GDDR5 in new 250 M and 260 M

According to the blog outlets, the new mobile GPUs are getting the latest architecture at last and some GDDR5 memory! :) Let’s hope there is something radically new on desktop side soon as well!,2817,2348694,00.asp

Wow, interesting news for two reasons. First, these are 40nm parts! So there’s progress with getting that new process working.
Second, the memory size for 4 out of 5 these mobile parts is 1GB! Wow… that’s really nice for mobile!

There’s also a vague reference to “half the power use” which sounds promising but is way too vague to make any conclusions.

However, the article does say that the new parts use GDDR3, not GDDR5.

Yeap, you are right. As usual I just looked at the chart and didn’t read the whole thing, so passed on their typo in the chart:)

This is indeed very good news, and that is one thing towards a GT300 with 40nm technology.

Hey, Demq, you’re not only right, but the NVidia official slides say so too!

Probably still a typo. But maybe not. GDDR5 gives you double bandwidth so you’d expect to see narrower memory width and slower clocks, but the values in the slide are consistent with GDDR3.

Hmm, and an article about them says “We should also note that these new mobile GPUs are DirectX 10.1 compliant–a first for NVIDIA.”

But that implies design changes for sure, since that’s a hardware issue. Unless maybe it’s a design fix and not really a new addition?

BTW, NVidia spec page clearly says GDDR3.
That same page only says DirectX 10, though.

And the next question is: Are these compute capability 1.1 or 1.2? We haven’t seen a 1.2 device yet, but a new mobile chip is the logical place for one to appear.

Yes, but that page is clearly for the GTX260M which is clearly not the same product as the GTS260M :argh:

Anyways, Washington Post mentions that the 96 core GTS250M has a 28-watt TDP:…9061501460.html

Whoah! You’re right. It’s confusing.

But look at the GTS260Mversus the GTX260M. Suddenly it looks like there may indeed be GDDR5!

GTX 260M :

  • memory clock “Up to 950 Mhz”

  • Memory Config up to 1GB GDDR3

  • Memory width 256 bit

  • Memory Bandwidth 61 GB/sec

  • DX 10

GTS 260M:

  • Memory clock “up to 2000 MHz”

  • Memory Interface Width 128 bit

  • DX 10, Compute

The spec page for the GTS does not specifically state the type of RAM nor the bandwidth, but look at what has changed. The bus width is halved, and the clock rate is doubled.

If they are counting the “effective” clock rate of the RAM and not the real clock rate, then that is very consistent with the GDDR5 doubling of bandwidth per pin.

So I would actually have to guess that yes, this is GDDR5! And the fact that the low end parts use GDDR3 means that the memory controller can handle both types. And the DX spec change shows that this is indeed not a trivial shrink, but some re-engineering and redesign.

If these are using GDDR5, that also means that desktop parts would also be very useful… GDDR5 has fewer board traces and makes things much simpler (and cheaper).

seibert’s question of whether this is compute 1.1 or 1.2 is now also a very good one.

Compute 1.1 does not form part of the question, I think:

  • We also got questions about which architecture these GPUs came from and if we were simply renaming old GPUs. The answer to that is definitive–these are new GPUs based on a new design and 40nm process. At a high-level these chips are based on the same architecture used in the high-end NVIDIA desktop GPUs, so they share the same compute benefits.

Rene Haas, General Manager, notebook business at NVIDIA…tions-answered/

Hmm, wow, look at the power use of the GTS 250M. It’s only 28 watts. That means you’re getting 12.9 GFlops/watt.

That’s better than anything we’ve ever seen. The last generation of mobile parts (from only about 4 months ago) is about 7 GFlops/watt.
The best ratio for desktop parts is the GTX295, with 6.2 GFlops/watt.

So, half the power use. GDDR5 means half the memory traces needed. Half the pincount on the chip. Smaller chip. Put this together and you get the possibility of making dual GPU single-width cards.
Or quad GPU double width cards??? Hmmmmm, tasty!!

And this isn’t even talking G300, this is just with chips that NV just showed it has.

That statement is too vague to be definitive. Mobile and desktop GPUs could be thought of as sharing the “CUDA architecture”, despite having different compute capabilities. (A technical detail I would not expect a manager to discuss in a generic Q&A session.)

Compute 1.2.

WOOHOO! Double the register count and the versatile coalescing are such a big help in designing kernels.

This is awesome news. Kudos to the hardware guys for getting it working… in 40nm and with GDDR5!

Great, and then the next qustion becomes “when?”, which I believe might be answered by some projections of the near future for TSMC’s 40nm process:

TSMC Promises to Improve 40nm Yields in Two Months

Nvidia plans to ramp up production its recently announced graphics processing units (GPUs) using 40nm process tech at TSMC in August. Meanwhile, ATI is trying hard to supply any amount … According to the report, Nvidia has said that its 40nm GPU production at TSMC will be limited to OEM parts initially. 40nm GPUs destined for the channel market will be introduced when the process is more mature, likely around the end of the year.…Two_Months.html

So no back-to-school-brand-spanking-new-MacBook - but when did that ever happen? - and those hoping to upgrade an existing system will have to wait untill Christmas … Which suits me just fine - the weather here is absolutely gorgeous these days :-)