Nvidia 790i comes to show 780i owners the error of their ways

WHEN, OVER A year ago, the Nforce 680i made its debut, we were all ecstatic - Nvidia had filled a high-end chipset platform void which Intel hadn't really covered since the 975X chipset in 2004. At last, we had good overclocking combined great memory performance - the 8.7GB/s Sandra bandwidth. Results on a FSB1667 Asus Striker Extreme then were the closest Intel CPUs came to Athlon64 FX memory benchmarks.

There were also three PCI-E x16 slots, easily overclockable, for full-speed SLI plus some high-speed peripherals, plus hardware-accelerated dual Gigabit Ethernet with teaming and firewall acceleration. Not bad at all, except for the awful heat which cried for liquid cooling from day one - remember EVGA's Black Pearl?

Then, Nvidia took a long nap - both GPU and chipset-wise, while Intel (and AMD) worked feverishly on new platform generations. The P35 subsequently arrived, soon to be upstaged by the all-round 2007 chipset winner, the X38, itself to be "upgraded" to X48 now - at least the very top bins.

Even the Xeon bus licence didn't help, as Nvidia didn't bother to fill in another void, the high-end dual-socket 3D workstation market, allowing Intel Harpertowns on Seaburg chipsets become the dominant workstation platform. And yes, Seaburg - that same North Bridge as in Skulltrail - supports AMD Crossfire, dual PCIe v2 x16 at that, by default. So, AMD can earn some emergency dosh by selling pairs of Crossfired FireGL high-end cards for couple a thousand bucks per Intel workstation. Nvidia, well, can't.

So, Nvidia held it for quite a long time. Then - as can happen when one holds it for too long - a violent outburst followed. The Nforce 780i was the first thing out - basically, a kludge combining the 680i North Bridge and a Nforce 200 PCI-E v2 fanout creating two PCI-E v2 slots out of one - factory-overclocked, mind you, remember Nvidia LinkBoost - PCI-E v1 one from the northbridge?

Still as hot as before, with minor, if any, performance boosts - the only plus for the 780i was resolving 45nm Penryn CPU generation support. Which, somehow, they couldn't get done on the 680i, pity for all the owners. We don't count Triple SLI on GeForce 8800 Ultra as a plus, as, firstly, you will need a kilowatt-plus PSU for that, and secondly, if you really have that much money, you may as well replace the mobo and CPU too.

The added heat from the NF200, coupled with the already oven-like northbridge nearby, forced vendors to create some really scary heat sinks and pipes, using enough copper per board to wire up one of Saddam's palaces in Baghdad and still have some spare.

The 780i boards started trickling out before Christmas but, by then, everyone knew that soon after, Nvidia would continue coming - with 790i, that is. The chipset that you'll see a lot more of starting this month.

What is 790i? In one sentence: a good reason not to look at the 780i at all. Now, the multiple PCI-E v2 is native within the northbridge, avoiding all the bandwidth and latency roadblocks in the 780i. Then, you have native, dual-channel DDR3 memory support, the first for Nvidia, with DDR3-2000 OC enabled on the Nforce 790i Ultra. Nvidia seems to stick with its EPP enhanced performance profiles for that memory, let's see how it goes against - or together with - Intel's already-entrenched XMP profiles.

Now, if comparing DDR2 controllers in, say, Intel 975X and Nvidia 680i, all things being equal, Nvidia easily took a 10-12 per cent performance lead in both bandwidth and latency. On P35 and X38, that lead became nearly 15 per cent, as all newer Intel chipsets' memory controllers seem to focus more on DDR3 high-bandwidth, high-latency approach.

Then, as you saw from our Asus Maximus Extreme reviews - MSI X48 Platinum shows similar results with the new BIOS - a FSB1600 bundle with low-latency dual-channel DDR3-1600 gives some nasty high results now, above 9GB/s in Sandra. And, with X48, these numbers may scale well towards DDR3-2000 and FSB2000, depending on your board and memory.

Can Nvidia match or exceed it? It was master of memory controllers, after all. However, a year ago it had a two-year old Intel chipset against it - now, it faces a more aggressive Intel which also knows memory well, and which owns the darn CPU and FSB. And Nvidians tell us they expect it to be the "top overclocker".

Let's see about that. We'll be looking at the numbers that the 790i boards like Asus Striker II Extreme and similar MSI and Gigabyte offerings provide, in the next few weeks - just nice, a pretty much simultaneous roll-out of both X48 and 790i, might be a testers' wet dream.

Now, what does it all mean for those few new 780i owners? Well, yeah, you've been had, guys. Left out cold on the obsolescence stone, right at the point of purchase.

So why buy the 780i at all? Was it just to keep the DDR2 memory, with its present pricing? It's not justified to waste more electricity, cooling and so on with compromised performance on an extremely expensive platform anyway, just for the memory saving.

The extra bandwidth that dual-channel DDR3 provides may actually be beneficial for those dual and triple SLI setups, where multiple PCIe x16 slot data streams start competing with the CPU for that memory "attention".

Anyway, we hope the Nvidia's outburst of chippery, coming after all this holding back, could continue and last for a while longer, with a few more offerings including the dual-socket ones. They need it now - once the Nehalems come onto the stage (same applies for Shanghai and Fusion by AMD later) the space for a third party chipset supplier at the high end will narrow substantially, as most of the 'added value' north bidge functionality will move to the CPU.

Of course, Nvidia could produce some nifty I/O chipsets there, with hardware-accelerated Ethernet, wireless, RAID, USB v3, Firewire3200 and even HD audio all running off an I/O processor without disturbing the CPU. But hold on, something like that seems to already be in Intel's plans for Nehalem high end QuickPath chipsets.

Talking about Quickpath interconnect: if Nvidia really annoys Intel, it might end up without a Quickpath licence altogether, cutting it out of future Intel platform chipset business. Intel won't need Nvidia then: Chipzilla's got the CPUs, chipsets too - and some really nasty GPUs not far from now. µ