This board uses the older 32-bit model CPU, 64 registers, 3-way VLIW, 5-deep pipeline, in-order, 64KB of L2 per core, 24KB of L1, cache coherence with an automatic subscription model and all the L2 cache combining to the effective L3 cache, 2 user-space mesh interconnect lines available between cores, full MMU with up to 64GB physical memory (still 32 bit, so 4GB per process), etc.
If your load isn't very computationally expensive, then you can have weaker cores running them "fast enough" at a fraction of the power, rather than a slice of a very fast machine churning through it very quickly while taking lots of power. It's all about performance ratios against power.
But I don't think the product's intent is talking along the same lines as people here, the latter namely being consolidating unused computing potential. Rather, it's more about getting machines taking less power while having the same effective computational capacity.
So run your big threaded app or VM array on N fast cores with X power consumption, vs getting the same (or greater!) overall effective speed on N+M slower cores at X-Y power consumption, as the promise has been for a long time. And systems like web servers and such can make do with tons of requests, each being handled on a slower core. And^2, the speed of the various individual low-power cores is nothing to sneeze at nowadays.
However, this is all pretty newly forged territory, so we'll have to see what this means in the real world at the moment.
Tilera describes their instruction set as "64-bit VLIW with 64-bit instruction bundle", so that doesn't sound quite like stock MIPS if they truly mean multiple independent parallel instructions per 64 bit instruction word. However, MIPS is also desirable for their low-power features, so their licensing might have been for something along those lines.
That 12 seconds does not include restarting all your applications, typing in ssh agent passwords, and all the other foo required to resume your normal working environment.
I've been googling around; their site is just a "Contact us" sort of sales deal, and no other hit seems to mention a price, like can sometimes be found in articles about headlining sales.
Last I looked at Tilera's offerings, their core count to memory controller ratio was really high. It seemed to be really focused on purely streaming data applications, like packet inspection or video conversion.
Anybody know if how Facebook is using them is actually memory-constrained, or it's just low power enough not to matter when 80 quadjillion requests are eventually handled quickly enough, regardless of actual latency?
How many people actually reboot their Linux systems? I guess if you're on a laptop you might sometimes, but I just use Sleep functionality instead of cycles.
Still, a good (even if by now esoteric) achievement.
I believe it's been measured that in times of economic struggle, people tend to engage in paid entertainment activities to a greater extent, from all slices of life.
Regarding the GP, I've recently rubbed elbows with people who Make Things Happen(tm), and while I'm not going to don any tinfoil hats over it, it's spookily applicable when all things are considered.
I agree with the sentiment about how it should help. I'm big into driving sims & practicing technique, and also thought this would be a legitimate benefit. The 3D was unnoticeable. It looks like the vanishing point of the road is about 5 inches behind the screen if you do focus on it, which really doesn't add anything in terms of perceiving the spatial configuration of the upcoming length of track.
I looked through the game options to no avail; not sure if the console itself has configuration for that, though I'd suspect it'd need to be game-specific depending on how each individual game made their geometry.
With two posts on this, I'm expecting somebody who owns it to jump on me and say how it's actually good/beneficial if you set it up properly, but it doesn't look like that's happening either.:-/
Mind you, I've only seen the 3D portions of Gran Turismo 5 and Sly 3, but each of those games only seemed to have a divergence of about 5 horizontal pixels onscreen between the 2 views even at the farthest Z-buffer depth. The actual 3D effect was incredibly understated and pointless. Sure as a graphics geek, I'm all for having superfluous 3D just for random kicks once in a while, but even from that end of things it did not deliver.
Every 3D game should have a configuration for adjusting the "strength" of the parallax divergence, especially as display sizes and other factors could benefit from them. Neither of those 2 games I tried seemed to have that at all. Trying to make a "safe" default divergence strength makes the gimmick effectively disappear.
(If I understand correctly, the 3DS has some sort of depth adjustment slider. Does it affect the rendering convergence, or just help focus at the hardware level?)
Um, no. I have two 200ppi 3840x2400 22" displays, plus a "normal" 30" 2560x1600 on my desk. That's 22.5 megapixels and taking up a fair amount of space, but I'd still take more pixels regardless of size. Oh, there's also a laptop typically stationed under the high-sitting 30" as well.
Unless you mean a single 30" at 300ppi, which would be more like 5-6k pixels horizontally, and I'd still want 3 on my desk.:)
the fact that those who did see it coming are few and tend to be drowned out by the din of reactive fools.
Unfortunately, a prevailing quote heard in Wall Street and Washington is IBGYBG: "I'll be gone, you'll be gone". They know what's up, and try to make a buck off it while attempting to kick the problem down the road for the next guy.
