Slashdot Mirror
Samsung to Produce Faster Graphics Memory
Samsung has announced a new line of GDDR5 chips that will supposedly be able to deliver data at speeds of up to 6 Gbps. In addition to faster data delivery the new chips also claim to consume less energy than previous versions. "Samsung said the new chips consume 1.5 volts, making them about 20 percent more efficient than GDDR 3 chips. Samples of the GDDR 5 chips began shipping to graphics-processor makers last month, and Samsung plans to begin mass production of the chips during the first half of next year. GDDR 5 memory should first appear in high-end gaming systems where users are willing to pay a premium for better graphics. Samsung did not disclose pricing for the chips.
Memo From Ki-Tae Lee
To: All Samsung Employees
CEO and President,
Samsung
December 3rd, 2007
Would someone tell me how this happened? We were the fucking vanguard of graphics memory in this country. Samsung's GDDR3 was on the card to own. Then the other guy came out with a GDDR3 graphics chip. Were we scared? Hell, no. Because we hit back with a little thing called XDR. That's GDDR3 on crack. For cokehead gamers. But you know what happened next? Shut up, I'm telling you what happened--the bastards went to GDDR4. Now we're standing around with our cocks in our hands, selling XDR & GDDR3. Cokehead gamers or no, suddenly we're the chumps. Well, fuck it. We're going to GDDR5.
Sure, we could go to GDDR4 next, like the competition. That seems like the logical thing to do. After all, three worked out pretty well, and four is the next number after three. So let's play it safe. Let's make a more crackhead gamer RAM and call it the XDR3SuperTurbo. Why innovate when we can follow? Oh, I know why: Because we're a business, that's why!
You think it's crazy? It is crazy. But I don't give a shit. From now on, we're the ones who have the speed in the memory game. Are they the best a man can get? Fuck, no. Samsung is the best a man can get.
What part of this don't you understand? If GDDR2 is good, and three is better, obviously five would make us the best fucking memory that ever existed. Comprende? We didn't claw our way to the top of the memory game by clinging to the GDDR2 industry standard. We got here by taking chances. Well, GDDR5 is the biggest chance of all.
Here's the report from Engineering. Someone put it in the bathroom: I want to wipe my ass with it. They don't tell me what to invent--I tell them. And I'm telling them to stick enough transistors on there to call it GDDR5. I don't care how. Make the chips so thin they're invisible. Put some on the handle. I don't care if they have to make the ram hang halfway off the motherboard, just do it!
You're taking the "safety" part of "safety electronics" too literally, grandma. Cut the strings and soar. Let's hit it. Let's roll. This is our chance to make memory history. Let's dream big. All you have to do is say that GDDR5 can happen, and it will happen. If you aren't on board, then fuck you. And if you're on the board, then fuck you and your father. Hey, if I'm the only one who'll take risks, I'm sure as hell happy to hog all the glory when the GDDR5 card becomes the gaming video card for the U.S. of "this is how we game now" A.
People said we couldn't go to three. It'll cost a fortune to manufacture, they said. Well, we did it. Now some egghead in a lab is screaming "Five's crazy?" Well, perhaps he'd be more comfortable in the labs at Sony, working on fucking electrics. Cell processing chips, my white ass!
Maybe I'm wrong. Maybe we should just ride in SanDisk's wake and make flash USB drives. Ha! Not on your fucking life! The day I shadow a penny-ante outfit like SanDisk is the day I leave the silicon game for good, and that won't happen until the day I die!
The market? Listen, we make the market. All we have to do is put her out there with a little jingle. It's as easy as, "Hey, shaving with anything less than GDDR5 is like playing Warcraft on a Commodore 64." Or "It'll be so smooth, I could snort lines off of your monitor." Try "Your frame rate is going to be so friggin' fluid, someone's gonna walk up and confuse it with a urinal."
I know what you're thinking now: What'll people say? Mew mew mew. Oh, no, what will people say?! Grow the fuck up. When you're on top, people talk. That's the price you pay for being on top. Which Samsung is, always has been, and forever shall be, Amen, GDDR5, sweet Jesus in heaven.
Stop. I just had a stroke of genius. Are you ready? Open your mouth, baby birds, cause Mama's about to drop you one sweet, fat nightcrawler. Here she comes: Put another microcontroller on that fucker, too. That's right. GDDR5, two microcontroll
Qimonda already released GDDR5 Article from November 2: http://www.pclaunches.com/other_stuff/qimonda_gddr5_memory_now_available.php
How does it consume volts?
This lends a bit of credence to the rumored NVidia G9 series launch, although I still think February is unlikely.
