Rethinking Computer Design For an Optical World

holy_calamity writes "Technology Review looks at how some traditions of computer architecture are up for grabs with the arrival of optical interconnects like Intel's 50Gbps link unveiled last week. The extra speed makes it possible to consider moving a server's RAM a few feet from its CPUs to aid cooling and moving memory and computational power to peripherals like laptop docks and monitors."

Comments

DRM

[vlm]

moving memory and computational power to peripherals like ... monitors.

They mean ever more complicated DRM. Like sending the raw stream to the monitor to be decoded there.

Re:dumb monitor

[jack2000]

So you can buy a new monitor again, and again and again. I bet this is what went through Steve Jobs' head when he they made macs hard to upgrade, that and a huge thunder of Ka-ching Ka-ching Ka-ching Ka-ching Ka-ching Ka-ching ...

Re:dumb monitor

[bsDaemon]

you mean like an imac? /ducks (disclaimer: typed from a 24" imac while at work)

Here we go again

[overshoot]

This is eerily reminiscent of Intel's flirtation with Rambus: they were so focused on bandwidth that they sacrificed latency to get it. Yeah, the Pentium4 series racked up impressive GHz numbers but the actual performance lagged because the insanely deep Rambus-optimized pipeline stalled all the time waiting for the first byte of a cache miss to arrive.

Same goes for optical interconnect to memory: the flood may be Biblical when it arrives, but while waiting for it to arrive the processor isn't doing anything useful.

Now, peripherals are another matter. But if bandwidth were all it took, we'd be using 10 Gb/s PCI Express for memory right now.

Re:Here we go again

[demonbug]

This is eerily reminiscent of Intel's flirtation with Rambus: they were so focused on bandwidth that they sacrificed latency to get it. Yeah, the Pentium4 series racked up impressive GHz numbers but the actual performance lagged because the insanely deep Rambus-optimized pipeline stalled all the time waiting for the first byte of a cache miss to arrive.

Same goes for optical interconnect to memory: the flood may be Biblical when it arrives, but while waiting for it to arrive the processor isn't doing anything useful.

Now, peripherals are another matter. But if bandwidth were all it took, we'd be using 10 Gb/s PCI Express for memory right now.

I was thinking the same thing regarding latency and remote memory. If you've got your memory 1 physical meter away, you're already looking at something like 6.6 ns round-trip latency (in a vacuum) just for light traveling that physical distance; seems like once you include switching plus getting to/from the optical interconnect you're looking at some pretty serious latency issues compared to onboard RAM (I think DDR3 SDRAM is on the order of 7-9 ns).

Re:Here we go again

[chrb]

Same goes for optical interconnect to memory: the flood may be Biblical when it arrives

But it won't be - the system is fundamentally limited by all of the rest of the components. A top end front-side bus can already push 80Gb; scaling that upto the 400Gbit that this optical link promises will probably be practical within a few years, but the latency of encoding and decoding a laser signal and pushing it over several meters is going to be a killer for computational applications. It will be great for USBX, and for high end networking it will challenge Infiniband (which currently tops out at around 300Gb). Infiniband is already used for networking high-performance computational clusters, but nobody is using it for the CPU to memory bus because of the high latency. Even with high bandwidth, computation still has to be carried out on the data, and so it still makes sense to put the data and processor as close together as possible.

In the last decade there were many research papers proposing that co-processors would be placed on DRAM cards, or Embedded DRAM would allow CPU and processors to be fabricated on a single die (e.g. 1, 2). But if you have a processor and DRAM connected to similar units via an optical interconnnect, guess what - the architecture begins to look awfully similar to a regular network with optical ethernet. So, it looks likely that this will be just another incremental improvement in architecture rather than the radical shift that TFA envisions.

Re:Here we go again

[hackerjoe]

You people are not thinking nearly creative enough. The article doesn't make it clear why you'd want to move your memory farther away -- it would increase latency, yeah, but moreover, what are you going to put that close to the CPU? There isn't anything else competing for the space.

Here's a more interesting idea than just "outboard RAM": what if you replaced the RAM on a blade with a smaller but faster bank of cache memory, and for bulk memory had a giant federated memory bank that was shared by all the blades in an enclosure?

Think multi-hundred-CPU, modular, commodity servers instead of clusters.

Think taking two commodity servers, plugging their optical buses together, and getting something that behaves like a single machine with twice the resources. Seamless clustering handled at the hardware level, like SLI for computing instead of video if you want to make that analogy.

Minor complaint, the summary is a little misleading with units: they're advertising not 50 gigabits/s, but 50 gigabytes/s. Current i7 architectures already have substantially more memory bandwidth than this to local RAM, so the advantage is definitely communication distance here, not speed.

light speed lag leads to higher latency

[Chirs]

Without factoring in speed of light drops due to index of refraction changes, at a distance of 1 meter you're looking at latencies of 7 nanoseconds just for travel time. The bandwidth may be decent but the latency is going to be an issue for any significant distance.

The 1990s called.

[PPH]

They want their rats nest of cables back.

The extra speed makes it possible to consider moving a server's RAM a few feet from its CPUs to aid cooling and moving memory and computational power to peripherals like laptop docks and monitors.

Computer architecture must have the Bhudda-nature

[idontgno]

because this appears to be another aspect of Wheel of Reincarnation.

I'm old enough to remember a time where a computer was a series of bitty boxes tied together with cables. Then someone decided to integrate a lot of the stuff onto a motherboard, with just loosely-related stuff connected by cables to the motherboard. Then the loosely-related stuff got put into cards that plugged into the motherboard. Then that stuff just got integrated into the motherboard.

And now it's being reborn as stuff in bitty boxes connected together with cables.

