Welcome! Today I will take an in-depth look into GIGABYTE’s top of the line 990FX Chipset offering, the 990FXA-UD7. The new UD7 has the AMD 990FXA chipset, which supports AMD’s future Zambezi processors, better known as Bulldozer. There has been a lot of hype over the new AMD platform, as Bulldozer is supposed to bring AMD back to the top, and AMD is bringing back its almighty FX lineup. Only time will tell how it will pan out for AMD, but motherboard manufacturers aren’t waiting around. The 990FX chipset paired with the SB950 offers some great features; biggest of them all is SLI support and Hyper Transport 3.1. The 990FX chipset itself has 42 PCI-E 2.0 lanes, 32 of which are for the GPUs. To top it off the 990FXA-UD7 can support 4-way SLI, something that no other 990FX chipset board can do at the moment. Now while on the Intel side 4-way SLI needs NF200 Chipsets, on this AMD platform we can do it without any added latency. Of course you will be running 4-way at 8x per slot. Now this platform is built for the future processors, so there are a few features that aren’t supported by AMD’s current top of the line Phenom 2 X6 1100T which we will use in the review today. One such example is the HyperTransport 3.1, which runs at 6400GT/s opposed to HyperTransport 3.0 at 5200GT/s. Another example, and one worth noting is the vdroop that many have encountered with this motherboard. GIGABYTE states that it’s design (vdroop in particular) strictly adheres to AMD’s AM3+ guidelines for the Zambezi processors. Today we will venture far into the deep and darkest corners of the 990FXA-UD7, and uncover whatthis board is really made of.
Here is a video review if you are short on time
This review features a new section that I have been working hard on developing, the VRM testing section. We will use a DSO(digital sampling oscilloscope) hooked up to a separate system, to analyze peak2peak voltage (ripple), as well as run a pass fail test to see how often the ripple measurements surpass a set point. We will run these tests not only at stock, but also in the worst case scenario, something this board out of the box cannot do.
The review will be cut into segments:
Introduction (you are here)
Box, Accessories, Layout
Deeper look at the VRM and ICs
VRM, Voltage, and Oscilloscope testing
Performance Analysis + Overclocking
Conclusion
Accessories are exactly what the doctor ordered, a prized 4-way SLI bridge(not easy to find), 3-way SLI bridge, one 2-way SLI bridge, two CrossFireX bridges, four SATA6GB/s cables(two angled tips), a labeled back panel, and the manuals(GIGABYTE sticker too). Also they are all black except for the manuals and stickers, a shift towards all black to match the board.
Sin’s Take: Now with P67 early on we say that GIGABYTE didn’t really customize the parts, in terms of color their bridges were the stock brown color, the cables were blue matching their X58 boards, and their 3-way bridge was blue as well. Since then we have seen a shift towards black SATA cables, then black 2-way SLI Bridge, and now black 3-way and 4-way bridges. We see that GIGABYTE is putting in that extra touch, which one comes to expect from a Tier1 manufacturer.
There is our full-sized shot. You can see the heatsinks look even better than advertised, something fast-food never looks like. Now don’t worry I am going to cover everything, just hold on tight, because it’s going to be a fun filled ride. We will start from the upper left hand corner (backpanel) and move towards the lower left. Oh by the same this board is standard ATX size, a bit wider than others, but it does support 4-way SLI in ATX standard case.
The Southbridge on this recent AMD platform has no USB 3.0 connections natively, unlike the A75 chipset, so third-party controllers need to be used. GIGABYTE uses two on this board, one of them supplies us with 2 ports in the back, and that controller actually operates off one of the 42 PCI-E 2.0 lanes from the 990FX chipset.
Sin’s Take: From what we saw with the A75 chipset, it seems that the EtronTech controller is actually faster than native AMD USB3.0, so for now this option (off-die controllers) is one that works best. I would have thought that by now AMD had USB 3.0 in all their chipsets, but then again the same goes for Intel and they don’t either.