Demented But Determined.
Samsung said the new chips consume 1.5 volts, making them about 20 percent more efficient than GDDR 3 chips.
What poor science reporting. Nothing "consumes volts." Volts measure voltage -- difference in potential. Devices consume Joules -- units of energy. Also acceptable would be Watts -- energy per unit time. It would have been really nice to be given the Watts per Bandwidth per Size (W/Gbps/bits), but I realize that's asking way too much of the Times.
everything in moderation
That would be implied by the aggregate 24G bytes/sec later in the article. So I guess they are still keeping the 32 I/O pins that the previous generation used and not quite doubling the speed. The article is also missing the size, which is a spec that hardware designers frequently wish to see, but I think it's probably still 512 Mbit or we would have heard about it.
Intron: the portion of DNA which expresses nothing useful.
No. It will make it 20% faster.
No further comment.
The difference is we are talking about semiconductor devices. Losses from these semiconductor devices are primarily due to leakage and switching. As long as we're still using silicon, leakage will be roughly 0.5 V^2/R, no matter how much current you pump through the transistors. Switching losses occur in when logic changes from 1's to 0's due to the capacitance of the transistors. The power lost here is roughly 0.5 f C V^2, where f is the switching frequency and C is the capacitance (material dependent). The V^2's means that reducing the voltage has a significant impact on losses. If we note that R and C are completely determined by the material (silicon) and the fabrication process, we can see that as long as the frequency is held constant, the voltage is a reasonable metric for comparing power consumption in silicon devices.
Of course this analysis is purely approximate since there are a lot of there things going in the devices. And I'm assuming complete capacitive discharge (independent of switching frequency), and didn't consider the changes in refresh rate to this DRAM device. But suffice it to say voltage is still a pretty good metric for comparison (until you actually build the thing and test it).
This has been understood in the industry for decades: in a given silicon process, power consumption fits roughly within the envelop V^2 * F, where F represents frequency. Given a process shrink, this relation might or might not hold true. For a long time it was a good rule of thumb, but then in the era of excessively high leakage current it did not hold true, more recently with better control over leakage, the relationship is again a good rule of thumb. The upshot is that, over two decades, almost every reduction in voltage for a given class of part corresponds to a significant increase in power efficiency.
What the article failed to explain is this long history of voltage serving as a proxy for power efficiency.
The other relationship is that a given part will usually demonstrate a relationship where lower frequencies are stable at lower voltages. If increasing the voltage by 20% allows you to overclock a processor from 2GHz to 3GHz, you can estimate your increased power draw as 1.2^2 * 1.5, about double where you started.
It's almost pointless to convert this measure into watts, as so many other variables change in tandem. The new part has different bandwidth, different latency, different leakage, different dynamic consumption. There's no simple number that gets you apples vs apples. Most of the time, however, voltage is fair proxy. Peak consumption figures are mostly worthless from an efficiency perspective, except for sizing your power and cooling requirements.
On a side note, I'm wondering when we hit the floor on practical CMOS voltage levels. Surely the band-gap will come into play in the near future, and then what? Does the efficiency graph suddenly develop a crimp and stagger forward on a much reduced slope? This happened with hard drives, where there was a period of accelerated capacity increase (PRML/GMR/pixie-dust era) only to return to the more sedate curve once again later on. It wasn't long ago that F hit thin air (due to thermal issues) and now F is increasing at half the rate it sustained for a least a decade prior.
Long ago apparently respectable sources used to proclaim "silicon will hit the brick wall at 0.1um". In turns out S-curves hardly ever play out that way. The curve begins to taper downward when the easy gains are exhausted. The phrase "peak oil" is another one of those conceptual nightmares, much like the chimeric brick-wall on photo lithography. It's not going to be a peak, is it? It's going to be a wavy plateau. On any particular graph, you can point to a "peak" (though none of the graphs will agree), it's just that there won't be a momentous Alderan-disturbance that ripples though planet earth as the precocious metaphor suggests. Much like the silicon people had to finally confess, driving F higher and higher as your primary performance metric (at the cost of absolute efficiency) makes about as much sense in the long run as a single-occupancy air-conditioned Hummer in rush hour traffic.
Speaking of which, engine displacement is roughly as fair as a measure in the automotive sector as voltage in silicon. It's the nature of the internal combustion engine that these engines are far from their peak efficiency at low to medium throttle, which is why having a lot of power you rarely use is no free lunch. If you accept that a typical 2 liter engine is more efficient than a typical 3 liter engine, why would voltage as a proxy for power be any different?