Having lived in azn land for about a year now one thing can be said: the computing power is the same here as it is in any other developed metropolis. Yes, I know you’ve seen live-action movies and even some anime that suggests these time zones are super special, but the Pentiums work the same here as they do in San Diego and London.
And since I’m so bad at predictions I thought I’d make one regarding gizmos.
This is a 10 year prediction. That a ATX-sized desktop will consume the same amount of power yet be able to compute 10 teraflops (double precision) of combined general-computing (CPU + GPU); furthermore, the form factor will also hold a 20 TB SSD running at SATA 3.0 transfer rates. The system will include at least 128 gigabytes of memory running at GDDR 5 equivalent speeds/timings (this assumes that the GPU has moved to the same package as the CPU). As far as networking, the desktop will have a 100 GB ethernet female connector.
And for the record, I’m not pulling a Kurzweil because he likes to accelerate things.
For instance, looking back 10 years ago the fastest mainstream workstation you could put together included:
Two Pentinum II’s at 450 mhz. Because these did not have any of the cool SIMDs from the SSE family (MMX doesn’t really count), the dual-CPU solution is only able to achieve a maximum theoretical 900 megaflops. 1 flop per cycle, times two cores.
Fast forward to today, for the same price (actually less when adjusted for inflation) you can put together a dual socket solution that includes 8 cores altogether. Eight 3.2 ghz Penryn cores will clock in at around 40 gigaflops.
Nevermind the jump in bus transfer rates (at least a 16x improvement) or the increased bandwidth in system RAM (100 mhz SDRAM has a bandwidth of 6.4 Gbit/s versus DDR3 1600 which pumps through 204.8 Gbit/s).
The modern day professional workstation also boasts multiple GPUs from nVidia and ATI. The new 280 and 4780 purportedly compute at around 1 teraFLOPS (the 280 is just under that, the 4780 is just over it). The penalty for double-precision calculations is about 60%. Oh, and you can string 4 of these bad boys together as seen in the University of Antwerp FASTRA.
10 years ago the best thing you could buy off the shelf was a 12 MB Voodoo2. It’s difficult to really gauge just how many FLOPS it could perform (primarily because the pipeline was non-programmable unlike modern GPGPUs), but lets be nice and say its 90 mhz core could pump out 90 megaFLOPS. In SLI mode you could put a max of two together.
So 180 megaFLOPS versus 4 teraFLOPS. You also realize that we’ve skipped over the entire giga range, right?
While it could be argued that it costs a considerably more amount of money to put together a quad-GPU system (around $2000 or so), the performance differential on this one metric alone is around 20,000x.
That’s huge, especially since a Voodoo-based solution still required a 2D card of some sort. Remember, this is May 1998 we’re talking about — the original TNT didn’t come out till the end of that summer. In fact, it was that fall that my friends and I each built the venerable Celeron/BH6 overclock combo.
Anyways, while Kurzweil likes to turn technology into an exponentially growing curve, I would like to take the more conservative approach and simply multiply the best consumer tech of today by the differentials based on the 10 previous years.
The typical hard drive back then was around 20 gig (mine was a mere 6.4 in October 1998). It cruised without a buffer and could transfer at bursts of 33 MB/s.
Today you can grab a SATA-based 1 TB drive with a 16 MB buffer with bursts at around 300 MB/s.
While regulatory agencies can screw up wireless development, ten years ago 802.11 allowed you to surf the interweb at a blistering 2 mbit/s. Today 802.11n put some MIMO engines under the hood allowing you to theoretically hit 248 mbit/s.
I’ll be honest and say I doubt that there will be a 100x fold increase in wireless bandwidth for the same 100m range, but I wouldn’t put 10-20 out of the question (just look at Bluetooth 3.0 and Wireless USB rates of 480 mbits and 2.5 gbit/s).
So, if you carry the 7 and factor the 3 you get the following for May 2018:
Single-precision CPU: 900 megaFLOPS to 40 gigaFLOPS turns into 44x. So around 1.75 teraFLOPS a decade from now.
Single-precision GPU: 180 megaFLOPS to 4 teraFLOPS turns into 22222x. So approximately a gajillion FLOPS. I do not think this scaling will occur (it would turn into 88 exaFLOPS which is a tad unrealistic). A single-digit petaFLOPS range might not be out of the question…
System RAM: 512 MB single-channel SDRAM to 16 GB dual-channel DDR3 turns into a gain of 32x. That gain in a decade is almost believable if designers fully merge the GPU onto the same package with the CPU (and thus pooling the memory together). So around 512 GB of memory.
Hard drive: Two 20 GB to two 1 TB drives turns into 50x. This is only limited by the ability to fabricate logic gates smaller (Intel just announced the construction of a 32 Gb NAND module at a 34 nm process which can help them slap a bunch of these in a standard 2.5″ form factor). Who knows what kind of other storage medium will be used at that time (holographic?). And this is becoming increasingly less important with cloud storage and large removable drives.
Soundcard: back in the day you wanted to grab a Creative-based 3D positioning card or something with that hyped Aureal chip. With Vista, it can all be done with software, so you really don’t even need the latest X-Fi and I doubt that in 10 years this will be much of a retail market.
Physics: didn’t exist 10 years ago. Ageia was bought out by nVidia and based on every report about the upcoming 280 and 4870 the PPU has moved entirely back onto the main card. DirectX 11 is supposed to natively add physics routines to the shaders as well. I don’t think the PPU will leave the card GPU ten years from now.
Networking: as I mentioned above, wireless could easily jump another factor of 10-20. Wired lines in 1998 were still stuck at 100 Mbit/s. I remember the day in highschool when a guy from a large CAT-wire company showed up and talked about their copper-based gigabit solution. I laughed and thought fiber would rule the day. This of course hasn’t taken place as there are numerous 10 Gb copper solutions for blade servers and workstations. I’d like to think that in 10 years we’ll have moved onto fiber 100 Gb ethernet connections, perhaps even terabit if someone can miniaturize the multiplexing mirrors effectively (right now those intercontinental speeds you hear about in the news are dozens of 10 and 100 Gb streams combined into one).
Of course, all of this could be for naught if software doesn’t scale with the hardware. Perhaps Windows 9 will be out. Maybe most of our gadgets will really just be thin clients connected to the cloud (though I doubt hard core gamers, graphic artists or CAD specialists would be able to go that route).
For the typical consumer, barring the apocalypse, I will predict that engineers will be able to cram the most powerful consumer workstation today into a iPhone-like form factor in 2018. I don’t think that is too unfathomable considering I am a super duper expert on “room at the bottom.”
Actually, in six more months my prediction will look more realistic because a dual-socket octo Nehalem solution will run 16 physical cores (plus another 16-”virtual” ones through Hyperthreading) each of which are more powerful than the Penryn versions. Larrabee and other discrete GPGPUs will continue to develop at their fast paced rates.
The only thing that probably won’t develop at such drastic rates will be the landscape here in azn land. Japan, Korea or Taiwan won’t have any floating buildings or flying cars. Of course China will, because it will become the hyped tech landmass of the future once they fix that nasty pollution problem.
See also:
What a difference 36 years make
GPU versatility
Seth Lloyd’s Million Megahertz CPU
What is wrong with Moore’s Law?
Specialization, Centralization, and the Future of Chip Integration
What do Botnets and GPGPUs have in common?
Intel Has a Small Urethra
FLOPS, MIPS, Watts and the Human Brain
10 years later, where are they now?