ATI Radeon HD 6990 Review

Aaah, it's Monday morning, 9AM - with a cup of coffee in my right hand staring at the monitor, looking outside through the office window where there is a clear blue sky and the sun is shining. Sitting 3 meters away from me is "Diablo" our primary mean graphics benchmarking machine, sitting in it, is this red little devil armed by two GPUs and a lot of features. Heck yeah, last week the Radeon HD 6990 arrived here in the office, and is sitting like two buns in an oven waiting to be consumed. It's gonna be a good day.

Radeon HD 6990 ladies and gentlemen is AMD's latest ATI Radeon HD dual-GPU based graphics card. And for now it will be the fastest 'single' graphics card available on the planet. The performance numbers you will see are anything short from astonishing, breathtaking stuff for a wicked product.

For many weeks now the Radeon HD 6990 has been a product of much discussion. Nobody really could confirm what GPUs would be used, how much graphics memory it would get and so on. Well, rest assured. AMD stuck two Cayman XT GPUs (R6970) onto the PCB and allows them to be clock at R6970 speeds as well, in fact you'll get options in clock-frequencies and TDP with the help of a small micro-switch seated on the card, which leads to 2 vBIOS, one with more acceptable TDPs and the other enabling a higher clock frequency mode. Now I've stated it, Cayman XT GPUs, that means the full shader processor count inside that GPU is available, that sums up towards 3072 shader processors (!)

Memory wise, AMD decided not to skimp here either, the Radeon HD 6990 is a card that will be perfectly suited for Eyefinity solutions, say 3 to 5 monitors PER Radeon HD 6990. In such setup it's wise to have a little more memory per GPU, especially with stuff like high anti-aliasing levels in mind. As such the Radeon HD 6990 comes with a flabbergasting 4 GB of graphics memory, that's two GB per GPU.

All in all, we'll have a lot to talk about today, we'll have a quick chat about verbs like Barts, Cayman and Antilles, then we'll describe the architecture a bit better, we'll have a close look at the products with the help of a photo-gallery ... and well that's all followed by power consumptions, heat levels and performance measurements of course.

An Islands Euphemism & Family Tree

AMD/ATIs naming scheme has not been rather clear lately. Next to a recent change in the Radeon HD 6000 numbering scheme, you'll have heard codenames flying around as well. Next to that AMD silently eliminated the ATI branding, which in fact has now become AMD.

Let's first dig our teeth in that and explain what is going on.

A bit of confusion out there on the street, is that people call the one released today "Antilles" -- which in fact are two "Cayman" graphics processors slapped onto one graphics card. As you might remember in the past ATI (now AMD) assigned code numbers to the GPU used on these graphics cards AKA RV770 or something. Though we are quite confident that the design team still uses that numbering scheme, AMD marketing however wanted to give it a little more TLC and as such each GPU family has a codename, and each GPU deriving from that family has a codename as well.

The previous generation GPUs were named after Evergreen trees which was the family codename; products deriving from that range known as Cypress, Juniper, Redwood and Cedar.

For the Radeon HD 6000 series generation the products are codenamed after islands in the Caribbean, in this case Northern Islands depending on how you look at geographical location, of course.

When we subdivide the Northern Islands groups we get small segments of islands, each GPU range is named after an island for all new Radeon HD 6000 series products, ready? Here they are... Barts, Cayman, Blackcomb, Antilles and Whistler.

When the Radeon 6850/6870 was released in October 2010, the GPUs empowering them carried codename Barts. Bart is named after Saint Barthélemy island and will be the performance/upper-mid segment GPU series. But that still leaves Cayman, Blackcomb, Antilles and Whistler.

Today's high-end products are based on a GPU called "Cayman" (after the Cayman Islands) which is the high-end product in the AMD Radeon HD 6900 series, and that's a change as well, as previously the 5800 series, was the most high-end.

And while we're still on the island rollercoaster , today we test "Antilles", named after the Antilles Islands of course. Antilles is a group of islands, get it .. group ? as in multiple. As such Antilles is a dual-GPU graphics card that makes use of two Cayman GPUs, positioned in the Radeon HD 6900 series of product, carrying the consumer name Radeon HD 6990.

So in a nutshell, Antilles is based on two Cayman GPUs = Radeon HD 6990. Let's move onwards to the actual product. First a word or two on the Cayman GPU.

 

An Architecture Change

Last year's released cards in the 6800 "Barts" series were the 6850 and 6870. These cards merely received a small architectural optimization/tweak over the last generation architecture, Cypress. With the Antilles/Cayman products, things have changed a little bit as the fundamental section of the GPU, the Shader processor setup underwent a significant change, and we are still debating whether or not it was a good one.

AMD moved from a VLIW5 (also knows as VEC5) towards a VLIW4 SIMD shader processors setup. We are not going to discuss the VLIW4 thread processor setup in much detail but basically what this means is that AMD went from a VLIW5 configuration, that used four simple SIMD units and one complex t-unit (transcendental unit) in order to build a stream processing unit, to a VLIW4 configuration that uses four stream units which feature equal capabilities, two of them being assigned with special functions.

AMD however claims this change will bring them 10% more performance over the previous thread processor setup, better scheduling and register management. We think it was merely a design change to save on the number of transistors which you can re-use to add more shader processors on the processor die.

Next to this rather significant change, there are more changes to be found on the graphics card. It has upgraded render back ends (ROPS) with a redesigned Z-Stencil and ROP unit architecture consisting of 128 Z/Stencil ROPs, and 32 color ROPs, up to 2 times faster in 16-bit integer operations and two to four times faster in 32-bit floating point operations which will have you in AA performance, much faster GDDR5 memory, and we also spot a series of improved compute features that will help out in performance in that segment.

One other detail that you might find interesting is that when you look at the block diagram, you'll notice that the GPU pretty much looks like a dual-core processor. AMD calls this dual graphics engines.  Anyway, have a peek at the block diagrams if at all interested.

    

Alright, some more generic information to grasp. Each Cayman GPU itself is based on a 40nm fabrication process and harbors a blistering 2.64 Billion transistors. The graphics engine can have up-to 24 shader clusters, with each engine holding 64 shader processors. Do the reverse math and you'll quickly learn that the most high-end GPU will count 1536 shader processors. A bit of an unusual number and we just wonder if there isn't more to be found of them inside that die really.

The Cayman chip has up-to 96 Texture Units and can produce 2.7 TFLOPs of single precision performance.

Memory wise AMD of course stuck to it's fine working GDDR5 setup, and yes it is still based on a 256-bit memory bus.

So what about the R6990 ?

So with this knowledge in mind we can now look into the specifications or the R6990 a little more in-depth, next page please.

The Specifications

So then, it is based on two Cayman GPUs, now we can describe the product very easily. Take a Radeon 6970 and multiply everything by two. And that is roughly the R6990 in a nutshell, features and specification wise, yet now it's one card.

Now before I show you the final specifications of the Radeon HD 6990, you need to understand that the card has two modes, normal and unlocked.

The AMD Radeon HD 6990 graphics card features dual-BIOS capabilities. This feature is controlled by an “Unlocking Switch” sitting closely next to the CrossFireX connector on your board. The switch toggles between the factory-supported Performance BIOS of 375W TDP (BIOS1), and a more extreme Performance BIOS (BIOS2) that unlocks higher clock speeds and up to 450W TDP of performance.

For end users to enable this higher performance BIOS, they will have to remove a label covering the dual BIOS switch and set the BIOS switch to the desired position outlined below:

• Position 1 — 450W Extreme Performance BIOS (BIOS2).
• Position 2 — (default shipping position) — 375W factory-supported Performance BIOS (BIOS1).

Let's place that in a table:

	Radeon HD 6950	Radeon 6970	Radeon 6990	Radeon 6990 Unlocked
Fab Process	40nm	40nm	40nm	40nm
Die Size	389 mm2	389 mm2	389 mm2 x2	389 mm2 x2
Transistors	2.64 Billion	2.64 Billion	5.28 Billion	5.28 Billion
Core Clock / MHz	800	880	830	880
Memory Clock / MHz	5000	5500	5000	5000
Memory Bandwith	160 GBps	176 GBPs	320 GB/s	320 GB/s
Memory type	GDDR5	GDDR5	GDDR5	GDDR5
Shader processors	1408	1536	3072	3072
Compute performance	2.25 TFLOPS	2.7 TFLOPS	5.10 TFLOPS	5.40 TFLOPS
Color ROPs	32	32	64	64
z/Stencil ROPs	128	128	256	256
TDP Gaming	140 W	190 W	350 W	415 W
Idle power	20 W	20 W	37 W	37 W

The Radeon HD 6990 comes armed with an astonishing 3072 shader processors, thus 48 SIMD based shader clusters, split up in a twofold engine per GPU. The domain and shader clock is locked in at 830 or 880 MHz. The card comes paired with 4 GB of memory clocked at (effective) 5000 MHz (2GB per GPU). The TDP of this product is 350W in default mode, and in unlocked mode the card can consume 415W with a hefty game. There's room left for overclocking in the unlocked design though, you may take the card up-to roughly 450W.

