-
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.
Pascal To Support 8 Way Sli - (4x Dual GPU)
- Thread starter africanos23
- Start date
Caporegime
- Joined
- 18 Oct 2002
- Posts
- 33,188
They'll scale well..... in the environment in which they'll be available, extremely, EXTREMELY expensive IBM based HPC machines.
Nvlink is unlikely to ever be used on a desktop motherboard. Not really sure why they are calling it 8 way sli, in a super computer environment gpu's already effectively work, in different motherboards but all on the same problem linked up as one giant machine. It has no bearing on gaming as much as IBM boxes to go within a supercomputer has no bearing on gaming. NVlink brings with it higher bandwidth, but so does pci-e 4. NVlink is a superset of pci-e and IBM had a previous incarnation that had nothing to do with NVlink, a IBM interconnect that again was a superset of pci-e. IBM cpu's have some pretty immense bandwidth and throughput, but at huge power usage.
Nvidia also on their whole "hey we could claim higher performance if we talk about fp16 flops rather than what everyone else uses... so lets keep doing that".
For gaming mixed precision is likely a draw back. Much like for every 64bit capable shader you increase power/die size usage over a 32bit one, adding in a complete 64/32/16bit capable shader unit means overhead. You'd almost certainly be able to fit in more solely 32bit shaders/compute units than either 64bit, mixed precision or 16bit that can be coupled together do do 32bit(often the way 64 bit is achieved).
Mixed precision makes for an architecture that scales better between mobile, desktop and professional(anything needing DP performance) but will almost certainly reduce overall 32bit throughout. Can we use 16bit for gaming... very possibly but we had years of AMD or Nvidia(mostly the latter IIRC) 'cheating' in benchmarks using optimisation to do something at 16bit instead of 32bit and speed it up. But gaming has effectively for a long time been primarily 32bit. Mobile uses 16bit more, a lot more, because it uses a little less power and in general the games are smaller and more limited, less complex.
Ultimately mixed precision(as standard on all shaders/compute units) definitely seems like it will help Nvidia rather than gamers.
Nvlink is unlikely to ever be used on a desktop motherboard. Not really sure why they are calling it 8 way sli, in a super computer environment gpu's already effectively work, in different motherboards but all on the same problem linked up as one giant machine. It has no bearing on gaming as much as IBM boxes to go within a supercomputer has no bearing on gaming. NVlink brings with it higher bandwidth, but so does pci-e 4. NVlink is a superset of pci-e and IBM had a previous incarnation that had nothing to do with NVlink, a IBM interconnect that again was a superset of pci-e. IBM cpu's have some pretty immense bandwidth and throughput, but at huge power usage.
Nvidia also on their whole "hey we could claim higher performance if we talk about fp16 flops rather than what everyone else uses... so lets keep doing that".
For gaming mixed precision is likely a draw back. Much like for every 64bit capable shader you increase power/die size usage over a 32bit one, adding in a complete 64/32/16bit capable shader unit means overhead. You'd almost certainly be able to fit in more solely 32bit shaders/compute units than either 64bit, mixed precision or 16bit that can be coupled together do do 32bit(often the way 64 bit is achieved).
Mixed precision makes for an architecture that scales better between mobile, desktop and professional(anything needing DP performance) but will almost certainly reduce overall 32bit throughout. Can we use 16bit for gaming... very possibly but we had years of AMD or Nvidia(mostly the latter IIRC) 'cheating' in benchmarks using optimisation to do something at 16bit instead of 32bit and speed it up. But gaming has effectively for a long time been primarily 32bit. Mobile uses 16bit more, a lot more, because it uses a little less power and in general the games are smaller and more limited, less complex.
Ultimately mixed precision(as standard on all shaders/compute units) definitely seems like it will help Nvidia rather than gamers.
Soldato
- Joined
- 5 Sep 2011
- Posts
- 12,881
- Location
- Surrey
lmao @ Gibbo pushing those 2000w Leadex units. Not selling many? 
Soldato
- Joined
- 18 Mar 2012
- Posts
- 5,492
- Location
- Eastbourne
8-Way SLI you say?
Your gonna need a bigger PSU
It just so happens OcUK has the worlds best PSU!!!
Hahah,
Soldato
- Joined
- 2 Jan 2012
- Posts
- 12,406
- Location
- UK.
In theory, there is nothing stopping a rendering engine from supporting any number of GPU's attached to a PCIe bus, when using Vulkan or DX12. With a large enough mobo, you could support 4 295x2 and have the rendering engine use all of them.
Beyond plugging them all in and sending work to them, its a matter of how well the bus can cope.
Plus i am sure NvLink is only being supported on PowerPC processors currently. so HPC only for now. nvm, just read that it can be used on mobo for gpu to gpu connections. with pcie sending commands and data to the gpu's from a cpu.
but im sure i read that Nvlink will be directly incorporated into PPC CPU's
Beyond plugging them all in and sending work to them, its a matter of how well the bus can cope.