The VRM heatinks is connected to the Northbridge heatsink through a sintered heat pipe, gold accents match very nicely, the VRM is lined up straight in line with the Northbridge; this allows GIGABYTE to open enough space for 4-way SLI. We will cover the VRM in depth very soon. Here you see the DIMMs, 4 for dual channel. They are close to the socket to reduce parasitics caused by trace length, to preserve signal integrity.
Sin’s Take: This design with the VRM in one line and the Northbridge next to it seems to be part of AMD design, as most manufacturers have done this, but surprisingly 4-Way SLI support is nowhere to be found other than this particular board.
In this area we have 8(Eight) SATA6GB/s ports, 6 from the Southbridge, and 2 from Marvell SE9172. We also have that little SATA power connector, which provides extra power for the PCI-E slots. For internal headers along the entire bottom of the board, we have; front case connectors, TPM connector, USB 3.0 header, 3 x USB 2.0 headers, IEEE 1394A header, and audio header. There is also a Clear CMOS jumper right on top of the case connectors. A small POST display is also on this board to help with debugging. Here we can also see the buttons; Power, Reset, and Clear CMOS. The Clear CMOS has a cover.
Sin’s Take: I also like the buttons, they are well placed, and unlike their P67 counterpart they aren’t facing upside down (the lettering that is). The POST LED is most helpful to me, for instance once I forgot to put RAM in the board and the error came up as C1, and then I looked and the RAM wasn’t there. It’s a useful display, probably the most useful auxiliary feature on most of these boards. What are also important are Power, Reset, and ClearCMOS buttons. The ClearCMOS button being covered is less of a pain, because I found that I had hit that cover a few times when not looking. Luckily the clear cover is easily removable.
Here you see the PCI-E slots, we have 6 of them, 2 are full 16x slots, 2 are 8x slots, and 2 are 4x slots. If you scroll down to the in-depth section you will see which is which. The 16x slots are triple slot spaced so that you can run two triple slotted GPUs without issue.
PCI-E layout is like this:
16X Slot==============================================
4X slot==============
8X slot=========================
4X slot==============
16X slot==============================================
PCI slot======================================
8X slot==========================
Sin’s Take: This is the ONLY 4-WAY SLI/CF capable AM3+ board. What still puzzles me is how GIGABYTE pulled this off while ASUS didn’t, it seems that GIGABYTE really did their research when it comes to this board, it’s built majestically. I went to check out ASUS’s offerings as I usually do when I do my reviews, and on Newegg it states that 4-way SLI is capable with the ROG expander. Now I laughed, but not because of the fact that you can’t fit it in a case, but because you can’t buy it in a store in any of the territory Newegg caters too. The thing about the 4-way SLI is that it fits in ATX standard sizing, so there is no need to buy a new case.
Deeper look at the VRM and ICs
First I will show you the heatinks, as I took them off of the board.
They were held down by 6 screws and washers.
In this shot as you can see the shape of the heatsink.
Next we are going to mess with the VRM, I have a few pictures that show what is going on with the VRM, whether you have no knowledge of electronics or some, you can relate to either picture.
This one is a bit more centered on the flow of things rather than their actual functions, as it only covers part of the function.
This is more function:
This is a bare shot just in case you want to see it without the text. The VRM uses SiC769CD Generation 3 Driver MOSFETs. They can switch at upto 1 MHz or as low as 100 kHz. That frequency is referred to as the switching frequency the PWM outputs. In our case it’s ~280 kHz measured at the DrMOS. Now the Intersil ISL6630 is a 4+1 phase PWM, it uses Intersil Phase Doubling technology, which takes a phase from the PWM and doubles it. Intersil ISL6617 Phase Doublers are used for this purpose. (Doublers: http://www.intersil.com/data/fn/fn7564.pdf) Doublers have to be used, because we need 8+2 channels to drive 10 separate Drivers inside the DrMOS. Now these phase doublers have different operating modes, and I am unsure which is used. The final output switching frequency is ~280KHz, so it might be dividing the switching frequency or maintaining it. This low switching frequency allows the DrMOS output more current while running at higher efficiency. Check page 14 for more information on this: http://www.vishay.com/docs/64981/sic769cd.pdf . Each DrMOS can output 35A, including those for the CPU-NB output. The inductors are 80uH in our case and the output capacitors are rated at 560uF each. That is 6720uF of bulk output capacitance. The LC(inductor/capacitor) circuit acts as a low pass filter on top of being energy storage. The inductors effectively reduce the switching noise of the DrMOS and stop frequencies above something like 60 Hz to pass to the CPU. When the CPU needs more current, the capacitors discharge, and while this happens the inductors charge with more current, and then empty into the capacitors to start the cycle over again, this way enough current is delivered to the CPU when needed.
