Couple of weeks ago at the Siggraph 2012 conference held in Los Angeles, both AMD and NVIDIA introduced the next generation of its professional graphics cards based on 2012 graphics processors. Both companies picked up different approach, hoping to yield the best results. In this article, we explain what you can expect on the workstation market in 2012/2013.
The workstation market is of paramount importance for NVIDIA. The Quadro family of products represents one of biggest cash cows for the company and with margins exceeding 70% – Quadro is one of most lucrative products on the market. As a radical contrast, FirePro barely makes a dent in AMD's revenue.
To us, this is a weird decision on the side of old AMD management. Back in 2006, AMD paid $5.4 billion for ATI Technologies and had to pay additional 500 million dollars for all the IP related to its professional graphics line-up (formerly known as FireGL), bringing the total cost of acquisition to $5.9 billion. However, that investment wasn't matched by investment in new products, driver team in Germany was first laid off, then re-hired, then again reorganized to the point of inexistence… while NVIDIA kept on launching new Quadros and earning tons of greenbacks from safe corporate harbors.
NVIDIA and Intel learned long time ago that the strength of the company doesn't lie in the consumer and retail markets, but rather in channel and more importantly, corporate customers. Luckily, with the new management at AMD, there are new winds blowing inside the company and recent hires are slowly but certainly changing the stalemate from within.
AMD FirePro Arrives on Southern Islands
The new FirePro cards (W and S series) are the first truly new graphics cards in almost two years. With the exception of V5900 and V7900, the bulk of FirePro cards were based on the 2009 Evergreen GPU architecture, no match for GT200b and Fermi-based Quadros. With the Southern Islands GPU family (Graphics Core Next – GCN architecture) being more compute oriented than Evergreen or Northern Islands, AMD stands to capture higher market share than it was the case in the past.
The 2012 FirePro lineup includes a grand total of seven cards, all powered by the 28nm GCN chips – W600, W5000, W7000, W8000, W9000, S7000 and S9000. The W7000 and S7000, as well as the W9000 and S9000 are identical boards, with the difference being the cooling solution: S-Series is intended for installation inside servers (similar to NVIDIA Tesla Cxxxx and Mxxxx series of cards).
Cape Verde GL GPU Boards
On the consumer desktop, Cape Verde GPU (640-core) is known as the Radeon HD 7700 Series. Notebook users will know this GPU as the Mobility Radeon HD 7800M e.g. Heathrow. Workstation users typically receive the 'GL' suffix and that's about that.
W600 is a low-power card (75W) intended to drive as much displays as possible. By combining the Cape Verde GL GPU and 2GB of GDDR5 memory, this board will drive six displays through mini-DP 1.2 connectors. It is capable of supporting 4K resolution e.g. 4096×2304 pixels at 30 bits per pixel at 60Hz ('billion color' displays). The board retails for $599.
Second low-power card is the W5000. This board represents almost equally product as the W600, without the six display output. W5000 carries two full-sized DisplayPort 1.2 connector and a DVI-I. Power consumption remained the same, with a 75W thermal limit. The W600 and W5000 are the only cards in the whole 2012 line-up that can be entirely powered by the motherboard. Performance-wise, we're looking at a 640-core part with 2GB of GDDR5 memory connected via 128-bit interface (102.4 GB/s). Furthermore, this board is $149 more affordable than the W600. At $449.95, we could say the board represents a steal in its pricing category.
W7000 succeeds the V7800, a very popular card. According to our sources, the V7800 was a very popular part due to its single-slot heatsink. This board also allegedly inspired NVIDIA to launch Quadro 4000, their own single-slot board. W7000 is based on Pitcairn GL chip, enabling the full feature set of higher performing parts for system integrators with limited space in their chassis. Just like S7000, the W7000 is based on Pitcairn GL GPU (1280 core), paired with no less than 4GB of GDDR5 memory. AMD rates the performance of at 2.4 TFLOPS in single and 152 GFLOPS in double-precision (IEEE 754-2008). The W7000 will set you back for $899.95.
For the very top of the line we have three parts based on Tahiti GL GPU. W8000 combines Tahiti GL silicon with no less than 4GB of GDDR5 memory. For its price ($1599), this board has no competitors in terms of frame buffer, which is something a lot of professionals will appreciate. Its direct competitor, Quadro 5000 (Kepler) offers only 2.5GB of GDDR5 memory for its current price of $1,722.18. Do note that AMD reduced the memory interface to 256-bit, keeping the 384-bit for its top-end part, the W9000. The number of cores was also reduced, from 2048 to 1,792. From the outside, the board requires two 6-pin power connectors (225W TDP) and offers four DisplayPorts 1.2 (plus stereo out/3D).