Plus i am sure NvLink is only being supported on PowerPC processors currently. so HPC only for now. nvm, just read that it can be used on mobo for gpu to gpu connections. with pcie sending commands and data to the gpu's from a cpu.
but im sure i read that Nvlink will be directly incorporated into PPC CPU's
Last edited:
Caporegime
- Joined
- 18 Oct 2002
- Posts
- 33,188
Yup, the biggest gain from Pascal/nvlink will be when it's integrated on the cpu die and improves the speed of the link between cpu and gpu. Adding the link onto a mobo alone to let the gpu's talk to each other will still rely on the pci-e link/speed for data from cpu to each gpu. It will bring extremely limited benefit to gaming with a mobo only implementation. For gpgpu where something can run entirely in the gpu and just talk to each other somewhat bypassing the cpu, there could be more benefit but Nvidia don't have an architecture that allows the gpu to as easily work alone. HSA with memory and instruction queuing models basically allow that to happen, the gpu to be able to run as a cpu does.
With pci-e 4 coming along and providing 64GB/s as opposed to Nvlink's 80GB's, it makes it a touch pointless in terms of gaming/desktop usage.
Pascal/Nvlink's biggest performance gains will be when it's doing compute and ultra reliant on data throughput from the cpu and main system memory, an IBM processor(with more system bandwidth) and a higher bandwidth from cpu to gpu will artificially improve performance. Neither will provide a performance bonus to a gpu or even 3-4 gpu's when the pci-e bus isn't a bottleneck to begin with. It will also be ultra expensive mobos with extra chips specific to Nvidia... if they even bother desktop side will be interesting to see.
8-Way SLI you say?
Your gonna need a bigger PSU
It just so happens OcUK has the worlds best PSU!!!
SuperFlower Leadex Platinum '8Pack Edition' 2000W Fully Modular "80 Plus Platinum" Power Supply - Black @ £349.99 inc VAT
The collaboration of Super Flower with the world's number 1 overclocker, the multiple world record holder 'Ian "8Pack" Parry' has led to the worlds first consumer 2000W power supply on the market, offering extreme power without compromises, the 8Pack way! This Power supply can handle 4-way SLI & 4-way CrossfireX with ease, and with efficiency levels of up to 94% to boot.
All Leadex Platinum Power Supplies feature a fully modular design, 140mm silent DBB Fan, full Japanese 105°C capacitors, ECO Intelligent Thermal control system, a huge amount of connectors and single rail design.
For when you need the very best and biggest power supply on the market, the '8Pack' edition 2000W Superflower 80+ Platinum should be your first and only choice.</p>
Features
- 100% Full Modular Cabling Design For Easy Cable Management
- 80 PLUS ® Platinum Certified, 92% 94% 90% Efficiency At 20% 50% 100% Rated Load
- Comply With ATX 12V. V2.2, EPS 12V. V2.91 & SSI EPS 12V. V2.92 Specification
- Comply With Energy Star Standard & EuP (Energy Using Products) Regulation
- Over Voltage Protection / Over Current Protection / Over Power Protection / Short Circuit Protection / Under Voltage Protection / Over Temperature Protection
- Support All Intel/AMD Multi-Core CPU & ATX/BTX Compatible Motherboards
- Support NVIDIA® SLI™ & ATI CrossFireX™ High Performance Graphic Card
- More than 2,000 times On/Off test
- Unleaded Green Materials with RoHS/WEEE Compliance
- Perfect For Professional & High Performance Gaming System
Specifications:
- Dimensions: 150 x 86 x 220 mm (W x H x D)
- Fan: 140 mm DBB(automatic control)
- Colour: Black (power supply, fan)
- Fully Modular Design
- Efficiency: max. 94% (230V), 80 Plus Platinum certificate
- Form Factor: ATX12V 2.2 / EPS12V 2.92
- Power: 2000W
- Active PFC
- 3.3V: 24 A
- + 5V: 24 A
- + 3.3V & + 5V combined: 120 W
- + 12V: 1999.2 W / 166.6 A
- -12V: 6 W / 0.5 A
- + 5VSB: 15 W / 3.0 A
- Safety: cTUVus/TUV/CB/CE/FCC/C-Tick/BSMI/RoHS/CCC
- Partcode: SF-2000F14HP
Connectors:
- 1 x 20+4-Pin ATX12V/EPS12V
- 1 x 8-Pin EPS12V
- 1 x 4+4-Pin ATX12V/EPS12V
- 9 x 6+2-Pin-PCIe
- 5 x 6-pin PCIe
- 18 x SATA
- 5 x 4-Pin-Molex
- 2 x Floppy
- Warranty: 5yr
Only £349.99 inc VAT.
ORDER NOW
Do you sell mini nuclear generator to go with that?
Or a solar farm?
Kaap and some other posters are going to need all the juice they can get next year or so... perhaps a good market to get into for you guys
I sure they said next year 2016.
You should be okish depending on your system with 2 good 1200W PSUs for stock . If your looking to OC probably 2x 8Packs.
You need to be Onassis for this hobby
You say that again
You're right, 2016. I'm behind
A stop gap card isn't a year though, it's more like a few weeks or a month tops!
You're right, 2016. I'm behind
A stop gap card isn't a year though, it's more like a few weeks or a month tops!
For the punters with fat wallets may be.
Where are the individual link speeds? I can't see them in the article, though i could be blind, it's 8am and my eyes are still slightly crusty
PCI-E 4 isn't out yet but is apparently going to be: 31.51 GB/s (16 GT/s)
Last edited: