It feels as if it’s been a long, arduous wait for this new generation of video cards to come forth, and as 2013 came into fruition rumours were rife that we may not see anything new beyond the HD 7970 and GTX 680 until the end of 2013. But NVIDIA’s mega high-end (and mega-priced) Titan which launched last month put that one to rest, and reignited confidence that we may see movement far earlier than Christmas within more modest price bands. While there is still little word from AMD on the Radeon HD 8000 series, NVIDIA have done the expected.
Meet the GeForce GTX 780.
The GeForce GTX 780 is obviously the direct replacement for the GTX 680, but we are still looking at the Kepler architecture here. So whereas the GTX 680 was based on a fully functional Kepler GK104 GPU, the GTX 780 is based on a cut down GK110 GPU. It’s probably best to do a brief overview of NVIDIA’s GPU line-up at this point to help understand what we mean by this.
NVIDIA has two main GPU’s in its line-up – GK104, and GK110. GK104 was designed for NVIDIA’s consumer gaming range and is found in the GTX 680. In the High Performance Computing market (HPC) aimed at the science, medicine, engineering, and finance sectors, NVIDIA has GK110 and is found in NVIDIA’s Tesla products, K20X and K20, and most recently, Titan. Both are based on the Kepler architecture, so think of these two GPUs as different configurations of the same core design.
First GK110 in full fat form:
What you are seeing here is a block level diagram of GK110. At a high level, what you can see is:
- Five GP clusters.
- Each cluster has three SMXes (Streaming Multiprocessors).
- A total of 15 SMXes.
- Each SMX has 192 FP32 Streaming Processors (SP’s), 64 FP64 SP’s, and 16 Texture Units (TU’s).
Funnily enough, this fully functional GK110 configuration does not currently exist in any NVIDIA product, be it the consumer product range or its HPC Tesla brand. We’ll get to that just a little further down the page.
Now let’s look at GK104:
Here is GK104 with absolutely nothing disabled. This is the same GPU you’d find in the GTX 680.
- Four GP clusters.
- Each cluster has two SMXes (Streaming Multiprocessors).
- A total of 8 SMXes.
- Each SMX has 192 FP32 Streaming Processors (SP’s), 8 FP64 SP’s and 16 Texture Units (TU’s)..
For GK104, this is as big is it can go. If NVIDIA were to go any further with Kepler in the consumer space, NVIDIA had two choices – build something new, or use what they already had, GK110.
|Max Configuration Potential|
You can see how GK110 has ample headroom well beyond the GK104 based GTX 680. The first consumer (well, prosumer) GK110 based card was Titan. Titan is a slightly disabled GK110 with 14 SMXes instead of 15. The GTX 780 on review today is very much the cut down Titan, coming with 12 SMXes.
I won’t say the GTX 780 has one GP cluster disabled as is shown above, as technically NVIDIA have the ability to disable three individual SMXes from three different GPC’s if they wish, and due to binning requirements, you could end up with either.
|GTX Titan||GTX 780||GTX 680||GTX 580|
|Memory Clock||6GHz GDDR5||6GHz GDDR5||6GHz GDDR5||4GHz GDDR5|
|Memory Bus Width||384-bit||384-bit||256-bit||384-bit|
|Memory Bandwidth||288 GB/s||288GB/s||192 GB/s||192 GB/s|
|Die Size (mm^2)||551||551||294||520|
|Manufacturing Process||TSMC 28nm||TSMC 28nm||TSMC 28nm||TSMC 40nm|
|Launch Price||£830 /$999||£550/ $649||£430/ $499||£400 / $499|
Having 12 SMX units means compared to Titan, the shader count reduces from 2688 to 2304. Texture units from 224 to 192, yet surprisingly the ROP count is identical to Titan at 48 – great. The core clock speed of the GTX 780 has in fact increased over Titan. The GTX 780 also comes with a more sensible 3GB of VRAM, half that of the rather excessive Titan, but it still runs at 6GHz and still has a 364-bit memory bus, giving it the same memory throughput as Titan – 288GB/s
Interesting to note is that the GTX 780 still has the same TDP as Titan at 250 Watts. If you factor in that its GK110 implementation runs two less SMX units, much less FP64 shaders than Titan but has the same 250 watt TDP (which is rather strict thanks to GPU boost 2.0), it all points to it probably being a better overclocker than Titan.
GK110 is a big chip. With 7.1 billion transistors, it’s double the complexity of the GTX 680. Of course, with three SMXes disabled and much less FP64 units, some of this count is just there for the ride. As we seem to be stuck at 28nm these days, all this extra transistor count means we are back to NVIDIA’s usual strategy of producing massive chips in order to win the day. Die size is 551mm2. Compare this to 294mm2 for the GTX 680. GK110 is a far more expensive chip to produce than GK104.
And this brings us to pricing. Depending on how you wish to look at it, it’s either a really cut-price Titan, or a bloody expensive GTX 680 replacement. I’m leaning towards the latter camp. Titan was a nice idea for those who wanted compute performance (FP64) at a reasonable price. But NVIDIA did semi-market it towards gamers and it’s most certainly created a new tier in the consumer space. So now upon introducing the direct descendant to the GTX 680, they have priced it much lower than Titan, but much higher than the GTX 680 launch price and made us feel a whole lot more comfortable in doing so.The GTX 780 comes in at $650 (£550) whereas the GTX 680 launched at $499 (£430).NVIDIA are no stranger to this price range, but every time they attempt it, I find myself getting pretty peeved.
If Titan hadn’t have existed, I’m sure the GTX 780 pricing would be more difficult to justify by reviewers than it presently is. One thing you can always rely on at NVIDIA is very smart marketing, and this one was pretty smart. It’s also a very significant factor that AMD are nowhere to be seen right now at this performance level. Alas, NVIDIA currently has uncontested control of the sea, and this is the outcome. You cannot deny the GTX 780 is a slightly cut down Titan card for much less money. It is identical in it’s look, it has the same high quality cooler, and it offers 90% of it’s gaming performance. But it’s rather like introducing something to market that is desirable by many, but priced way out of reach, only to then offer something very similar at a price that is far lower but still difficult to justify, but hey, it’s now a bargain compared to the other one, right? Is it really?
We humans are suckers for a good deal, even when it’s been manufactured by a team of marketeers to appear that way. I would like to see the price come down, and it will in time, I’m sure. At $100 less, the GTX 780 automatically becomes untainted by pricing and on paper just looks rock solid in all departments. For now, the price is it’s only downside.The Card, Software & GPU Boost 2.0