Now, since we have all this knowledge let's just compare this product with some others. Let's have a quick comparative overview of some of the specifications representing a certain scope of other performance parts, you'll notice that the differences are just HUGE, I've inserted the 'default' performance mode in there, not the unlocked one :

Specifications	Radeon HD 5770	Radeon HD 5850	Radeon HD 6850	Radeon HD 6870	Radeon HD 6950	Radeon HD 6970	Radeon HD 6990
GPU	Juniper XT	Cypress Pro	Barts Pro	Barts XT	Cayman Pro	Cayman XT	Cayman XT
Manufact. tech.	40nm	40 nm	40nm	40nm	40nm	40nm	40nm
GPU frequency	850 MHz	725 MHz	775 MHz	900 MHz	800 MHz	880 MHz	830 MHz
Stream processors	800	1440	960	1120	1408	1536	3072
Memory Datarate	4800 MHz	4000 MHz	4000 MHz	4200 MHz	5000 MHz	5500 MHz	5000 MHz
Memory bus	128-bit	256-bit	256-bit	256-bit	256-bit	256-bit	256-bit x2
Memory buffer	1 GB GDDR5	1 GB GDDR5	1 GB GDDR5	1 GB GDDR5	2 GB GDDR5	2 GB GDDR5	4 GB GDDR5
Power consumption	108W	151W	127W	150W	150W	190W	350W
Idle Power	-	-	19W	19W	20W	20W	37W

Impressive stuff huh ? The card of course is an up-to-date DX11 class product with a couple of new features. Features wise, the graphics cards will be very similar to the last generation products and is merely an advanced, updated model. However some features like one DVI and four DisplayPort 1.2 monitor connectors are present for a full Eyefinity experience up-to five cards with just one R6990.

Now before you get concerned about the four DP connectors, the card will ship with 3 adapters to enable Eyefinity gaming out of the box for all users. The product will ship with:

• 1x mini DisplayPort to passive single-link DVI adapter
• 1x mini DisplayPort to active single-link DVI adapter
• 1x mini DisplayPort to passive HDMI adapter

The adapter configuration will enable 3x1 gaming out of the box with DVI panels but with additional display adapters or Display Port displays you will be able to drive up to 5 displays in portrait Eyefinity (5x1 Portrait mode) for the rather grand gaming experience.

Power Management - AMD PowerTune

AMD PowerTune is a new technology that opens up a new direction for maximum performance versus TDP. Pretty much AMD can now limit the maximum TDP applied to a card.

AMD can regulate the TDP with the help of active monitoring. Basically you can lower TDP, have it at normal, or increase the thermal headroom of the graphics cards though the Catalyst drivers. The new feature allows the GPU to be designed with higher engine clock speeds which can be applied on the broad set of applications that have thermal headroom.

So at default it will try and keep your power consumption and TDP at a pre-defined baseline, say 200 Watt. But you'll also have a margin to increase that TDP to say +20%, or vice versa, when you want to limit your power draw you can lower the power usage by -20% (or anything in-between).

• AMD PowerTune can enable higher performance that is optimized to the thermal limits of the GPU by dynamically adjusting the engine clock during runtime based on an internally calculated GPU power assessment.
• AMD PowerTune technology also deals with applications that would otherwise exceed the GPU’s TDP like OCCT, Furmark or 3DMark's perlin noise tests. It does so by dynamically managing the engine clock speeds based on calculations which determine the proximity of the GPU to its TDP limit.
• AMD PowerTune allows for the GPU to run within its TDP budget at higher nominal clock speeds than otherwise possible.

So a monitoring function on the graphics card can be used to downclock your card when needed. And the flipside of the coin is that it can be directly adjusted by the user using the AMD Catalyst Control Center and used for tweaking and overclocking as you can allow for a more aggressive power containment (and therefore more aggressively limit power and heat) or be used by enthusiasts to relax the enforcement of factory thermal constraints on their AMD Radeon HD 6900 Series Radeon graphics card and squeeze every last bit of performance - at the cost of a much higher TDP of course.

Mind you that if you use ATI Overdrive for tweaking the R6990, you'll need to tweak and apply the setting for each GPU independently.

What's That New Switch?

If you look at the photo below you'll notice a tiny micro-switch next to the Crossfire connector. The R6990 cards have one firmware flashable BIOS and one (non flashable) default BIOS, with the switch you can select BIOS 1 or 2.

AMD initially implemented the feature likely to prevent the RMA rate. They know very well that the enthusiast community often re-flashes their cards, often unsuccessfully after which they enter a very expensive RMA procedure at AMD's cost.

With the R6990 AMD decided to expand that function a little as you'll get control over clock-frequencies and TDP. So the dual-BIOS feature now became an “Unlocking Switch”. The switch toggles between the factory-supported Performance BIOS of 375W TDP (830 MHz - BIOS1), and an Extreme Performance BIOS (880 MHz - BIOS2) which unlocks higher clock speeds and up to 450W TDP of performance. it also applies a slightly higher voltage.

• Position 1 — 450W Extreme Performance BIOS (BIOS2).
• Position 2 — (default shipping position) — 375W factory-supported Performance BIOS (BIOS1).

It's definitely not a bad thing to have on any graphics card really and we certainly appreciate the implementation.