Ever since its auspicious debut as a technology demo at last year’s fall IDF, Lucid’s Hydra chip has been an object of curiosity for us. Could this small start-up firm really create a GPU load-balancing chip that would function as smoothly as SLI and CrossFire, yet allow more leeway in mixing GPUs of different types? They’d taken on a daunting challenge, but they seemed to have a pretty good start on the problem.
Now, a little more than a year after that first IDF showing, Lucid says its Hydra 200 chip is ready to ship in consumer systems. To underscore that point, the firm recently invited us to its San Jose, California offices to experience a Hyrda-based solution first-hand. We came away with our impressions of the Hydra solution in action, along with some of the first performance numbers to be released to the public.
If you’re unfamiliar with the Hydra, I suggest you read our original coverage of the chip, which introduces the basics pretty well. The basic concept is that the Hydra chip can sit on a motherboard, between the north bridge (or CPU) and the PCI Express graphics slots, and provide real-time load-balancing between two or more GPUs. The Hydra accomplishes this task by intercepting calls from a graphics API like DirectX, dynamically dividing up the workload, and then assigning a portion of the work required to draw each frame to each GPU. The Hydra then combines the results into a single, hopefully coherent image, which is then sent to the display.
Several things have changed over the past year, as the Hydra has moved from a technology demo toward a real product with proper driver software. Most notably, perhaps, the first Hydra silicon demoed supported only the PCIe Gen1 standard, whereas today’s Hydra 200 series is PCIe Gen2-compliant.
In fact, the Hydra can support up to 48 lanes of PCIe 2.0 connectivity, with 16 “upstream” lanes to the host north bridge or CPU and 32 lanes intended for graphics cards. Those 32 lanes can be bifurcated into as many as four PCIe x8 connections, with several other configurations possible, including dual x16 connections and a single x16 plus dual x8s. The chip can auto-configure its PCIe connections to fit the situation, so this full range of connectivity options can be exposed on a single motherboard with the proper electrical connections and slot configuration.
To execute Lucid’s load-balancing algorithms, the Hyrda chip also includes a 300MHz RISC core based on the Tensilica Diamond architecture, complete with 64K of instruction memory and 32K of data memory, both on-chip. The chip itself is manufactured by TSMC on a 65-nm fabrication process, and Lucid rates its power draw (presumably peak) at a relatively modest 6W.
Bringing the Hydra to market: Big Bang fizzles?
That’s the hardware, pretty much, which is relatively straightforward. The story of the Hydra’s current status is considerably more complex. Lucid says it has been working with a number of motherboard makers on products that will employ the Hydra chip. MSI has been furthest along in the process and, to our knowledge, is so far the only partner to reveal its plans to the public. Those plans center around a gamer-oriented motherboard based on the Intel P55 chipset dubbed the Big Bang Fuzion.
During IDF, MSI and Lucid announced plans for a public launch of the Big Bang motherboard on October 29. As we understand it, the idea was for the Big Bang board to be available to consumers for the holiday season, complete with a driver that supported both symmetrical—two or more identical video cards—and asymmetrical—a GeForce GTX 260, say, and a GeForce 9800 GTX—configurations. A new capability enabled by Windows 7, the ability to mix Radeons and GeForces in the same GPU team, would be enabled by a driver update at a future date.
However, October 29 came and went, and nothing happened—the Big Bang Fuzion wasn’t launched. Rumors flew that the board had been delayed to the first quarter of next year. A certain someone pinned the blame on Nvidia, to the surprise of no one. The charges were plausible, though, because the Hydra’s capabilities threaten Nvidia’s SLI licensing regime, whereby motherboard makers must pay tribute to Nvidia in order to enable GeForce multi-GPU configurations on their products. It’s conceivable Nvidia might have pressured MSI to delay the product.
According to a source familiar with the situation, MSI’s requirements for the Hydra driver changed abruptly after the press blitz at this past IDF, with the schedule for support of mixed-vendor GPU configurations pulled into October, a dramatic acceleration of the original plans.
When asked for comment on this story, Nvidia spokesman Ken Brown told us that Nvidia welcomes new technology and innovation, especially those that improve gaming, and said he wasn’t aware of Nvidia playing any role in the Big Bang Fuzion delay. Brown reiterated Nvidia’s long-standing position that what Lucid is attempting to do is “very ambitious,” “an enormous technological challenge,” a position the firm has rather curiously communicated at every opportunity. Furthermore, though, he confirmed to us that Nvidia will not block its partners from producing motherboards that incorporate Lucid’s technology.
