Wednesday, December 17th 2008
RV740 Taped-Out, First 40nm GPU From AMD
The manufacturing process technologies for graphics processors that dominated the 2008 were TSMC 65nm and 55nm. While AMD made the transition to 55nm over an year ago, it was a little later when we started seeing 55nm versions of existing NVIDIA GPUs.
A month into TSMC's announcement of of its 40nm bulk production node, AMD has reportedly taped out its first 40nm GPU, the RV740. While this is no high-end GPU, it is supposed to be the first successful port of AMD's GPU architecture to the new node. The RV740 succeeds the RV730, the GPU that went into the making the Radeon HD 4670. It is a mainstream GPU that ideally should make it to the sub-$100 graphics card segment. With RV740, AMD gains some experience as a manufacturer as it works on the RV870 "Lil' Dragon", the next generation GPU from the red camp. A product based on the RV740 can be expected only after Q1 2009.
Source:
NordicHardware
A month into TSMC's announcement of of its 40nm bulk production node, AMD has reportedly taped out its first 40nm GPU, the RV740. While this is no high-end GPU, it is supposed to be the first successful port of AMD's GPU architecture to the new node. The RV740 succeeds the RV730, the GPU that went into the making the Radeon HD 4670. It is a mainstream GPU that ideally should make it to the sub-$100 graphics card segment. With RV740, AMD gains some experience as a manufacturer as it works on the RV870 "Lil' Dragon", the next generation GPU from the red camp. A product based on the RV740 can be expected only after Q1 2009.
23 Comments on RV740 Taped-Out, First 40nm GPU From AMD
And I doubt AMD/ATi would roll out a faulty product. Mainstream will be happy with it's new and improved power efficiency and therfore thermal envelope.
Most engineering obstacles they'll get from die shrinkage should be ironed out through harvesting and repair techniques that both ATi and Nvidia rely on. It's not nearly as big of a gamble as the HD 4k concept was to begin with.
We'll see anyway:)
EDIT:
What I am worried about is the next generation of cards... this thing is pretty neat right now for us with all this competiton. I hope this never changes
tick-tock after all ;)
I also think a chip with ~rv730 (or even 2x) performance on 40nm would be ridiculously small, probably the size of RV710, certainly under 100mm2. It couldn't have a 128-bit bus, let alone a 256-bit bus needed to make an upgrade over rv730. It just doesn't make sense.
I think it becomes the perfect upgrade to rv770, and 40nm should allow it to be a similar size to rv730 (150mm2), at little over a billion transistors. A 128-bit bus with 4000mhz gddr5 could match 4850, and 7gbps gddr5 part could equal 4870 in bandwidth.
If we do the math, 1.5TF would only take 781mhz, 1.25TF 650mhz. A 25% upgrade over existing parts.
Now wouldn't that be a great "value" - oriented part? A upgrade to RV770, but to be placed below RV870. It also would make the value parts roughly follow the "2.5x" strategy I think ATi is going for with new processes vs rv730...while retaining the same die size, as well as floating technology down from performance to mainsteam.
On top of that, I wouldn't be surprised to see such parts, if it is what they are doing, residing in the 75W without-a-connector territory. RV730, at roughly the same size uses 59W @ 750mhz.
Such a part could probably get pretty close to taking on a GTX285 (at 400mm2+), and with a die size probably around 1/3.
It'd look something like this:
RV740pro:
150mm2@40nm
128-bit
650/4000
960 shaders
48 TMUs
16 ROPs
1.25TF
64gbps bandwidth
RV740XT:
150mm2@40nm
128-bit
775/7000
960 shaders
48 TMUs
16 ROPs
~1.5TF
112gbps bandwidth
Now, I'm not saying that's what they're doing...But wouldn't that make sense? It could compete now, and perhaps compete with the 40nm GT200 when it arrives on price, because it would be a smaller die and cheaper to make. I think GT212 will be 200-225mm2, 256-bit, gddr5.
But it's okay to call me crazy, I'm the guy that still thinks 870 is 2000 shaders @ 205mm2... 2.5 and 3TF parts (625 and 750mhz).
It may not be the direction they go, but if it is, I will find it amusing how well they continue to spin the press to confuse Nvidia (just as they did with RV770, and seem to be with RV870.)
/Dream
AFAIK ATi, Nvidia and pretty much every chipmaker has been releasing high-end first for a reason, and the reason they always give is the one I am "defending". It's not as if I knew that first hand, if it's wrong it's because THEY ALL were wrong, until know...
32 rops, 1280 sp's, 512-bit gddr5, hello!
Multi CPU on a die works for CPUs because you can't scrap its units if they fail, just the cache and because it's the better way of reaching parallelism. But GPUs are inherently parallel already.
Another issue is that if your elemental unit (each of those 4 GPUs) is a low-end part what are you going to do wth the ones that fail? Create an even lower end card? Foresee that will happen and use self-repairability in conjuntion to harvesting? A bad idea in any case IMHO, because you would inevitably end up with too many different options, this kind of chips (following your example):
1 - Chips with 4 full cores.
2 - Chips with 4 cores with 28 ROPS (448 bits) and 1280 SPs.
3 - 4 cores with 32 ROPS and say 800-960 SPs.
4 - 4 cores with 24-28 ROPS and 800-960 SPs.
[5 to 8] Repeat 1-4 but with 3 cores.
[9-12] Repeat 1-4 with 2 cores.
13 - Chips where each core is different.
*Note that the clockability of each core has not been mentioned, but that would just add one more layer of complexity, which is not good.
As you can see, there are too many different chips there, so you would have to scrap a lot of cores in order to create a defined lineup. You think Nvidia has released too many cards lately? That's a joke compared to what would they have to do for the above to be anything close to profitable.
sitting on a GTX260 anyway so very happy.