Mega computing power up next.

[Wayne Morellini]

L.Langer posted about a workstations HPC manufacturer doing the sort of thing I have been wishing (to do with dealing with the bottle necks in PC architecture that greatly speed things up. But I stumbled across even better things coming in future.

17 exaflop optical computer by 2020 that is designed to replace GPU cards.

http://www.nextbigfuture.com/2014/08/startup-with-optical-solver.html

Note this IBM proposal for a one petaflop 10 liter computer is the same sort of thing I have been proposing, except using liquid to distribute power (I kid you not, all these things sound like April fools jokes but weren't written in April). I was proposing using such a system for quantum cellar automata that uses 1000'th the power and doesn't strictly need such a power system, but would be good) and using it on present tech, like below.

http://www.nextbigfuture.com/2013/10/ibms-goal-is-petaflop-computer-size-of.html

Finally, something I have been advocating and in Nvidia circles, seems to be coming true, but in much larger form. A 50k Nvidia Xavier 20 watt core 1 Megaw exaflop computer in 2018. The proposal was for a series of computers based on arrays of the latest industry leading Nvidia ARM chips, each with cuda array.

http://www.nextbigfuture.com/2016/11/50000-nvidia-xavier-chips-would-deliver.html

The proposal was for smaller sizes, like 200w-1kw up to 10kw towers for lower end research and commercial purposes. It went a bit further than the design above, actually using most chips stripped of excess functionality (actually I don't see them not doing that, the thought of just using standard designs) to provide more processing density and lower power fitted on as few dies as possible. So, 200 watt machine could house upto, then, 20 10 watt chips, a 1kw version 100 chips. Anyway, the Nvidia proposal is using the latest Xavier over 7 billion transistor 8 core, with 512 core cuda pascal gpu on chip at 20 watts. So 50 of these is over 25 thousand cuda cores, and 400 controlling processors. Doing the maths to chuck out what's not needed and rebalance the numbers and placements of the types of chip could lead to even higher performance (note, I have long advocated that GPU chips could be expanded to replace host processor functionality, in this case you simply would have less processors for control and mostly independent stripped down gpus (only a few have to retain display 3D and rendering functionality) which I call workstation processing units. Just one of these chips running Linux with container and x86 binary translation software would make on NLE, or 8k camera. This is on something like 14nm tech, and future Volta GPU version will likely hit sub 10nm.

Now, that Nvidia machine is likely to be made up of a lot of circuit boards likely in blades, but being able to purchase just one of those is likely to be more than enough for our purposes. This talk seems real, but I wonder why not the newer chip, I wonder if there is a special chip version for this.

 

[Antony Newman]

One of the big problems for 'ExaScale' systems is that many behave as though they are collections of a large number of computers linked by a (relatively) slow network.

Many companies predictions on exaflop capabilities in 2013 required that someone solve the energy consumption issues getting memory to-and-from CPUs, and raise the bandwidth 100 times between CPU ( or GPUs).

Nvidia's TOPs metric is starting point in Visual processing systems. If it is not a FLOP ... what sort of Operation is needed to performed by the computer?

In order for new 'visual' systems to greatly increase their efficiently (ie useful actions per watt) - we almost need a 'GPU-like' SoC dedicated for that task.

Intel are hoping that they will be able to offer this by directly connecting one of their CPUs to Mobileye IP on one piece of silicon. This will go some way to lower power requirements.

ARMs view of the world in 2 years is that the extra blocks of IP cannot be considered 2nd-class citizens; ARM want to rearchitect the new system Around a very high bandwidth bus in which the critical components can ramp up their interactions by 50 times without burning a hole in the power budget (DynamIq architecture).

The CCIX consortium is what many OEMs are betting on to stay relevant. Its is a Mid-speed bus the companies are hoping will be fast enough (and OPEN) that they will not be dealt out of future designs.

For the next innovations in PCs - perhaps they will be courtesy of the Auto industry?