For its part, MSI issued a statement here (at the bottom of the page) citing a two-fold reason for the delay related not to the Hydra hardware but the drivers: the need for better optimization and stability in Windows 7 and in multi-vendor GPU configs.
Everyone involved seems to agree that the Hydra 200 hardware is ready to go. Based on our brief hands-on experience with the Hydra in Lucid’s offices, though, we think MSI’s trepidation about the drivers may be warranted. Lucid gave us a preview of the mixed-vendor mode in action, and predictably, we ran into a minor glitch: the display appeared to be very dark, as if the gamma or brightness were set improperly, in DirectX 10 applications. This was a preview of that nascent functionality, though, so such things were expected at this stage.
More troubling was the obvious visual corruption we saw in DirectX 9 games when using an all-AMD mix of a Radeon HD 4980 and a Radeon HD 4770. The Lucid employees we spoke with about this problem attributed it to Windows 7, and indeed, Lucid VP of R&D and Engineering David Belz told us that Windows Vista had been the driver team’s primary focus up until the last month. Belz said they had found few differences when moving to Windows 7, but forthrightly admitted the firm might need to look into those differences further. Belz seemed surprised when he asked what percentage of prospective Hydra buyers might wish to run Windows 7 immediately and we answered, “Uhh… 99%.” The Hydra comes attached to a new motherboard, though, so one would think that answer would be rather obvious at this point in time, even if our estimate might be overstated by a few percentage points.
Belz did express confidence that the issues we saw were rather trivial, likely not difficult to fix with software tweaks. Given what we’ve seen of the Hydra in action, we’re not inclined to disagree with that assessment.
Hands on with a Hydra box
Our testing time with the Hydra in Lucid’s offices was limited, but we did have the chance to gather some preliminary performance data and record our impressions of the solution in action.
Pictured above are the guts of a demo system based on the MSI board that Lucid had running. Although it was fully operational, the Big Bang Fuzion board is an unreleased product, so we didn’t get a chance to test with it. Instead, Lucid offered up its own test chassis, which looks like so:
This box below is a regular system with a PCIe cable connected to one of its PCI Express x16 slots. The box above contains a Lucid test board with a Hydra chip and some PCIe slots. Although this looks very different from a fully integrated solution, the system topology in fact is very similar. As you can see, we had a pair of GeForce GTX 260 cards installed in the system.
We were able to test mixed-vendor performance, as well. Above is the Device Manager view with a Radeon HD 4890 installed next to a GeForce GTX 260.
Here’s how the Hydra Engine software indicates that it’s using two GPUs from different vendors. In this case, the display is connected to the GeForce rather than the Radeon, although the choice of display GPU is apparently flexible.
Lucid’s control panels for the Hydra are pretty straightforward. The second one, as you can see, allows the user to enable or disable the Hydra on a per-game basis. Lucid’s software detects installed games on the system and offers this list. Obviously, this particular box has a considerable number of games installed simultaneously. Heck, I was a little surprised it generally worked properly. This system was prepared by Lucid’s own internal QA group, and they have been testing a big swath of the most popular games internally to ensure compatibility and performance.
A preview of Hydra performance
Below are the performance results we managed to squeeze out of the Hydra during our session in Lucid’s offices. Going on-site like this to conduct testing is never our preferred situation, and as one would expect, we were limited by time and circumstance in various ways. We had to choose quickly from a limited selection of games, to limit the number of repetitions in our test runs (we tried to do two for each test, if possible), and to test with the hardware Lucid made available to us.
Still, we had time to test a handful of games on a number of different GPU configs. We even got a preview of the Hydra’s mixed-vendor capability by pairing a GeForce GTX 260 with a couple of different Radeons.
The PC we used for testing was based on a Core i7-920, a Gigabyte EX58-UD3R motherboard, and the 32-bit version of Windows 7. Oddly, the system had only two 1GB DIMMs installed, so one of the Core i7-920’s memory channels wasn’t available. In fact, when we first started testing, only 1GB of memory was available, because the second DIMM was installed in the wrong slot—an artifact of the limited setup time Lucid had for this press demo. We corrected that problem, though, and moved on with our testing with the full 2GB at our disposal.
The first game we’ll look at is Operation Flashpoint: Dragon Rising. We used FRAPS to record frame rates in this game and in FEAR 2, and to keep things very repeatable, we simply stood still at a fixed point in the game and recorded frame rates for 30 seconds. That’s not how we usually test, but it should suffice for our purposes here. Also, all of the multi-GPU configurations below make use of the Hydra for load balancing. We didn’t have the chance to test with CrossFire or SLI for comparison.