Sin’s Take: I have seen a lot of VRM designs, and I want to talk about true phases and this and that. Most of the time, I see the use of a pair of Low RDS(on) MOSFETs instead of DrMOS, and this is for a reason, most companies can use half the number of drivers to drive these FETs. Instead of using a driver per two FETs they use a driver per four FETs, so in fact each of the two “phases” is working off a single driver. This does cut down costs, but it also doesn’t allow for the use of DrMOS, because of the DrMOS IC needs its own PWM signal. Instead of switching on two FETs 1/8 of the cycle (with DrMOS); they will switch on 4 FETs ¼ of the cycle. I like to think of true phases as; a PWM signal, a driver, a pair of (high-side & low-side) MOSFETs, and an inductor. Most would just count phases by the inductors present, but this isn’t the most correct way to do it. Personally I know DrMOS are the future, as they are much smaller and output more current. They are an all-in-one solution, and it looks like VRMs are becoming smaller and smaller with each generation. I read on forums that many are blaming the DrMOS for their vdroop, and this is untrue. How do I know? Well I did a vdroop modification on this board, and the voltage actually increased under heavy load, which disproves that the DrMOS are the reason for the vdroop. The reason for the vdroop directly from GIGABYTE:
“GIGABYTE 990 series motherboards strictly follow the AMD AM3+ load line calibration design guide, and so CPU V-core voltage will drop according to loading. Such calibrations are built into the platform to protect the user’s purchase and prevent damage to the PC system.“
I will show you guys the droop mod in a second.
Above are the RAM and Southbridge VRMs, the RAM uses two phase design, and the Southbridge uses a single phase. Both use the ISL6546 as their PWM, with Low RDS(ON) MOSFETs. The RAM VRM also seems to have a fairly large amount of capacitors. The CPU PLL is powered by an LDO, as is the Northbridge.
Sin’s Take: I feel as if the RAM VRM has been upgraded from previous generations of AMD boards. We also have very precise voltage control, 5mv divisions through the BIOS for EVERY voltage. GIGABYTE says this is because of their precision OV ICs.
Following this picture are 3 pictures that explain everything else on the board. Please enjoy.
Sin’s Take: This board is loaded and that is that, it uses high quality components, from power delivery to its features it doesn’t hold back. I expect no less from a flag-ship motherboard.
VRM, Voltage, and Oscilloscope Testing
To begin this section I need to explain everything that is going to happen, but testing will proceed also.
Voltage Read Points Identified
990FXA-UD7 VDroop Modification Identified and Explained
Digital Multi Meter Used to Capture Voltage Outputs
Oscilloscope Measurements and VRM Testing
Voltage Read Points aren’t the most abundant on this board. The back of the CPU socket with the MLC Capacitors have the major read points, but then under the Northbridge itself we have two more, and I also found a VDimm read point on top, as well as the CPU PLL voltage.
So let’s begin, some read points have more than 1 point.
As you can see the voltage read points are most easily attained at the backside of the board. Unlike the P67 generation where you can measure the vcore off the choke’s leg, these inductors are through-hole and not SMD, so you can try again at the back of the board. Of course the back of the socket at the MLCC for read points of ground and vcc is the best measurement spot as it is the voltage right before the CPU’s pins, you can also try the capacitor legs.
The 990FXA-UD7 VDroop mod.
Disclaimer- Please do not do this mod unless you read the explanation and the output numbers below, and do not do this mod if you do not know how to solder, or if you want to keep your warranty. GIGABYTE does NOT endorse the modification, and thus you will probably be denied RMA if something goes wrong.
I made a video for you showing the vdroop mod and its affects:
The way this mod works it that it changes the resistance a certain pin on the PWM and ground. This pin controls vdroop. On the Intel side, for PWMs there is no vdroop pin, there are only an FB and COMP pins which are voltage feedback regulation pins, and that is how LLC is usually implemented on Intel side. On AMD PWMs from Intersil we have this “drpcntrl” pin #4. Intersil describes this pin as:
“Droop Control for Core and Northbridge. This pin is used to set up one of four user programmable selections via a resistor tied to ground: Core Droop On and Northbridge Droop On; Core Droop Off and Northbridge Droop On, Core Droop On and Northbridge Droop Off; Core Droop Off and Northbridge Droop Off.”
Later Intersil states that :
So at stock it’s a 0K resistor, we change it to 100K and we have droop disabled for both.
Here is an idea of what happens when vdroop is disabled, as well as other voltages:
You can see how voltage increases under load without droop! It’s actually extremely high. Now I have to say a warning, you do notice that after you do this mod the voltage will be 75mv higher than what you set at idle and 30mv higher than that under load. Well at stock turbo core is enabled and your stock vcore is 1.475v, 1.375 if you disable the turbo core. So watch out when you reset the CMOS, as you will need to lower the vcore.
Sin’s Take: The voltage increments through the BIOS are excellent. 5mv for every voltage we have control over. The voltage regulation in terms of vdroop and no droop is about the same, but reversed. I find it odd that voltage increases so much when droop is disabled, so I recommend not doing so, and just compensating for the droop itself. Let me show you what happens under load with current.
Explanation of the Oscilloscope:
For Channel 1 the probe is soldered to be back of the socket to the MLC Capacitors, I use two points soldered together to make sure I have good contact, using a shielded coaxial cable to reduce noise, and this cable is routed straight up and out of the underneath the board, as to not pick up noise from the DIMMs or VRM. Still it is not perfect, it can never be, as the noise created will be picked up, but we can always try and make that noise smaller. I am using DC coupling as I want to measure the DC voltage change over time.
For Channel 2 we hook the probe upto a current meter. This current meter is made by Zalman, and it is pretty accurate. Accuracy is not the first objective of this test though; the objective of this test is to see how voltage changes compared to current, so we need to be able to graph current at the same time as voltage. Usually the easiest method to measure current is by using a current sensing resistor, and measure the voltage drop across of it. In this case an IC outputs that voltage change, and I am able to hook my probe upto that difference and graph it. The difference is proportional to the change in current, and using math functions I could probably graph current to the tee in terms of current measurement, but that is not the point of this test, we just want to see how current changes. This probe is in DC mode as well.
Here are picture of where the probe is hooked up too for current measurement. (notice how nothing is shielded and the ground cable can create problems, this isn't how you want to set it up for ripple measurements)
Notice I am not caring about the ground cable and its noise causing loop, I am worried about just hooking up the probe as I please. Results were actually very good.
As you can see the current is in yellow and the voltage in green. I used IBT, and you can see the current step as well as the initial drop in voltage. You can see every time current changes the voltage dropped if the current rose and the voltage rose if the current dropped. If we were to multiply the voltage and current together we would probably see that the small changes resulted in the same output. Current X Voltage = Watts (power)(P=IV), and Vdroop is implemented to drop the voltage under load when we might have great overshoot from the large change load to idle and idle to load, so that overshoot remains within spec. You can also see the vdroop helps keep the CPU watts withing a specific range. You can easily see that here.