Maybe Apples next MacPro will have a steering wheel ;-)

AJ

 

[Phil Holland]

The next relevant things coming down the pipe for "us", both of which are not this year will be PCIe 4.0 (experiencing delays) and DDR5 on track for 2018/2019.

NVMe has shown it's potential and I'd like to see a bigger focus there as well as larger (gasp) more affordable SSDs. Some of which we'll see May/June of this year.

Intel needs to look at it's "core" strategy very hard right now. AMD seems to be gearing up for a 16 core release and it's going to be rather affordable.

There is a very, very hard focus on mobile and notebooks right now, but we will still need forward thinking motherboards for workstations, which is where I feel things are lacking at the moment.

Nvidia has some plans to follow up it's latest generation of GPUs (wonder if there will be a mobile 1080 ti actually), but the Pascal based stuff is indeed kicking butt overall. Again, once PCIe 4.0 drops it's going to be a different game all together. The trick there is we might not even be using PCIe 4.0 if they delay it much longer.

We do indeed need more USB-C and Thunderbolt 3 adoption overall IMO. I show up at gigs and people look at me like I'm a time traveler from different distant future time. But hopefully more of this trickles out and is adopted quicker. Saves time/money at the end of the day for a lot of workflows.

 

[Wayne Morellini]

One of the big problems for 'ExaScale' systems is that many behave as though they are collections of a large number of computers linked by a (relatively) slow network.

Many companies predictions on exaflop capabilities in 2013 required that someone solve the energy consumption issues getting memory to-and-from CPUs, and raise the bandwidth 100 times between CPU ( or GPUs).

Nvidia's TOPs metric is starting point in Visual processing systems. If it is not a FLOP ... what sort of Operation is needed to performed by the computer?

In order for new 'visual' systems to greatly increase their efficiently (ie useful actions per watt) - we almost need a 'GPU-like' SoC dedicated for that task.

Intel are hoping that they will be able to offer this by directly connecting one of their CPUs to Mobileye IP on one piece of silicon. This will go some way to lower power requirements.

ARMs view of the world in 2 years is that the extra blocks of IP cannot be considered 2nd-class citizens; ARM want to rearchitect the new system Around a very high bandwidth bus in which the critical components can ramp up their interactions by 50 times without burning a hole in the power budget (DynamIq architecture).

The CCIX consortium is what many OEMs are betting on to stay relevant. Its is a Mid-speed bus the companies are hoping will be fast enough (and OPEN) that they will not be dealt out of future designs.

For the next innovations in PCs - perhaps they will be courtesy of the Auto industry?

Maybe Apples next MacPro will have a steering wheel ;-)

AJ

Lol, you could mount s steering wheel on the macpro and it would look like a game controller.

Yes even the old bus looked very fast compared to what we were seeing from the normal spagpu in those days. Yes, Dynamiq, and the new CPU architecture.

You did read my WPU proposal, stripping down unseeded bits of a gpu and enhancing it functionality, this is the basic need. Once you have one or two or so GPU cards, all the extra stuff becomes redundant for a lot of people but the base functionality still remains important for us. I'm going to say this potentially unwanted thing. Some people in the wide world could do with a cheap GPU card, even integrated, and put the money towards a WPU like card. You could even have a GPU card with mostly wpu cores, with the gpu providing basic performance.

The point of the above articles is not that we ourselves need exascale computers, but that even a fraction of that could be a lot better than what we have.


I should have mentioned, about the optical computer link. It may not suit us, but in a hybrid machine, some of those techniques used in parts of a circuit could come in handy.

 

[Stacey Spears]

The next relevant things coming down the pipe for "us", both of which are not this year will be PCIe 4.0 (experiencing delays) and DDR5 on track for 2018/2019.

.

NVLINK is a MUCH bigger pipe than PCIe 4.0. Just wish companies would support it. There is a Supermicro specialty board, but not much else. Here is the press release for the Supermicro.

 

[Wayne Morellini]

(Pss, Shss Stacey, he doesn't seem to realise it yet)

 

[Wayne Morellini]

So Stacey, how is that useful for us. Is it possible to do a cheap version with 1080 ti?