In this game, the Hydra delivers performance scaling in earnest, even with asymmetrical and mixed-vendor configurations. No, we’re not seeing linear performance scaling when, say, going from a single GeForce GTX 260 to two of them, but I doubt this game is entirely GPU-limited at this resolution.
Surprisingly, the mixed-mode config with a GeForce GTX 260 and a Radeon HD 4770 outperforms dual GTX 260s. That’s unexpected, but otherwise, the Hydra’s performance is pretty much as advertised.
As I noted before, we did run into visual corruption problems with the 4890+4770 config in the two DX9 games, Operation Flashpoint: Dragon Rising and FEAR 2. We’ve reported the performance results anyhow, for the sake of completeness, but they come with that caveat.
The visible problems of the 4890+4770 config translate into performance issues here, as the pair of GPUs turns out to be slower than a single 4890. Otherwise, though, the Hydra does its thing quite well.
These last two benchmarks use DirectX 10, and as I mentioned, the mixed-mode configs with DX10 apps had much darker displays than normal, for whatever reason. They looked fine otherwise, though, and the performance scaling pictures for these two DX10 apps are very similar. Generally, the Hydra achieves good results once again, although the 4890+4770 pair’s scaling issues remain.
So what now?
Although we encountered some glitches with Lucid’s Window 7 drivers, the Hydra appears generally to work as advertised and to be tantalizingly close to ready to ship in consumer products. No doubt a start-up like Lucid would have benefited from having its products out on store shelves in time for the holidays, but that now seems unlikely. That’s unfortunate, but it doesn’t negate the enormity of Lucid’s apparent accomplishment. Even if we have to wait a few months to enjoy it, the flexibility to mix any two reasonably similar graphics cards and achieve proportionally better performance should be a very nice thing to have.
And make no mistake, Lucid has undertaken a monster challenge. The keys to success are in Lucid’s load-balancing methods, the details of which remain largely a secret. SLI and CrossFire predominantly use alternate-frame rendering, where the load is interleaved between GPUs. This method presents several problems, including the fact that most newer games have frame-to-frame dependencies that will limit performance scaling. Also, AFR doesn’t scale well when the GPUs aren’t evenly matched. Lucid may employ AFR in certain cases, but its best load-distribution methods are clearly finer-grained than AFR and involve per-polygon or per-object divisions of labor. We’ve gotten a taste of how they might work with some visual demos, but for obvious reasons, Lucid is guarding the particulars of their operation. Lucid’s Belz would only reiterate that his solution seeks to understand what the application is doing and then applies the appropriate load-balancing algorithm. The method used may change from frame to frame, as the application’s needs change.
Mixing different GPU architectures, whether it be different generations from one GPU maker or a cross-vendor config, presents additional issues. For things like internal mathematical precision or antialiasing techniques, Lucid must restrict itself to exposing the lowest common denominator between the GPUs. Among other things, that means specialized antialiasing methods like AMD’s custom filter AA or Nvidia’s coverage sampled AA won’t always be available, although the base multisampled modes appear to work just fine.
Lucid’s goal in all of this is to maintain what Belz calls image integrity. This is no small thing. For instance, we’ve documented the differences in texture filtering algorithms from one GPU architecture to the next. If Lucid isn’t careful about how it divides up the workload for a scene, either within one frame or from frame to frame, it risks exposing visual noise caused by the differences in image output. We spent some time peering skeptically at Operation Flashpoint: Dragon Rising with a mixed-vendor config trying to detect any such problems, though, and none were apparent. The on-screen image seemed solid and looked quite good.
Maintaining pristine image integrity will naturally limit the Hydra’s load-balancing options, and so it may present a performance tradeoff. Belz tells us Lucid has chosen image integrity over raw performance in its default driver tuning, but the firm plans to expose an option to allow the user to favor performance, instead. Some folks may not be annoyed by the subtle differences between GPU output, so giving users the option seems sensible.
The reality of these technical issues underscores a point: the Hydra will require continual driver support in order to maintain compatibility with future games. The Hydra doesn’t yet support DirectX 11, for instance, and the company will have to develop that. Lucid remains focused on solid overall support for graphics APIs, though. Belz says Lucid’s approach to fixing any problems its QA team finds in a specific game is to make a general tweak to its driver. Although that fix might resolve an issue with a particular game, he insists that game-specific profiles are not employed.
Nevertheless, Lucid the company will have to succeed in order for the Hydra to become and remain a viable, useful consumer product. We’re hopeful on that front, and we look forward to getting a Hydra-based motherboard into our labs for testing soon.