• Competitor rules

    Please remember that any mention of competitors, hinting at competitors or offering to provide details of competitors will result in an account suspension. The full rules can be found under the 'Terms and Rules' link in the bottom right corner of your screen. Just don't mention competitors in any way, shape or form and you'll be OK.

40nm cards might not be so brilliant - leaky 'sisters

Lightnix

Lightnix

Lightnix

Soldato
Joined
11 Sep 2007
Posts
5,744
Location
from the internet
At least according to this article on The Inquirer (watch out, it's by Charlie Dermerjian! :p), a possible reason for the numerous delays on the upcoming 40nm parts (which were originally scheduled for around December/January sort of time) is that the TSMC's 40nm process leaks power like mad - similar to the 80nm process that the 2900 XT suffered from, and Intel's 90nm process that didn't really help the already power hungry Prescott Netburst parts.

If this is to be believed, the 40nm parts won't actually bring any real advantages in terms of energy efficiency, the main problem with this being that the two graphics giants will have problems producing high-end graphics products for the next generation due to the high power consumption of existing chippery, what with DirectX 11 around the corner and all that.

http://www.theinquirer.net/inquirer/news/878/1050878/tsmc-40nm-revealed

Pinch of salt required obviously, but it would go some way as to explaining why there've been delays on the 40nm refreshes of both graphics companies' GPUs.
 
This could be where Larabee comes into its own, in the longer term. Intel will always be pushing the boundaries of what's possible before anyone else. They're close to releasing 32nm processors and are already testing 22nm.
 
So, what are these new cards - are they refreshes (e.g. 4970) or actual new ranges (e.g. 5870)? I'm a wee bit behind on this one.
 
reflux said:
So, what are these new cards - are they refreshes (e.g. 4970) or actual new ranges (e.g. 5870)? I'm a wee bit behind on this one.
Click to expand...

Here's how it's supposed to play out, in around March-May sort of time frame, we're going to see several refreshes of current generation parts from both graphics manufacturers. The cards are as such:

Radeon RV740 - 640 shaders, 128-bit bus, GDDR5 memory. Mostly a repacement for the 4850 and 4830, probably with higher clock speeds.

Radeon RV790 - 960 shaders, 256-bit bus, GDDR5 memory. The upgrade to RV770.

GT218 - 32 shaders, 64-bit memory bus, DDR2/GDDR3 memory. The replacement for the 9400 GS/8500 GT. Should perform more like a 9500 GT.

GT216 - ? shaders - probably around 160, 192-bit memory interface, GDDR3 memory. The 9600 GSO's successor.

Those're the ones I know of anyway. I think there's also a GT212 part that's supposed to have 384 shaders but I'm a bit sketchy on the details (512-bit bus + GDDR5 memory seemed a bit far-fetched, but hey, anything can be true in Nvidia land. :D)

Edit: Then in Q4-Q1-2010:

RV870, the new DirectX 11 radeon part. Not a lot is known about it, what specs were 'revealed' about it before are apparently now applicable to the RV790 so not a lot is known.

GT300, the new DirectX 11 GeForce part. As far as rumours go, it's supposed to be implementing MIMD technology as opposed to SIMD technology for processing data.

Both DirectX 11 parts will definitely have a hardware tessellation unit as defined in the DirectX 11 specifications and will be compliant with Shader Model 5.0.

On the official naming front, the upcoming releases will probably still be under the G200/4000 naming scheme we see today. The DirectX 11 parts will probably be GTX3-whatever and 5870. That's just a guess though so don't quote me on it. :P
 
Last edited:
Vertigo1 said:
This could be where Larabee comes into its own, in the longer term. Intel will always be pushing the boundaries of what's possible before anyone else. They're close to releasing 32nm processors and are already testing 22nm.
Click to expand...

not really, what you can make with one chip doesn't go ahead and mean you can use the same process on another chip. Their 32nm isn't too far off, but its also tested on their cpu's which use quite a radically different design to a gpu. The same way they tested 32nm sram chips they could build probably minimum of a year or more likely two years ago is only just about ready for use in a full scale cpu.

Also their plants are very much not geared to making gpu's. Yes they have fab's setup to make chipsets but they won't be geared out for making gpu's and would cost a LOT of money to refit, at a time where they are cutting fab's down and moving a couple of others the likelyhood of them pushing billions into a purely larabee fab with their highest quality newest more expensive tech are, well, slim to say the least.

TCSM has been testing this 40nm for, well, ages. Remember they are currently manufacturing various orders of chips from various companies probably ranging from 55nm through to 90nm stuff still.