The absolute top end is a breathtaking part. Breathtaking price, that is. Rivaling Quadro 6000 6GB ($4,125 MSRP, available in OEM package for as low as $2,750), the W9000 is the first professional part from AMD that really goes for NVIDIA's jugular. This board carries several firsts – the first time in history to see a professional card clocked as high as the desktop versions, which is a no small feat: both Radeon HD 7970 GHz Edition and the W9000 are clocked at 1GHz, resulting in amazing performance of 4 TFLOPS Single / 1 TFLOPS Double-Precision. 6GB of GDDR5 memory is clocked at 1,375 MHz in QDR mode e.g. 5.5 billion transfers per second. 264GB/s of available compute/video bandwidth is the highest single-GPU bandwidth in history.
Thus by any account, AMD should be able to beat Kepler-based Quadro line-up like no tomorrow.
Kepler-based Quadros Carry Two Faces: Desktop and Mobile
Unlike AMD, NVIDIA is battling the different battle altogether. As we wrote in our previous articles, Kepler architecture is divided into three different GPUs: GK104, GK107 and GK110. With the GK110 arriving as Tesla K20 only in December of this year (Quadro variant isn't expected before April 2013), it is obvious that NVIDIA had to shift its professional workstation lineup.
The company launched a six mobile (K500M, K1000M, K2000M, K3000M, K4000M and K5000M) and single desktop part. Given that we are looking into desktops, we'll focus on the K5000 4GB. Available in September (AMD is shipping all of its parts now, and if you're a member of privileged media – you might have received those parts already), NVIDIA pitches the K5000 4GB as a part of its Maximus platform, giving you incentive to combine Quadro and Tesla parts. Here also lies the major difference between the two companies – AMD guns for highest single-board performance, while NVIDIA is pitching for pairing multiple cards (earning more money in the process).
Quadro K5000 4GB is based on GK104GL GPU, which you already met as the chip behind GeForce GTX 660/660Ti/670/670M/680/680M/690, as well as Tesla K10 8GB. This board utilizes all 1536 cores packed inside the GK104 chip, offering raw performance of 2.15 TFLOPS in Single and mediocre 90 GFLOPS in Double-Precision (IEEE 754-2012). Even though this will not be the fastest Kepler-based Quadro, it easily beats Fermi-based Quadro 5000 2.5GB and 6000 6GB. The board consumes mere 122 Watts, enabling NVIDIA to use just a single 6-pin power connector. Memory bandwidth stands at 173 GB/s and if you're interested, you need to prepare no less than $2249, slotting between FirePro W8000 and W9000.
In our conversations with several key software partners, they were not happy by looking at NVIDIA's empty Quadro roadmap – meaning AMD has a pretty good chance to fill the void of the 6GB cards, given that 6000 6GB is around two years old.
Initial Results – FirePro Can't Compete Even Against Fermi-based Quadros
Given that AMD skipped us from their FirePro briefing, yet alone review list, we had to wait until the first results came out. Thanks to our colleague Joel at HotHardware, you are able to see that a $4000 FirePro W9000 cannot keep up with Kepler-based Quadro 6000 in numerous tests, and in some tests the W9000 suffered an indignity of being defeated by a $749.99 Quadro 4000. Naturally, in some tests AMD went forward, but the general feeling is that the Southern Islands family has trouble competing against Fermi. When Kepler-based Quadro arrives in our lab in a few weeks' time (early to mid-September), we'll revisit the platform.
There's plenty of other SpecViewPerf 11 tests where both parts were competing more teniously against each other, but in order to really see how the Southern Islands performs against Kepler, look at the quite an unusual comparison.
Apples to Oranges – Desktop AMD FirePro W9000 Beaten by Laptop-based Kepler – K5000M
While we aren't currently able to measure apples-to-apples e.g. workstation parts, we can give you an insight into what Kepler-based Quadros can do. K5000M 4GB is almost identical to its desktop version, with some sacrifices made in order to fit the 100 Watt thermal limit. 22% less power means that NVIDIA had to fuse off one SMX cluster from the GK104, resulting in 1344-core part. Furthermore, the 4GB GDDR5 memory is clocked down to just 750MHz QDR (three billion transfers per second), giving 96GB of compute/video bandwidth. Our benchmark system was consisted out of Intel Core i7-2960XM, 16GB DDR3-1333 memory and a single SSD. Note that this system is significantly weaker than Intel Core i7-3770K processor. 16GB DDR3-1333 memory and the solid state drive were almost identical in both cases.
In SPEC ViewPerf 11, the results are as follow:
As you can see, a notebook part with a 32nm notebook CPU operating at 2.7GHz (3.7GHz Turbo) was competitive against a 22nm desktop CPU operating at 3.5GHz with a 3.9GHz Turbo mode.
There's plenty of battles coming through, and we'll closely follow to see just how AMD can lead this battle with NVIDIA. From thes initial results, AMD has plenty of work ahead to optimize the drivers in order to get their parts competitive against almost two year old Quadro cards, yet alone the brand new, Kepler-based Quadro K5000. NVIDIA has proven its position as the undisputed leader (over 90% share) and AMD has to go on an aggressive tangent to build its challenge. For now, we cannot recommend investments valued in thousands of dollars for parts which do not beat two-year old competitors. Then again, we're not exactly happy to see NVIDIA continuing to sell numerous Fermi-based boards well into 2013.
Then again, by looking at the performance above, who could blame them?