backtop

Hardware Samsung Makes Speedy GDDR4 Even Faster
Marcus Yam (Blog) - February 26, 2007 3:00 PM
Print E-mail del.icio.us 21 comment(s) - last by scrapsma54.. on Mar 1 at 7:16 PM
Samsung claims to have broken speed records with its new graphics memory

Samsung Electronics announced over the weekend that it has increased the data transfer speed on its extremely fast GDDR4 (Graphics Double Data Rate, version 4) graphics memory. The South Korean electronics giant claims that it has increased speeds by two-thirds.

Using 80-nanometer production technology, the 4Gb/s (2.0GHz) is 66 percent faster than today’s fastest commercially available memory – the 2.4Gb/s GDDR4. The new 4Gb/s graphics memory, offered in 512Mb density, has a 32-bit data bus configuration. GDDR4 uses JEDEC-approved standards for signal noise reduction to help attain the highest possible speed.

“Our new GDDR4 memory will add even more zip in video applications, making gaming, computer-aided design and video editing faster than ever before,” said Mueez Deen, marketing director, graphics memory, Samsung Semiconductor, Inc. “This will enable ultra-smooth movements in animation and make games incredibly realistic, resulting in a truly immersive user experience,” he added.

Samsung cites analyst expectations for GDDR4 to significantly boost demand for high-performance graphics memory over the next 12-18 months. The company said that it will begin customer sampling of its new chips this month.

Most video cards on the market today from NVIDIA and ATI are still using GDDR3, including solutions inside the Xbox 360 and PlayStation 3. GDDR4 memory thus far has seen limited use in recent high-end cards from ATI, and is expected to be paired with the upcoming R600 GPU.

Comments     Threshold


This article is over a month old, voting and posting comments is disabled
<no subject>
By Scabies on 2/26/2007 3:07:46 PM , Rating: 2
I knew there was logic in ATI/AMD waiting to start printing R600-based cards. I just hope this decreases bottom line PSU requirements.




RE: <no subject>
By FITCamaro on 2/26/2007 3:25:18 PM , Rating: 2
I really doubt it will help much since I don't think the memory is whats sucking up all the power on current GPUs. Its the several hundred million transistors on the GPU itself. Granted 512MB-1GB of graphics memory is a ton of transistors as well.


RE: <no subject>
By Calin on 2/28/2007 3:02:12 AM , Rating: 2
No matter how many billion transistors could be on the memory side of the things, the heat is produced on the chips that need the most cooling - that would be the graphics core itself.
On many cards that need active cooling for the graphics chip, memory chips don't even have passive heat sinks.


RE: <no subject>
By ajfink on 2/26/2007 3:49:02 PM , Rating: 1
There is also the rumor that AMD is waiting to have the entire R600 lineup on 65nm. If they waited to have it all on 65nm and this memory, I can understanding pushing it back to May. But, please, no more delays after this.


RE: <no subject>
By Dactyl on 2/26/2007 10:43:59 PM , Rating: 2
That's ridiculous. If AMD could have gotten their cards out sooner, they would have. The simple, obvious truth is that the silicon wasn't ready.

Maybe you think NVidia has a top secret strategy reason for delays in their Windows Vista/DX10 drivers?

Further, it's not hard to stick faster DDR4 memory on an existing DDR4-based card. They could call it the "X2900XTX XXX OC Edition" or something, and let their AIB partners release it as a special version.

There won't be a decrease in GPU requirements, because both kinds of DDR4 operate at the same voltages. Increasing the DDR's speed by 20% increases the power use by 20%.


RE: <no subject>
By InsaneScientist on 2/26/2007 11:03:45 PM , Rating: 2
quote:
Increasing the DDR's speed by 20% increases the power use by 20%.


Actually, as we saw with Prescott, power usage tends to go up geometrically with a linear progression in clock speed... so it's probably more than 20% more power for 20% more clock speed... But that just makes your point all the more relevant.


speed?
By lamestlamer on 2/26/2007 4:49:02 PM , Rating: 2
I don't get it. At 4 gb/s per module, the standard 8 modules for 512mbytes only yields 32 gb/s over a 256bit bus, which is quite slow in modern graphics throughput. Are you sure that isn't 4GB/s instead? That would be 256 gb/s which is very very fast.




RE: speed?
By WayneG on 2/26/2007 5:03:51 PM , Rating: 2
the latter is correct, yes :). The same speed as the eDRAM connected to the 360 but on a much larger scale and used for more than simple AA tasks and the likes.


RE: speed?
By FITCamaro on 2/26/2007 6:40:41 PM , Rating: 2
R600 also has a 512bit bus, not 256bit.


RE: speed?
By shimman on 2/26/2007 11:03:41 PM , Rating: 2
those byte & bit conventions are all messed up. samsung always use 'b' as byte when it comes to memory, so it is 4gbyte/sec/chip & i think the chip can be overclocked considerably. question is how much samsung is charging for it as samsung is the only one making it that fast.

i really hope everyone is clear about whether b stands for bit (as usual convention) or byte (usually B) by spelling out unless it is bps which should be bits per sec


RE: speed?
By ET on 2/27/2007 11:10:14 AM , Rating: 2
Actually, 4Gb/s at 2GHz seems right. It means that each line can transfer 2 bits per clock, or a total of 4G bits per second. A 256 bit interface would then transfer data at 1Tb/s, or 128GB/s.