Remember also, despite being Intel, there is no support in any way or form for Larabee yet, neither would their be on release(unless the rumours of PS4 having a larabee prove to be true which is a great move by intel, if true of course). In other words, Larabee will most likely start off with small sales, VERY small, maybe really only minor sales in GPGPU based rendering rigs or whatever big companies use those comps for till more graphics side support in games comes in. But with a small selling gpu, Intel would waste a lot of money putting the best and newest process into it, in a new fab with massive capacity. So expect small runs, on a cheap high yield process that works best for them and cheapest. Otherwise you put hundreds of million into the kit to produce them, and only sell 10,000's of units. hell, they might even farm them out to somewhere like TCSM until they need to produce more where using their own fabs becomse better value to them.
 
I think we'll see Intel producing their own Larrabee chips, if only for a matter of pride over financial sense. They hold the title for largest semiconductor manufacturer on earth, after all. I place no doubt in that Intel barely even considered outsourcing their manufacturing for Larrabee. I'd imagine they're much more kitted out than you might think for producing a GPU, particularly as Larrabee has been in the pipeline for quite a while now.
 
Vertigo1 said:
This could be where Larabee comes into its own, in the longer term. Intel will always be pushing the boundaries of what's possible before anyone else. They're close to releasing 32nm processors and are already testing 22nm.
Click to expand...
Will the 32nm chips require a motherboard upgrade from a C2D board?
 
MikeHunt79 said:
Will the 32nm chips require a motherboard upgrade from a C2D board?
Click to expand...

I would say from a core 2 board yess......but they might be the same as the i7 socket
or the new mainstream socket that is comming.

i would definatly be suprised to see the currant core 2 and core quad range on 32nm.
 
drunkenmaster said:
not really, what you can make with one chip doesn't go ahead and mean you can use the same process on another chip. Their 32nm isn't too far off, but its also tested on their cpu's which use quite a radically different design to a gpu. The same way they tested 32nm sram chips they could build probably minimum of a year or more likely two years ago is only just about ready for use in a full scale cpu.
Click to expand...

Larrabee is very very different to the GPU's ATI/nVidia are making though, it is essentially a very large cluster of tiny CPU's, so the fab process may transfer very well. The individual processors in Larabee are based on a modified implementation of the old P3 architecture iirc.

nVidia/ATI design GPU's within the limits of what TSMC can produce. Intel know very well what they can produce so there is no reason to assume they would design a GPU that doesn't work well with their own fab technologies.
 
p0rks0da said:
Larrabee is very very different to the GPU's ATI/nVidia are making though, it is essentially a very large cluster of tiny CPU's, so the fab process may transfer very well. The individual processors in Larabee are based on a modified implementation of the old P3 architecture iirc.

nVidia/ATI design GPU's within the limits of what TSMC can produce. Intel know very well what they can produce so there is no reason to assume they would design a GPU that doesn't work well with their own fab technologies.
Click to expand...

yes there is a reason, even in a best case scenario larabee won't sell more than 1/100th of what ATi/nvidia will sell at the same time because, larabee's an unknown, dell won't stick a Larabee gpu in any of their computers, it won't be competitive low end which is where the massive sales come, and it probably won't be supported high end very well to start with.

In that case, theres no reason known to man that spending money on the top kit to make a very small run of chips would be worth while in terms of money. They will however have lots of fab equipment "left over" as they move cpu's to newer smaller processes which will mean "free" equipment to transfer over to making gpu's in small quantities.

Remember, this is a time that latest info is I believe 3 fabs are to close and a 4th fab is being moved from a very expensive area of, shaghai is it, over to a cheaper area with cheaper labour/costs.

The actual design of larabee, the individual chips might be more similar to cpu's, but all together, with all the core logic involved and the interconnectivity is where current leakage and massive signalling issues become a significant difference to a "basic" cpu. Cpu's don't tend to move 100gb/s of data througout the core, gpu's have far more data moving in and out and that all leads to being much more complex. You can't look at the individual part and say its similar, because break any gpu/cpu/memory chip down far enough and its all single transistors, it doesn't really mean much. You look at the whole of the chip, how its used, and then what can give you a profitable production run.

Using a fully kitted out 800million cost plant to produce 100k gpu's that only make £30 profit a sale is a complete waste. Using the same kit to making 5million cpu's at £50 profit a pop is worthwhile. if you're only going to be starting in the gpu market you'll use the old tech that you can pull together for 50mil, stick it in a corner of a plant and make 100k cpu's to get your nose into the market making as few losses as possible.
 
You must log in or register to reply here.
Back
Top Bottom