VRM Ripple Testing
Before we begin again I would like to give you the scenario I used to measure ripple. I set the vertical division to 20mv per division (I can do 10mv but opted not too), I used a 1ms horizontal division (I can do 4us if I want) so we can see a good amount of time. Then we use something called a trigger, a trigger waits for a certain time when certain conditions apply, in my case I set the trigger to capture the highest peak2peak ripple, Vpp. If we do not use a trigger we will see a large variety of ripple measurements. Next after I capture the max ripple/noise spike, I then want to know how often that spike occurs. I thought very hard about how to measure the average ripple, but that number is just so small it’s unbelievable. My scope can capture 10,000 ripple points in one excel file, but the output for average was very small, and varied too much. So I opted to use the special pass/fail function my scope has. The pass/fail function records how many times the measured voltage exceeds the maximum threshold I set. In my case I used about 80mv Vpp, with roughly 4 division. The test then tosses out how many times it passed, how many failed, and the total number of points it measured. Then I do each test in best case scenario and then worst case scenario. The best case in our case is at stock (1.475v set with Turbo) with vdroop enabled. The worst case is at 4.2 GHz (on air), 1.63v load(much more than needed), and almost 250watts of current into the VRM input. Of course in both cases I run wPrime, SuperPI, and 3DMarkVantage. All of these tests have greatly varying loads. I will show you the Vpp(peak2peak) ripple. I use a single channel of my scope and set the probe to AC coupling. These same variables will be conserved for future ripple tests on other VRMs.
(please click to enlarge pictures)
Here is an example of how I tested the VRM with the Scope, it’s for the first result, wPrime at stock:
First off Stock Wprime:
MAX Vpp(ripple spike)=64.4mv
Percentage of time above 80mv= 0.000%
Stock SuperPI:
MAX Vpp(ripple spike)=41.2mv
Percentage of time above 80mv= 0.000%
3DMark Vantage CPU Test 1:
MAX Vpp(ripple spike)=57.5mv
Percentage of time above ~60mv= 1.111%
3DMark Vantage CPU Test 2:
MAX Vpp(ripple spike)=52.2mv
Percentage of time above ~60mv= 0.360%
Sin’s Take: The ripple measurements were very decent. In constant load situations with large current steps, such as wPrime we see the most ripple. When situations with smaller variable current steps we start to see load transients, but the VRM maintained good composure. As long as we stay under ~70-80mv at stock with full load situations we are in good company. In fact the ripple spikes are totally acceptable, but what is impressive is that these spikes only occur very rarely. Now we look at when the VRM is under extreme load.
Extreme load VRM Testing:
1.63v under load, with NO vdroop, instead 1.60-1.63v increase under load. ~250-260 watt pull at 4.2ghz.
Worst Case wPrime:
MAX Vpp(ripple spike)=111mv
Percentage of time above ~80mv= 2.341%
Worst Case SuperPI
MAX Vpp(ripple spike)=88.8mv
Percentage of time above ~80mv= 3.870%
Worst Case 3DMark Vantage CPU Test 1
MAX Vpp(ripple spike)=85.0mv
Percentage of time above ~80mv= 3.545%
Worst Case 3DMark Vantage CPU Test 2
MAX Vpp(ripple spike)=83.7mv
Percentage of time above ~80mv= 0.755%
Sin’s Take: Now these results are most interesting. First off with this kind of loading under ~100mv is acceptable. We see that in wPrime we reached 111mv, with big transient loads. Then we look at the other tests, and they are each averaging about ~80mv max ripple spikes. What is most interesting is the pass/fail test results, we see that wPRime even with its high max Vpp still has lower percentage of spikes than SuperPI and Vantage CPU Test 1. Then we see that Vantage CPU Test 2 has some of the lowest percentage of spikes, even though it has almost the same spike as CPu test 1. Then we look at CPU test 1 and SuperPI, CPU test 1 has higher spike and lower percentage of spikes, than SueprPI with highest spikes.
So what can we take from this? We can take that it is important that we see the results of the pass/fail test, as they give us a better idea together of how good the VRM handles each test. Spikes are great knowledge, but having a way to average out the result of what is happening with the spikes is even better. All together I am going to pass this VRM as a whole, as it is just 8 phases for the CPU core. Even though the switching frequency is lower than competitors they used enough bulk output capacitance to counter the ripple. We aren’t seeing spikes over 100mv all the time, and it’s handling transients the best it can. I really don’t have anything to compare it to, but I will in the future. So for now I will only say that it does well, but I cannot call it the best or the worst b/c this is the baseline test. Next I have to figure out how to compare VRMs based on their loading and phase count and capacitance. But compared to the oldVRMs in the P5K deluxe motherboard I used as my first test subject this VRM did great, considering they are both 8 phases for CPU power.
