
Posted 2010-03-05 14:34 by Nate
As expected, CeBit was a total gold-mine for Fermi news. This is the chip that will live inside the new GTX 470 and GTX 480 chips from NVIDIA. It is a beast of a chip too, more than doubling the number of processing cores from 240 in the current GT200 to 512 in Fermi. Here's what we're talking about, in terms of block diagrams:

GT200 Fermi
It is rather tough to tell from these little images, but Fermi is a massive departure from the old way of doing things at NVIDIA. This has been mostly setup for DirectCompute and CUDA operations, with a massive L2 cache that the whole chip can access, which is one reason why NVIDIA began this product's launch with an architecture discussion back in late September of last year, that Anandtech covered expertly. They wanted everyone to know that their cards were going to be the kings for scientific applications, and they appear to have delivered that promise. They did also promise to have massive gains in gaming too, and that's what we're here to talk about today.
As we should all know by now, Windows 7 is officially king after shipping more than 90 million licenses, and that means that DirectX 11 is the gaming platform of choice for PC developers. We've already seen some titles hit the ground running, like Aliens vs. Predator and Battleforge, but we haven't seen the huge opportunities that are out there yet. I'm preparing an article that might just blow your mind if you don't know what is really possible with just one piece of the DX11 pie, but the point is that anyone who is anyone will be gaming with DirectX 11-capable cards in the near future. NVIDIA is running quite a bit more than fashionably late to this party, but Fermi changes all of that.
It would appear that NVIDIA had a little trouble with TSMC's 40nm process with this chip, as the working silicon we're seeing at CeBit is "A3" silicon, with the "A" being the first main design, and the "3" being a third spin. Spins come from small changes to the design, usually to increase yield or frequency headroom. If there was anything wrong with the way the chip computed things that could not be fixed in software, we would see a "B" revision. The A1 chip was probably just coming back in September when NVIDIA blew open the bag on Fermi, the A2 spin would've been ready around Christmas, and now this third one is finally ready to start shipping on March 26th. However, from the reports at CeBit, all is not well with A3 silicon.
Here's what we can show about GTX 480 and GTX 470 versus the Radeon 5870 thus far. As Fudzilla reported just yesterday, this might all change before launch, but given the way GT200 turned out, I would be willing to bet on these specifications, compared to current Radeons:
The first really scary thing is that power rating. We saw boxes at CeBit that showed a 600 watt power supply requirement, with a total of 42 amps on the 12 volt line, for the GTX 480. That's 500 watts on the 12 volt, so don't expect to be running one of these off just any 600 watt PSU. Since our lowest-end gaming machines have shipped with Corsair's 650 watt PSU, they'll do just fine with one of these big guys. But still, what do you cool a 300 watt card with? Here's what they showed at CeBit:

It looks like NVIDIA has resorted to a pretty serious heatsink here, however Fudzilla claims this isn't the one that will actually ship out. What's notable is the 4 8mm heatpipes on board, which would be be a massive departure from the old way of doing things. However, desperate times call for desperate measures, right?
There's no doubt in my mind that the increased complexity of this heatsink will also be passed along to the public, which means we're probably looking at launch pricing pretty similar to the GTX 280 and GTX 260 prices. They came out at $650 and $400 respectively, and I wouldn't be surprised to see that again here. In fact, if NVIDIA charges much less for these boards, they will be losing way too much money on each one to make it worth their while. They have an extra 6 months in development to pay for in addition to the extra RAM, big heatsink, and huge die size. So, what does that mean for the shopper?
Well, NVIDIA finally showed us some numbers, namely the Unigene Heaven benchmark:

image used from Hexus.net's story
If this is all that NVIDIA can muster against a stock 5870, they are in a heap of trouble. At 70 seconds or so, the two cards make a serious departure, which coincides with the heavily-tessellated dragon statue scene in the benchmark. While that's great for NVIDIA down the road, having a very fast geometry processor isn't going to win benchmarks on today's games. In fact I'd say that we can expect to see the GeForce GTX 480 pull a 10% win in today's DirectX 10 games. If you guess at a 10% win on this graph in the non-tessellated areas, you can guess at how much faster Fermi is than their previous generation. Looking at Crysis:Warhead, for instance, I see the GTX 480 coming in just between the GTX 295 and two GTX 275's in SLI at 1600P with 4x AA.
For having twice the number of cores, more than twice the memory bus and a more advanced core design, they are only getting around 40% more performance than their old GT200. I'm guessing that means clock speeds are down, considerably, and other reports online show that also. This means to fit in the 300 watt envelope, NVIDIA seriously crippled their doubling efforts, and are going to ship a card that hopes to compete on features and not necessarily performance. What's interesting is that around Christmas, NVIDIA started to dry out supplies of GTX 275's, 285's and 295's. They were probably expecting to ship the A2 silicon, and decided to force customers into Fermi. When that didn't happen, they probably did an emergency batch of GT200's to hold up until A3 came around, so if you still want the best DX10 combo around, get your 285's while you can.
I'm still on the fence about whether or not DirectX 11 is worth the jump yet. The GTX 480 is shaping up to be a very hot, very power hungry, and very inefficient card. It's also going to be very expensive. NVIDIA will, without a doubt, shape this chip up with a B1 or A4 revision that will allow for much higher clocks and lower power consumption, and it will probably come as a GTX 485 at the end of the summer. So, with the unforgivable issues plaguing the Radeon 5800-series, the question really comes down to whether or not any titles will be out that really need DX11 between now and then. So far, AvP and Battlefield: Bad Company 2 are the two big titles that use DX11, and they don't seem to have really hit the nail on the head. Will Starcraft 2? Probably not. LOTRO is my current favorite for DX11, since their custom engine can be tweaked much easier than say the Unreal Engine, but even then it's a single title in the big sea.
I hope I'm wrong about all of this, and that NVIDIA has an ace up their sleeve, but this time I think only one word can describe what's coming out: FAIL. Go grab a GTX 260 Core 216 or two and be happy until that DX11 title hits that really gets your blood pumping.


UPDATE: This EVGA box shot confirms the memory size and, thusly, the memory bus for the GTX 470. The big question is how many cores and what frequencies?
ANOTHER UPDATE:
If this screenshot is to be believed, things are not as good for GTX 470 as I hoped. It only shows 448 (14 x 32), which means two packs of cores have been disabled on this chip. That would go inline with the two memory interfaces being disabled. I have a feeling we'll be seeing a GTX 470 Core 480 towards the end of the year, much like the GTX 260 Core 216 that came once GT200 yields improved.