nForce 4 SLI X16: Full Speed PCI Express
by Derek Wilson on August 8, 2005 1:56 PM EST- Posted in
- CPUs
Final Words
The first motherboards are only available from Dell, and other vendors will be bringing out products shortly. NVIDIA is shipping chips now, so it shouldn't be too long before we are able to buy an SLI X16 based board. Normally, we would lament the break in NVIDIA's recent trend of announcing a product with immediate broad availability, but if that's what it takes for NVIDIA to get a deal with Dell, then more power to them. Getting the SLI X16 into the new XPS 600 is quite a big deal for NVIDIA.NVIDIA is touting the added bandwidth as a performance increasing factor for the future. They acknowledge that current games don't see much benefit from the added bandwidth, but stand firmly behind the assertion that future games designed for PCI Express and SLI will take advantage of the increased bandwidth offered. We certainly aren't expecting to see any major gains under current titles, but we will have to wait until we get our hands on a board to play with before we can say for sure what we think of the performance.
We did take a couple of shots at comparing workstation class hardware to current nForce 4 SLI (nForce Professional enables 2 full x16 PCI Express connections with 2 processors), but due to the processor and memory configuration difference, we couldn't devise a fair test that could narrow down the impact of PCI Express bandwidth on real world games. Hopefully, we will have an SLI X16 board in our labs sooner rather than later.
There are no downsides to the convenience and market impact factors of this launch. On the high end, not having to flip a paddle is a welcome change. The added configuration options that motherboard makers will have for PCI Express slots can only help speed adoption. As motherboards based on the SLI X16 chipset will come in at current high end prices, the current SLI boards will fall in price and mainstream users will find little reason not to grab an SLI based system over the NF4 Ultra. More people with SLI systems means more impact from things like budget and mainstream GPUs supporting SLI. Here's the new price structure for NVIDIA based motherboards as NVIDIA sees it.
This is also a preemptive blow to ATI's Crossfire. With Crossfire motherboards not offering more than 2 x8 slots and many manufacturers going with the paddle design rather than ATI's recommended ICs for auto configuration, the new ATI enthusiast board will be stuck competing with a now mainstream product from NVIDIA. To be sure, the Crossfire motherboards are very good performers, but with this new option from NVIDIA and the price drop in current SLI products, ATI will have an even harder time getting their motherboards to end users.
We also see some possibilities for the GPGPU crowd with this addition of bandwidth to SLI configurations. This setup does more than just make it easier to move data around on each card. GPGPU stands for "General-Purpose Computation on Graphics Hardware" and with this setup, there is more than twice as much bandwidth dedicated to graphics as there is to system RAM in an AMD system with DDR 400. Depending on the latency, this opens up quite a number of possibilities for using the storage space available on graphics cards when they aren't in use. If some program (or even Windows itself) decided to treat unused graphics RAM as a separate memory node, we could see a total theoretical memory bandwidth of well over 20 GB/sec. Since system busses don't get guaranteed latencies, we would have to expect something like this to only benefit very high bandwidth stream processing. There was some talk at NVIDIA about GPGPU projects that they have under wraps, but they didn't give us any indication of what they are looking at doing right now.
In the end, even if performance impact isn't great (and even if NVIDIA's promised future performance benefits never materialize), this launch is a very good thing. Pushing prices on SLI systems down to more affordable levels and offering the potential for more PCI Express expansion slots (each with higer bandwidth) is definitely welcome. This is not the kind of solution that will entice current nForce4 SLI users to upgrade and the verdict is still out on whether or not there will be a real world performance difference, but the existance of the nForce4 SLI X16 solidifies NVIDIA's position in the core logic market and helps to push prices down across the board.
Facebook
Twitter
RSS
61 Comments
View All Comments
PrinceGaz - Tuesday, August 9, 2005 - link
You can easily test to see if there is any performance difference between x8 and x16 PCIe with a standard nF4 SLI board. Just drop one card (ideally a 7800GTX) in the first graphics-card slot, and run tests with the paddle set to single-card mode. That gives you the PCIe x16 results. Now set the paddle to SLI mode and re-run the tests with the same single card. It will now be running at PCIe x8 and you can see if there is any drop in performance. Voila! :)Fluppeteer - Tuesday, August 9, 2005 - link
The thing about graphics slot bandwidth is that it's *always* much less thannative on-card bandwidth. Any game which is optimized to run quickly will,
therefore, do absolutely as much as possible out of on-card RAM. You'd be
unlikely to see much difference in a game between a 7800GTX on an 8 or 16-lane
slot (or even a 4-lane slot). If you want to see much difference, put in a
6200TC card which spends all its time using the bus.
There *is* a difference if you're sending lots of data backwards and forwards.
This tends to be true of Viewperf (and you've got a workstation card which is
trying to do some optimization, which is why the nForce4 Pro workstation chipset
supports this configuration), or - as mentioned - in GPGPU work. It might also
help for cards without an SLi connector, where the image (or some of it) gets
transferred across the PCI-e bus.
