Samsung joins Qimonda and Hynix in the GDDR5 race

Tuesday 04th December 2007, 01:45:00 PM, written by Arun

Samsung has announced the company's first GDDR5 chip, and while its capacity only reaches 512Mbit, its effective memory clock reaches an astonishing 6GHz (versus 2.4GHz for the GDDR4 on the Radeon HD3870). This news comes 2 weeks after Hynix's announcement, and one month after Qimonda's.

Data

  • Samsung claims 6GHz for its highest-end GDDR5, while Hynix and Qimonda both claim 5GHz. Samples from all three manufacturers have been shipping since last month.
  • Hynix is using a 66nm process, while Samsung is on a slightly more advanced 60nm process. Qimonda looks to be on 75nm, but will presumably transition to 58nm sooner rather than later.
  • Power consumption is claimed to be 20% lower than GDDR3 by Samsung. We presume this means 1GHz+ GDDR3, so it's still higher than low-end GDDR4 but beats it awesomely in perf/watt.
  • Mass production is expected to start in the first half of 2008 for both Samsung and Hynix, while Qimonda didn't mention it but the company is expected to ramp up production in early 2008.

Analysis

  • Based on the above timeframes, the first GPUs based on GDDR5 could possibly be released in late Q2 or Q3. Of course, that also depends on NVIDIA and AMD's roadmaps, and whether that makes sense in their current plans.
  • With GDDR5's levels of bandwidth, you really want to aim at the ultra-high-end even with a 256-bit memory bus.
  • And in that market, 512MiB of memory just won't be enough - so you'd need 16 chips with 512Mbit memory, which is more expensive. Thus, 1Gbit chips might have an advantage in the market.
  • Larrabee, which is slated for early production in 2009, is also widely expected to use GDDR5. The bus width is currently unknown, however.

With this kind of bandwidth, we can expect to see GPUs with more than twice the overall performance of the current highest-end products within less than a year. After a slow year during which NVIDIA's G80 remained the world's fastest GPU, things look like they're about to heat up big time.


Discuss on the forums

Tagging

graphics ± gddr5


Latest Thread Comments (4 total)
Posted by Tim Murray on Tuesday, 04-Dec-07 15:20:48 UTC
Mmmm doggy, that's some fast RAM! 128-bit bus for everybody!

Posted by Jawed on Tuesday, 04-Dec-07 15:43:45 UTC
Quote
And in that market, 512MiB of memory just won't be enough - so you'd need 16 chips with 512Mbit memory, which is more expensive. Thus, 1Gbit chips might have an advantage in the market.
Don't forget clamshell: http://forum.beyond3d.com/showpost.php?p=1085735&postcount=24 http://forum.beyond3d.com/showpost.php?p=1086223&postcount=38 which means Samsung's 512Mbit chips are not necessarily disadvantaged. Jawed

Posted by Arun on Tuesday, 04-Dec-07 16:00:45 UTC
Hmm, yes, there are advantages to clamshell - I'm not sure if it really reduces costs though? I mean, it's still 16 chips either way, so that ought to still be more expensive than 8 (larger) chips. Am I missing something here?

Posted by Jawed on Tuesday, 04-Dec-07 16:22:26 UTC
Quoting Arun
Hmm, yes, there are advantages to clamshell - I'm not sure if it really reduces costs though? I mean, it's still 16 chips either way, so that ought to still be more expensive than 8 (larger) chips. Am I missing something here?
I interpret this as a cost reduction arising from the reduced pin count per chip for data, 16, instead of 32. Which makes routing, per chip, easier (which I presume means less circuit board layers). The chips, themselves, will prolly cost the same for a given total board RAM size. So it's a question of board cost. The more I think about clamshell, the more it seems like this would suit low-end or midrange cards. Where one circuit board is configured as 128/256 bit or 64/128 bit (i.e. XT and Pro SKUs). Tricky subject... Jawed


Add your comment in the forums

Related graphics News

[Analysis] TSMC 40G to deliver up to 3.76x the perf/mm^2 of 65G & Power Implications
Old News: AMD CTO resigns, NVIDIA CFO retires, DDR3 for MCP7A, S3, etc.
SwiftShader 2.0: A DX9 Software Rasterizer that runs Crysis
S3 launches DirectX 10.1 Chrome 400 GPUs
GPGPU and 3D luminaries join 3D graphics heavyweights
The Technology of a 3D Engine - Part One
Stanford previews new F@H GUI and GPU client at SuperComputing 2007
Crytek postpone Crysis beta until October 26th
New Ageia PhysX processor on PCIe pictured
Free in-game ad supported PC games from Ubisoft available in the US