I wonder what enlightement will be like, because karma appears to have been a bitch.

Re:Interesting, but...

[derGoldstein]

It would allow you to use components an a more modular way, especially around an office. If you're not big enough (of a company) to have dedicated rendering/encoding servers, you could move the GPU around depending on who's currently doing the work that requires it. Even on a more casual basis, you could have a bunch of laptops with mid-range GPUs, and have an external GPU for whomever if gaming at the moment. Just like people take turns in a household with the home-theater rig in the living room -- you don't need to install a huge LCD + amp + speaker system in every room, you just need to take turns.

Re:a few extra feet

[Sarten-X]

By my understanding, it's not so much the travel time as the decoding/switching/other electronic time. As one example, consider the switching time of a transistor/photodetector. The gate must collect enough energy to switch from "off" to "on". Increased speed means having fewer electrons enter the gate. Higher energy per electron means raising the voltage. That's why overclocking often involves fiddling with voltages. Unfortunately, with more voltage comes more induction, breakdown, and other headaches I don't know enough about to list.

In contrast, light is much simpler to work with. You can make a light beam brighter without affecting other beams much. There's little chance of a beam breaking through its cable. We can send higher energies to gates with ease. Higher energy means less time to switch, and faster operation.

Note that I am not a physicist, and not much of an electrical engineer. I may be entirely wrong.

Speed of whatever

[overshoot]

I don't think that light travels that much more quickly than electrons do.

Yes and no. In a vacuum, electrons aren't terribly useful unless you're driving them with a particle accelerator. In wires, electrons aren't really doing the work anyway: electrical signals effectively travel as waves in the dielectric surrounding the wires and in particular between signal pairs. In that case, the signal travels at around half the speed of light in a vacuum (faster if you use expensive insulation like Teflon, slower for other plastics.)

Light in optical fiber is also slowed by the refraction coefficient of the material and by path-length extension in multimode fiber. However, on balance it's a bit faster.

The real gotcha is that electrical signals at outrageous bandwidths suffer from some really horrible losses due to both skin effects on the wires and dielectric losses in the insulation. At 50 Gb/s and 30 cm, you're doing well to detect the resulting signal, never mind decode it. Worse, the losses are highly frequency-dependent, so you have to do all sorts of ugly things to pre- and post-condition the signal to make it usable. Some of this can be overcome by cranking up the transmit power, but then you get into that property of wires known as "antenna." All of that processing at both ends takes time, too.

Just not worth doing, generally.

Likewise, putting a bunch of streams out in parallel requires all sorts of cleverness to put the separate lanes together again after transmission skew. A single optical stream is much easier to use, sort of like the communications equivalent of Amdahl's Law.

Re:a few extra feet

[Mordok-DestroyerOfWo]

Note that I am not a physicist, and not much of an electrical engineer. I may be entirely wrong.

I'm not qualified enough to say whether you're right or wrong, but you stated your case eloquently and if there's one thing that Hollywood, politics, and Star Trek have taught me, sounding right is more important than being right.

Re:a few extra feet

[bennomatic]

Huzzah for the Internet-age realist and/or snarker. Nice complement, back-handed or otherwise.

Re:LightPeak

[somersault]

CPUs have high speed cache that is faster than the mainboard RAM for high speed processing on a set of data, and swap the cache to/from RAM as necessary (kind of like how you page RAM to your hard drive when you run out of RAM).

Such a small cache would be useless for GPUs though, so they need faster RAM to read the massive amounts of texture/vertex/shader/whatever data they have as quick as possible. They also benefit more from stuff like RAM that is optimised for high sequential read speeds, so it does make sense to use RAM that has been specially designed for GPUs if you actually care about graphics performance (I doubt most Mac Mini users do).

Re:a few extra feet

[smooth wombat]

Or, as our esteemed Professor Farnsworth remarked:

Yes, yes, anything with that many big words could easily be the solution.

Re:Speed limit

[vlm]

Now admittedly electricity usually only travels at about 0.5c, IIRC, but I think she was giving the speed-of-light delay, not the speed-of-electrons delay.

Don't confuse propagation velocity of electromagnetic waves, which depends on dielectric constant and is around 0.8c in normal conductors, with drift velocity of electrons which is maybe a meter per hour.

http://en.wikipedia.org/wiki/Speed_of_electricity

http://en.wikipedia.org/wiki/Drift_velocity

http://en.wikipedia.org/wiki/Velocity_of_propagation

Electrons really move slowly in metal. In a vacuum tube like a CRT, pretty quick.

Re:LightPeak

[Yvan256]

Most people don't want to mess around inside a computer case, just like most people don't want to mess with the engine of their car or truck, or with the insides of their televisions, etc.

Such a modular system would be similar to huge LEGO bricks, nothing to open up, just connect the bricks together. Hopefully they would make the modules in standard sizes and allow multiples of that standard size. A CPU module could be 2x2x2 units, optical drives could be 2x1x2, etc.

The system could allow to connect to at least four faces, so we don't end with with very tall or very wide stacks. Proper ventilation would be part of the standard unit size (you need more heatsinking than the aluminium casing allows? Make your product one unit bigger and put ventilation holes in the empty space). A standard material such as aluminium could be used so that machining/extruding could be used and would allow the modules to dissipate heat.

Re:LightPeak

[The Master Control P]

I recommend reading the programmer's guide to a modern graphics architecture; Caching is essential to them.

Modern GPU architectures face the same clock speed/bus speed disparity and memory latency problems as CPUs and have taken their response much farther. They have several thousand registers per core and an L1 size & speed cache per processor group. Cache misses carry a typical penalty of several hundred cycles.