Too bad you won't find it working for the big, policy-burdened companies, either. Might as well go for that which is more interesting to work at, given that job stability was being called a myth at least 10 years ago and has only gotten worse.
And this is why it's very beneficial to go towards startups and small companies instead of large, established companies seeking some false "security" or "stability" notion. In a budding place, meritocracy tends to prevail, and tend to be desperate for people who would actually bring the company forward. Though you do run into dumb seniority politics, I think it's well worth the tradeoffs in work culture.
Note that a good portion of the worry about shipping the spent fuel around is that the rails themselves actually need to be upgraded to support the weight of how the nuclear cargo needs to be shipped. The standard lines can't handle it.
As far as I understand DHTs, they still need routing underneath. The identity at IP 1.2.3.4 is closer to your hash target than IP 1.2.3.5, so you choose to go talk to 1.2.3.4 even though the real node is down (say it was on 1.2.3.6). But you still have to have the ability to get to 1.2.3.4. And there are no locality guarantees about routability, just liveness & hash distance on top of the existing IP infrastructure.
I think if you try to make a DHT equivalent of routing, you'd quickly end up with distinct local minima that would never find each other.
Corrupt police are being caught, and convicted. It's getting reported. Sure, some holes might need to be closed in terms of accessing their system, or requesting permission to use it, but the fact that these guys aren't getting away with it is a very good thing.
Too bad physical reversible computing hardware isn't practical. You still need non-reversible guts outside of it to track state, and reversible hardware typically can't do things like arbitrary-length loops.
Right, most accidents are in more populated areas. It's a combination of not paying attention because the area or route is very familiar, and needing to get across a short distance quickly thus taking risks.
On longer trips, you're through unfamiliar territory, so you're going to be more conscientious about your surroundings and probably take things a little more carefully. However, an additional problem is monotony of sensory inputs across large empty areas and fatigue, but those are much more noticeable, so can be handled much better. Take a break, see some sights, stop at a restaurant, etc.
Slashdot Mirror
Ah, one of the comments on the article linked to the original paper: http://gigaom2.files.wordpress.com/2011/07/facebook-tilera-whitepaper.pdf
This board uses the older 32-bit model CPU, 64 registers, 3-way VLIW, 5-deep pipeline, in-order, 64KB of L2 per core, 24KB of L1, cache coherence with an automatic subscription model and all the L2 cache combining to the effective L3 cache, 2 user-space mesh interconnect lines available between cores, full MMU with up to 64GB physical memory (still 32 bit, so 4GB per process), etc.
If your load isn't very computationally expensive, then you can have weaker cores running them "fast enough" at a fraction of the power, rather than a slice of a very fast machine churning through it very quickly while taking lots of power. It's all about performance ratios against power.
But I don't think the product's intent is talking along the same lines as people here, the latter namely being consolidating unused computing potential. Rather, it's more about getting machines taking less power while having the same effective computational capacity.
So run your big threaded app or VM array on N fast cores with X power consumption, vs getting the same (or greater!) overall effective speed on N+M slower cores at X-Y power consumption, as the promise has been for a long time. And systems like web servers and such can make do with tons of requests, each being handled on a slower core. And^2, the speed of the various individual low-power cores is nothing to sneeze at nowadays.
However, this is all pretty newly forged territory, so we'll have to see what this means in the real world at the moment.
Tilera describes their instruction set as "64-bit VLIW with 64-bit instruction bundle", so that doesn't sound quite like stock MIPS if they truly mean multiple independent parallel instructions per 64 bit instruction word. However, MIPS is also desirable for their low-power features, so their licensing might have been for something along those lines.
That 12 seconds does not include restarting all your applications, typing in ssh agent passwords, and all the other foo required to resume your normal working environment.
I've been googling around; their site is just a "Contact us" sort of sales deal, and no other hit seems to mention a price, like can sometimes be found in articles about headlining sales.
Anybody in the know here?
Last I looked at Tilera's offerings, their core count to memory controller ratio was really high. It seemed to be really focused on purely streaming data applications, like packet inspection or video conversion.
Anybody know if how Facebook is using them is actually memory-constrained, or it's just low power enough not to matter when 80 quadjillion requests are eventually handled quickly enough, regardless of actual latency?
Also, lack of ARM is disappointing.
How many people actually reboot their Linux systems? I guess if you're on a laptop you might sometimes, but I just use Sleep functionality instead of cycles.
Still, a good (even if by now esoteric) achievement.
I believe it's been measured that in times of economic struggle, people tend to engage in paid entertainment activities to a greater extent, from all slices of life.
Regarding the GP, I've recently rubbed elbows with people who Make Things Happen(tm), and while I'm not going to don any tinfoil hats over it, it's spookily applicable when all things are considered.
Welcome to the club.
- USA
I agree with the sentiment about how it should help. I'm big into driving sims & practicing technique, and also thought this would be a legitimate benefit. The 3D was unnoticeable. It looks like the vanishing point of the road is about 5 inches behind the screen if you do focus on it, which really doesn't add anything in terms of perceiving the spatial configuration of the upcoming length of track.
I looked through the game options to no avail; not sure if the console itself has configuration for that, though I'd suspect it'd need to be game-specific depending on how each individual game made their geometry.
With two posts on this, I'm expecting somebody who owns it to jump on me and say how it's actually good/beneficial if you set it up properly, but it doesn't look like that's happening either. :-/
Mind you, I've only seen the 3D portions of Gran Turismo 5 and Sly 3, but each of those games only seemed to have a divergence of about 5 horizontal pixels onscreen between the 2 views even at the farthest Z-buffer depth. The actual 3D effect was incredibly understated and pointless. Sure as a graphics geek, I'm all for having superfluous 3D just for random kicks once in a while, but even from that end of things it did not deliver.
Every 3D game should have a configuration for adjusting the "strength" of the parallax divergence, especially as display sizes and other factors could benefit from them. Neither of those 2 games I tried seemed to have that at all. Trying to make a "safe" default divergence strength makes the gimmick effectively disappear.
(If I understand correctly, the 3DS has some sort of depth adjustment slider. Does it affect the rendering convergence, or just help focus at the hardware level?)
Um, no. I have two 200ppi 3840x2400 22" displays, plus a "normal" 30" 2560x1600 on my desk. That's 22.5 megapixels and taking up a fair amount of space, but I'd still take more pixels regardless of size. Oh, there's also a laptop typically stationed under the high-sitting 30" as well.
Unless you mean a single 30" at 300ppi, which would be more like 5-6k pixels horizontally, and I'd still want 3 on my desk. :)
Madness? This is VESTA!!!!!
the fact that those who did see it coming are few and tend to be drowned out by the din of reactive fools.
Unfortunately, a prevailing quote heard in Wall Street and Washington is IBGYBG: "I'll be gone, you'll be gone". They know what's up, and try to make a buck off it while attempting to kick the problem down the road for the next guy.
Too bad you won't find it working for the big, policy-burdened companies, either. Might as well go for that which is more interesting to work at, given that job stability was being called a myth at least 10 years ago and has only gotten worse.
If he's a great programmer, then no, it doesn't matter.
And this is why it's very beneficial to go towards startups and small companies instead of large, established companies seeking some false "security" or "stability" notion. In a budding place, meritocracy tends to prevail, and tend to be desperate for people who would actually bring the company forward. Though you do run into dumb seniority politics, I think it's well worth the tradeoffs in work culture.
Note that a good portion of the worry about shipping the spent fuel around is that the rails themselves actually need to be upgraded to support the weight of how the nuclear cargo needs to be shipped. The standard lines can't handle it.
You do realize that this has very similar contact characteristics to motorcycle tires, right? And they still work, both on pavement and dirt.
As far as I understand DHTs, they still need routing underneath. The identity at IP 1.2.3.4 is closer to your hash target than IP 1.2.3.5, so you choose to go talk to 1.2.3.4 even though the real node is down (say it was on 1.2.3.6). But you still have to have the ability to get to 1.2.3.4. And there are no locality guarantees about routability, just liveness & hash distance on top of the existing IP infrastructure.
I think if you try to make a DHT equivalent of routing, you'd quickly end up with distinct local minima that would never find each other.
But have you ever actually seen a business successfully happy with what resulted from getting in with SAP?
Yes, I was going to post something like this.
Corrupt police are being caught, and convicted. It's getting reported. Sure, some holes might need to be closed in terms of accessing their system, or requesting permission to use it, but the fact that these guys aren't getting away with it is a very good thing.
Not just a TCP packet, but a mass of XML!
Too bad physical reversible computing hardware isn't practical. You still need non-reversible guts outside of it to track state, and reversible hardware typically can't do things like arbitrary-length loops.
Right, most accidents are in more populated areas. It's a combination of not paying attention because the area or route is very familiar, and needing to get across a short distance quickly thus taking risks.
On longer trips, you're through unfamiliar territory, so you're going to be more conscientious about your surroundings and probably take things a little more carefully. However, an additional problem is monotony of sensory inputs across large empty areas and fatigue, but those are much more noticeable, so can be handled much better. Take a break, see some sights, stop at a restaurant, etc.