose. It's no different to bandwidth on different sized gpus. It would be a waste putting a 384bit gddrx 10gbps bus on a 1080, let alone a 960, which is why you know... Nvidia didn't. There is no point putting a 200GB/s Nvlink bus on a motherboard designed to run either IGP or for 99.9999% of users 2 or less GPUs, it wastes power, space, money for zero gain to the target market. In server you do want more bandwidth, you routinely use more than 2 gpus and the extra power, cost and space are completely acceptable.
There is nothing wrong with NVlink, it's just a interconnect designed for server usage, with server costs and power level and is completely unsuitable and more important, completely unneeded for consumer/desktop computing.
PCi-e will be replaced, not by Nvlink, not least because Nvlink uses pci-e, but because things get old over time and better things are developed. pci-e will eventually be replaced by an entirely new standard, not something built on top of it.
Second, PCI-E isn't a dinosaur, you don't know what you're talking about, pci-e isn't small or fast or big, pci-e as you would call it, is the controller within the cpu, it's not the slot. The slot is just a standardised way agreed for everyone to work together to build their own cards that they know can fit into any consumer motherboard.
Nvlink uses a different connector because it's not going in consumer motherboards, it's going in systems mounted in racks, they are far bigger and they have space to bolt them all down flat into a motherboard. This isn't an option for consumer desktop computers at all. The nvlink is also a controller built into the CPU, the connection is nearly irrelevant. You can make an NVlink card with a slot like a pci-e slot, and you can make a pci-e card with a custom connection that looks identical to an NVlink one. They are connectors, they are literally just copper traces arranged in different ways. You're arguing pci-e is a dinosaur because it uses copper traces arranged one way over another, that is how ridiculous an argument you are making.
Again, NVlink is built on top of pci-e, as the IBM version prior to Nvlink also was.
Lastly, where did I champion pci-e, where did I say I want it forever or for that matter where did I say Nvlink sucked? I said it's designed for a different purpose. It's no different to bandwidth on different sized gpus. It would be a waste putting a 384bit gddrx 10gbps bus on a 1080, let alone a 960, which is why you know... Nvidia didn't. There is no point putting a 200GB/s Nvlink bus on a motherboard designed to run either IGP or for 99.9999% of users 2 or less GPUs, it wastes power, space, money for zero gain to the target market. In server you do want more bandwidth, you routinely use more than 2 gpus and the extra power, cost and space are completely acceptable.
There is nothing wrong with NVlink, it's just a interconnect designed for server usage, with server costs and power level and is completely unsuitable and more important, completely unneeded for consumer/desktop computing.
PCi-e will be replaced, not by Nvlink, not least because Nvlink uses pci-e, but because things get old over time and better things are developed. pci-e will eventually be replaced by an entirely new standard, not something built on top of it.
Soldato
- Joined
- 17 Aug 2003
- Posts
- 20,167
- Location
- Woburn Sand Dunes
lolkaap.
PCI-E or PCI express (Peripheral Component Interconnect Express) is a standard (actually a high-speed serial computer expansion bus standard), it is certainly not just the controller in the CPU, but your right in saying it is not just the slot either.
Soldato
- Joined
- 19 Feb 2007
- Posts
- 15,556
- Location
- Northampton
Pretty cool but we'll never see this on desktops, Just faster iterations of PCI-E.
Caporegime
- Joined
- 18 Oct 2002
- Posts
- 33,188
The slot is just traces, pci-e isn't tied to that slot, it's not dependent on that slot design, it's just the connector they chose. But it's also not a bad design, the size of a slot makes it bad... seriously Kaap just says some really odd stuff to shout at certain people, me included.
If you change the controller and the coding of the message nothing using the pci-e standard would be compatible with that controller, if you roll up that signal in something else, send it over a serial bus, and decode it at the other end you'd still have pci-e.
I mean, look at thunderbolt, or... honestly I keep forgetting which stupid standard is which, the Intel one, you can support pci-e over a external cable. pci-e on a motherboard is just a standardised way to allow internal pci-e connections so companies know in advance how to connect their device in a standard way they know will be available by the time their device is finished and available. pci-e isn't at all dependent on a motherboard pci-e slot to work as show by the things that support pci-e over an external cable.
Last edited: