Slashdot Mirror
Nvidia's Kal-El Tegra Will Have Fifth "Companion Core"
Blacklaw writes with an article in Thinq about the upcoming quad-core Tegra chipset. Quoting the article: "Nvidia has released a few technical details of its upcoming 'Kal-El' Tegra processor, including a secret it's done well to keep under its hat thus far: it's a five-core, not four-core, chip."
The fifth core will be clocked lower and is intended to let the system use little power without having to fully suspend. A few years ago Openmoko had a vaguely similar idea to include a microcontroller for low-resource idle tasks (e.g. GPS logging), but this design is superior since it should be more or less transparent to user space programs.
I thought Supergirl was Superman's cousin.
Oh, it's a chip. :)
I'm a "sidekick" core, not a "companion" core...
Or something about elephants...
The Enrichment Center reminds you that the Companion Core cannot speak. In the event that the Companion Core does speak, the Enrichment Center urges you to disregard its advice.
I hope it has a small pink heart image on the sillicon.
Last time I had a companion, I had to destroy it in an incinerator. I felt like a monster.
A fraction of the speed, and only one core, they should have codenamed it Clark.
Hopefully Nvidia will deign to support h.264 High Profile this time. Sure, the Tegra2 can play back 1080p happily at unreasonably high bitrates (for something you'd watch on a phone), but only if you don't use weighted p/b frames or CABAC when encoding. Guess what the majority of video you'll find in h.264 uses? It's a real glaring omission.
we call him "comrade core"
This sig is not paradoxical or ironic.
I don't even have 4 cores in my main PC, what am I going to do with 4 cores on a phone? The companion core is an interesting idea to increase battery life. But I have the feeling that as soon as the 4 main cores kick in, I would be left with a dead battery and burns on my hands. I also wonder how smooth the transitions between the companion core and the main cores will be...
Nigel: ...the chips have five cores. Look...right across the board. ...and most of these chips go up to four.... ...all the way up. You're on four on your processes...where can you go from there? Where? ...these have five.
Marty: Ahh...oh, I see....
Nigel: five...five...five...
Marty:
Nigel: Exactly.
Marty: Does that mean it's...faster? Is it any faster?
Nigel: Well, it's one faster, isn't it? It's not four. You see, most...most blokes, you know, will be running on four. You're on four here...all the way up...all the way up....
Marty: Yeah....
Nigel:
Marty: I don't know....
Nigel: Nowhere. Exactly. What we do is if we need that extra...push over the cliff...you know what we do?
Marty: Put it up to five.
Nigel: Five. Exactly. One faster.
Marty: Why don't you just make four faster and make faster be the top number...and make that a little faster?
Nigel:
Obligatory:
http://www.theonion.com/articles/fuck-everything-were-doing-five-blades,11056/
Seriously, a low-performance core doing administrivia type work sounds great, but won't this require OS support? I can't imagine this detail is completely abstracted from the kernel.
Make sure everyone's vote counts: Verified Voting
even the "dual-core" tegra2 had a companion core. it's hard to say that this extra management core is a real core, since it's not a peer of the others in, for instance, cache-coherency.
still, sure, asymmetric cores are a nice way to take further advantage of extreme variance in load. even after you've downclocked a normal core as far as it can go, a "designed for slow" core is going to dissipate less power. I'm not sure why supporting this kind of asymmetry would be all that hard for the linux kernel, though.
My first question was, "Why the hell not just underclock an existing core, like my existing phone does?" (from 1.1GHz under load down to 125MHz or 250MHz display-on "idle", depending what apps are open and/or displaying). I understand it's "special", optimized for power consumption at low speeds rather than high-speed capability, but I wouldn't expect huge gains over a standard core at the same clock, and it has to add a fair bit of price... I'm no expert, but it doesn't seem like the benefit would justify it
A possible answer, 20 seconds of thought later, was latency. Switching speeds (from low to high) typically burns a few milliseconds while the core does nothing useful. This can be quite annoying when the upclock-inducing load is in response to user interaction -- first it redraws or whatever slowly, because it's in a low-frequency state. Then it freezes for an instant as it changes clock speed, and then once it works through any event backlog, it finally runs smoothly.
Switching a core on takes less time, and background processes can continue execution on the weighted companion core -- I bet it reduces the perception of UI lag ~50% on a standard multi-core, plus the power gains from a low-leakage companion... Anyone know if this is really part of the reason, or am I just underestimating the benefits of a special low-leakage core alone?
Will it be a weighted companion core? I suppose it is, it's a lighter core than the others. They totally need to put a heart on the diagrams for it.
A Portal reference had to be made.
The Companion Core will never threaten to stab you, and, in fact, cannot speak.
frequency switching is fast, especially when you're switching by integer multiples of the memory clock speed. but dropping frequency gives little benefit compared to lowering voltage. but switching voltage is slow. and there is a limited range of voltages that a module may support.
“Common sense is not so common.” — Voltaire
I'm making a note here: HUGE SUCCESS!
Userspace. Userspace. Want to go to Userspace. Can we go to Userspace? Userspace. Look at me. I'm in Userspace. Userspace. Userspace. Userspaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaace. You know what's slow? You know what's low power? Userspace. ... Userspace. Want to go to Userspace. Userspace. Userspace. Userspace. Userspace. Userspace. Userspace. Userspace. Userspace. ... Userspace.
Check out my sci-fi book "Lacuna" at http://goo.gl/MVxX8
It's because the companion core uses different transistors. The four main processors use TSMC's 40nm general purpose (G) process while the companion core uses their 40nm low power (LP) process. (Though it's two different "processes" it's made on the same die, just with different transistor design).
To reach >1GHz for the main cores you have to use the faster but leakier and power-hungrier transistor design, so even if you underclock one of those cores to match the frequency of the companion core it'll still use a lot more power than the companion does.
I don't have any space programs. Oh, you mean user-space programs. All right then. Feel free to argue. While I watch. Mind if I slip into something lighter? Ah, okay, that's better.
(should have said this in my post: as you say, just tweaking the processor design doesn't lead to big enough gains to make this worthwhile, but the different transistor type is enough to make it worthwhile)
"Companion Core"
Is it weighted? I'm still suffering trauma from being forced to incinerate my last weighted companion. It was my only friend *sniffle*
Join the Slashcott! Stay away entirely Feb 10 thru Feb 17! Close all tabs to prevent autorefresh!
Their 'power saving' bar chart has gratuitously chopped off the bottom 20% of the graph.
Quattuor res in hoc mundo sanctae sunt: libri, liberi, libertas et liberalitas.
This is a nice development, and i'm sure it'll help at least a little. I have a Tegra2 based tablet, the Asus Transformer, and according to my battery stats most of the power is used to run the screen and WiFi, rather than the CPU. More efficient screens and WiFi would make a far bigger difference than a low power core.
My battery stats are: Screen 32%, Tablet Idle 22%, Wifi 19%, Android OS 14%, everything else is below 10%.
Yeah, I actually knew that, should have said "switching operating points", not "switching speeds".
Your pedantry is noted and appreciated. ;-)
"Nvidia boss Jen-Hsun Huang has stated that he aims to make Nvidia's Tegra the company's main focus, moving away from the discrete graphics that has been the company's bread and butter in the past."
Sounds like they'll balls up their graphics chips just to become another low power CPU firm. Bad plan. Bad for gamers. Bad for employees. Bad for everyone.
These posts express my own personal views, not those of my employer
^ ~
This was what we were discussing the other day another thread - how does the Tegra compare performance wise against the iCore5 and iCore7? Or against the Atom? The 4 or 5 cores gave me the impression that the performance was competitive w/, or exceeded Intel's.
Yes, Tegra 2 is also a tri-core. Two Cortex A9s, and an ARM7. If I recall, they were going to clock the ARM7 at a maximum of 100mhz (at which point it should use like 10 or 20mw.) This is smart from a power consumption standpoint. The Linux kernel (including the one in Android) does run "tickless" so there are no timer interrupts 100x or more a second like in older kernels. This helps the CPU power down entirely as much as possible. But if something DOES happen.. well, on my Droid 2 Global, the 1.2ghz ARM's minimum clock speed is 300mhz (and on the phones with 1.0ghz cores, it's 250mhz.) That is WAY more speed than is needed just to blink an LED, or store an incoming text, or whatever.
So is it called Kelex or Jimmy Olsen in the internal design docs?
"While it has been a faithful companion, your Companion Core cannot accompany you through the rest of the test."