Performance Testing (benchmarks)+ Overclocking:
To begin this section I would like to say that these new 990FX chipset boards are most likely optimized to perform best with the Zambezi processors. When those processors come out I will most likely redo this review, and see how it compares! For now here are the awaited benchmarks!
First my MAX OCs:
MAX BLCK: 300mhz
MAX Memory and CPU Clock(4.4ghz and 1920mhz):
http://valid.canardpc.com/show_oc.php?id=1912601
The test systems:
That concludes our benchmarks for this review. They should give you a slight view of how this board does overclocked vs. stock.
Conclusion:
It is hard to judge a board by how it looks, of course if I just did that I’d give this board a perfect score, as I don’t think it can look better than it does now. So to determine what makes the board tick, tock, and run we have to run tests, do analysis, and analyze our test data. That is what I have done for you. So I will start with what I really liked (pros) and then move on to what could have been better (cons).
What Sin likes:
I love the fact that I can run 4-Way SLI on this baby, that is something that X58 boards should have been able to do without the NF200, but NVIDIA decided otherwise, so I am pleasantly surprised to see it on this board. Of course I don’t have 4 GPUs on hand to test out the 4-way SLI, but I am going to say that it should work just fine for those who want it, even though I cannot say how it performs. I can however say how this board overclocks, which is one thing I really loved. This board was not just stable, it was extremely stable. It passed through all my stability tests without any trouble, and I was able to make changes and test out better timings, speeds, and multipliers without trouble. When my OC goes terribly wrong, the board automatically resets the CMOS, and takes you into the BIOS where your OC settings still are preserved for you to change. Apart from that I was pleased that the VRM took a nice beating and survived, and not only that, but performed upto par. I have to add that I like how feature filled this board is, from connectivity, to the new buttons, POST Debug LED, and SATA power connector the board is equipped for OCing.
What Sin thinks can be better:
I think that Vdroop has been improved upon through the new BIOS releases, but its still not good enough for extreme OCers. At first BIOS F3 I used, the voltage dropped 100mv under load, and with F4B it drops about 70mv. I would like to see this improved on, as compared to other board you can use less idle vcore. I would also like to see some voltage read points one day on all top of the line GIGABYTE boards, as soldering wires to the underside of the socket isn’t ideal. All other manufacturers seem to be implimenting their UEFI GUI based BIOSes into their 990FX boards, but GIGABYTE has not, they haven't even offered their TouchBIOS. I do think they can add TouchBIOS as they have used the 32MBit BIOS IC which is required for it, but we have yet to see it on this platform.
Last word:
GIGABYTE is on top of this board, from releasing BIOS updates, to giving a response on the vdroop, to paying attention to the small details. All black accessories are part of that attention to detail ordeal, as is the Clear CMOS cap, but the all black PCB with the new style heatsinks is like a cherry on top of the ice cream sundae to seal the deal. There is no doubt in my mind that this board lives up to its current expectations. There is not much I can say about this board’s performance until Bulldozer is released, but I can say that from my month of messing around with it, it’s a solid board. It has taken my beatings, my trace cutting mistake, and hours and hours of gruesome voltage testing. My RAM clocked to (1920) almost 2000 MHz stable for 24/7, which is impressive in my book for AMD. I really enjoyed working with this board, and while I try my best to stay objective and find issues, there are currently none to report. I am glad that at release Bulldozer will avoid the BIOS issues that Sandy Bridge had, as these boards are being tested evaluated and used months before release. This being the first 990FX chipset motherboard I have reviewed, I have to say that it is absolutely solid. Other than the vdroop issue, the rest of the cons I talk about are just things this board doesn't have making this board almost problem free.
I would recommend this board for anyone looking for a top of the line 990FX chipset board, as I have a feeling it lives up to all its expectations.
Rating : 9.5/10(Excellent)
I would like to thank GIGABYTE for making this review possible.