This is completely in line with current memory bandwidth. For example, the Radeon X1950 XTX has a 1GHz memory speed and 64GB/s bandwidth. An R600 with 1GB of such RAM will therefore have 256GB/s of bandwidth.


Latencies?
By lotharamious on 2/26/2007 3:19:30 PM , Rating: 2
Any idea what kind of latencies these chips have? I know that for graphics latencies aren't nearly as important as bandwidth, but it would be nice to know that the latencies were kind of low so that one day this technology could be ported well to main system RAM.




RE: Latencies?
By saratoga on 2/26/2007 9:51:17 PM , Rating: 4
GDDR latency is typically somewhere between "unimaginably high" and "astronomical".


R600?
By WayneG on 2/26/2007 4:27:32 PM , Rating: 2
One can only hope that this memory goes into the R600 but teh truth is that it is highly unlikely. Especially the rumor of the card being downscaled to 65nm is definitely not going to happen for quite some time due to the complexities involved. Low end cards are easier in this regard... Would be pretty awesome though if R600 came out with memory this fast, would completely destroy every single resolution that we could throw at it. Who needs crossfire? :D
We will probably only see this memory turn up at the end of this cards' life and so not get much out of it, maybe in a refresh at the end of the year just before R700 is delayed, i mean released.
(Eagerly anticipating the R600, I had enough money for it in December. Saving sucks! But as long as R600 is fast and comes with some fast memory like this then I don't really care :))

One thing that needs to be said is that this article explains that the RAM is "offered in 512Mb density", is that only one side of the graphics card or both? If it's only one side it means that it definitely won't be on R600XTX because of the 1GB memory that it is specced to have. It can only have 512mb or 1gb because of the 512bit bus nature. 512mb being limited to the XT models. Correct me if I'm wrong please :).




RE: R600?
By InsaneScientist on 2/26/2007 11:00:44 PM , Rating: 2
How much memory there is on a card is completely dependent on how many chips the manufacturer puts on.

Please note, it's not saying that they're 512 MB (in which case I would see why you think it's the whole card), they're saying it's 512 Mb.

When referring to GDDR memory densities, they're usually referred to with Megabit capacities rather than Megabyte.

I would assumme the main reason is marketing: it makes them look bigger. At a cursory glance, which would you think is bigger: 64MB or 512Mb?

Another point: GFX cards these days have a tendancy to have 8 GDDR chips on them (32bits per chip * 8 chips = 256 bit bus) so it makes it real easy to tell how much memory a card using those particular chips will have. I.E. since these are 512Mb chips, a card equipped with 8 of them will have 512MB of memory.

If R600 is equipped with these things, it will have 1GB of memory. (512bit bus / 32 bits per chip = 16 chips * 512Mb/64MB = 1GB memory)
Of course... I think we already knew that R600 would have 1GB VRAM... but I'm using this to illustrate the point.

I hope that all made sense. Basically, what it boild down to is that they're referring to the size of the chip itself and not how much memory goes on one side of a board (after all... how much memory goes on 1 side of a board changes. The 8800 GTX has 12 memory chips, wheras many of the lower model graphics cards have 4 chips, or even two)


If I had a dime...
By UnFaZeD on 2/28/2007 12:02:17 AM , Rating: 2
for everytime someone said this...

“This will enable ultra-smooth movements in animation and make games incredibly realistic, resulting in a truly immersive user experience,”

My middle name would be Mint...




RE: If I had a dime...
By staypuff69 on 2/28/2007 5:07:49 AM , Rating: 2
wow you all amaze me... i had no idea so many people here created RAM technology that was 66% faster than previous... excuse me for not jumping on all of your critical bandwagon but none of you seemed to have gotten what was intended with this technology.... FASTER RAM

Why is it people seem to criticize things they have no idea what it's about.... here's a hint for all the doomsayers.... ANY ADVANCEMENT IN VIDEO CARD TECHNOLOGY = GOOD..... guess what something we've seen ohhhh about 9 million times..... it's gonna be expensive and not a lot of companies are gonna adopt it at first..... guess what people welcome to the world of computers......


ATI Delays?
By Mitch101 on 2/26/2007 3:38:34 PM , Rating: 3
Anyone think this might be the reason ATI pushed back the release date of some of the DX10 cards? Possibly to release a few with these chips?




By Kuroyama on 2/26/2007 3:50:06 PM , Rating: 2
In answer to my subject line, probably quite a few, since Samsung seems to claim credit for anything and everything these days. While I like their stuff, they are certainly not the only innovative company in the world, but a naive person could think from their flood of PR that within a few years all technology will come from them.




Marketing Director
By ADDAvenger on 2/26/2007 4:39:56 PM , Rating: 2
Mueez Deen, I wonder if his middle initial is an H... Now that would be a name for the "funniest real-life name you ever heard" thread.




Cas
By scrapsma54 on 3/1/2007 7:16:38 PM , Rating: 2
yet no mentioned attempt to fix the cas latencies. I really hope they fix that, because eventually it will bite everyone once Os's start going into the 32GB memory range.




"DailyTech is the best kept secret on the Internet." -- Larry Barber