This chipset sounds like they've just taken an nForce4 Pro (2200+2050 combo)
and pulled one of the CPUs out. It does make my Tyan K8WE (NF4Pro-based dual
16-lane slots, dual Opteron 248s) look a bit of an expensive path to have taken,
even though I've got a few bandwidth advantages. Guess I'll have to save up
for some 275s so I don't look so silly. :-)
PrinceGaz - Tuesday, August 9, 2005 - link
I wasn't suggesting measuring the difference between x8 and x16 with a TC card, it was for people who are worried that there is some performance hit with current SLI setups running at x8 which this new chipset will solve. I'm well aware that performance suffers terribly if the card runs out of onboard memory, and was not suggesting that. Besides anyone with a TC card won't be running in SLI mode anyway so the x8 vs x16 issue is irrelevant there.I agree there is unlikely to be much difference between x8 and x16 in games but it would be nice to test it just to be sure. Any difference there is could be maximised by running tests at low resolutions (such as 640x480) as that will simulate what the effect would be of the x8 bus limitation on a faster graphics-card at higher resolutions. It's all about how many frames it can send over the bus to the card.
Actually my new box has a 6800GT in it and an X2 4400+ running at 2.6GHz, so I'll do some tests this evening then flick all the little switches (it's a DFI board) and re-run them, then report back with the results. I doubt there'll be much difference.
Fluppeteer - Tuesday, August 9, 2005 - link
Sorry, should've been clearer - I didn't mean to suggest abandwidth comparison test either, just to say that where
you don't see a difference with the 7800 you might with the
6200TC. Not that I'd expect all that many owners of this
chipset to be buying 6200s.
I'd be interested in the results of your experiment, but
you might also be interested in:
http://graphics.tomshardware.com/graphic/20041122/...">http://graphics.tomshardware.com/graphic/20041122/...
(which is the source of my assertions) - although not as many
games are tested as I'd thought I remembered. Still, the full
lane count makes a (minor) difference to Viewperf, but not
to (at least) Unreal Tournament.
Of course, this assumes that my statement about how much
data goes over the bus is correct. The same may not apply
to other applications - responsiveness in Photoshop, or
video playback (especially without GPU acceleration) at
high resolutions. Anyone who's made the mistake of running
a 2048x1536 display off a PCI card and then waited for
Windows to try to fade to grey around the "shutdown" box
(it locks the screen - chug... chug...) will have seen the
problem. But you need to be going some for 8 lanes not to
be enough.
It's true that you're more likely to see an effect at
640x480 - simulating the fill rate of a couple of
generations of graphics cards to come, at decent resolution.
The TH results really show when pre-7800 cards become fill
limited.
My understanding was that, in non-SLi mode, the second slot
works but in single-lane config. Is that right? I'd like to
see *that* benchmarked...
Ah, wonderful toys, even if we don't really need them. :-)
PrinceGaz - Tuesday, August 9, 2005 - link
Yes, when an nF4 SLI mobo is set to single-card mode, the second slot does run at x1 so it is still very useful assuming companies start making PCIe TV-tuner cards, soundcards, etc in the next year or two. Apparently Creative's new X-Fi will be PCI only at first which is lame beyond belief. The 250MB/s bi-directional bandwidth that a x1 PCIe link would give a graphics-card would have quite an impact I'm sure.Fluppeteer - Wednesday, August 10, 2005 - link
Re. the X-Fi, I don't see the bandwidth requirements needing morethan PCI (not that I know anything about sound); I'm sure they
can make a version with a PCI-e bridge chip once people start
having motherboards without PCI slots (which, given how long ISA
stuck around, will probably be in a while). If even the Ageia
cards are starting out as PCI, I'd not complain too much yet.
Apparently the X-Fi *is* 3.3V compatible, which at least means I
can stick it in a PCI-X slot. (For all the above claims about
PCI sticking around, my K8WE has all of *one* 5V 32-bit PCI
slot, and that's between the two PCI-Es. I hope Ageia works with
3.3V too...)
nserra - Tuesday, August 9, 2005 - link
"Obviously, Intel is our key processor technology partner and we are extremely familiar with their products. But we continue to look at the technology from AMD and if there is a unique advantage that we believe will benefit the customer, sure, we will look at it."jamori - Monday, August 8, 2005 - link
I'm curious as to whether or not they fixed the problem with displaying on two monitors without rebooting into non-SLI mode. I'm planning to buy a new motherboard this week, and am going with the ultra version instead of SLI for this reason alone.I figure I'll spend less on a motherboard and more on a videocard that will actually do what I want it to.
Doormat - Monday, August 8, 2005 - link
Anandtech, please make sure to test out the CPU Utilization when using the onboard Ethernet w/ ActiveArmor on production boards. I'd like to see if they revised that at all, since the CPU Utilization was so high on the current revision of the boards. In fact, most nVidia nForce Pro motherboards for Opterons dont use the included nVidia Ethernet, they use the Broadcom or some other chip because performance is so bad.Anemone - Monday, August 8, 2005 - link
Just cuz if you own the mobo a few years there will be things to stick in x1 and x4 slots I'm sure.Nice going Nvidia :)