Performance Retrospective: AMD’s Radeon HD 7970M

Closing Thoughts: Driver Woes and Looking to the Future

I’ve been accused over the past year or more of being overly harsh on AMD’s mobile graphics drivers, but even now it should hopefully be obvious why we still have concerns. Enduro has improved tremendously since its initial release in early 2012, but AMD isn’t out of the woods just yet. The good news is that starting with the 2013 driver releases, there has been a working Enduro driver available with every desktop driver update. That definitely wasn’t the case in 2012 and earlier, and that’s what will allow AMD to at least fix the driver issues for their user base in the future. There’s still more to the driver story, unfortunately.

As I pointed out, the latest three beta drivers from AMD (13.3, 13.5, and 13.6) have all had DX9 rendering issues on the Alienware M17x R4 when in Enduro mode. I didn’t recall seeing issues with an earlier Enduro 9.01 beta driver, so this is something that likely broke with the 13.x releases. I’ve specifically called out StarCraft II and Skyrim as showing the problem, but it’s also present in 3DMark03 and in fact 3DMark05 and 3DMark06 are now having problems for me as well. I spot-checked a few other DX9 games, and this appears to be something that isn’t working (at least on the M17x R4) across the entire spectrum of DX9 titles. You’d think that in 2013 the number of new DX9-only titles would be relatively limited, but all of the Mass Effect series, Valve’s Source engine games, and all of Blizzard’s titles to date fall into this category, along with a variety of console ports. (That last one will hopefully cease to be an issue when we get the next generation of consoles with DX11.1 era hardware.)

I’m not sure of the precise reason for DX9 causing problems, but AMD last year had to give me a special custom driver with DX9 DLLs that had to be installed in safe mode in order to properly run DX9 titles. Then there was the 12.11 beta Enduro driver…which once again somehow missed getting the DX9 hotfix DLLs included. Even now there are still issues I guess, and it’s hard not to be concerned when something that was fixed at one point keeps getting broken. Whether it’s a question of manpower, management, or money, the simple fact is that the Enduro drivers need to get to the point where every new release—including the beta drivers!—will at least run DX9 games without rendering problems. (I suppose it’s also possible that the DX9 problem is specific to the M17x R4—I know everything works fine on the GX60, but I don’t have any other Intel + Enduro notebooks around to test. If you happen to have a Clevo notebook with 7970M, let me know how it's doing!)

You might think about reverting to an earlier driver, but even that isn’t without problems. Installing a new AMD driver on top of an existing driver usually works, but I’ve had more than a few instances where OpenCL ended up broken in the process. Going the other way (e.g. installing an older driver on top of a newer driver) in my experience almost never works properly. The solution is that you need to first uninstall all AMD drivers, reboot, run a utility like AMD's Catalyst Uninstall Utility or Driver Fusion (the free version is sufficient) and/or manually delete some files and registry keys, reboot again, install the new driver (and reboot a third time), and if everything goes well you’re now running properly on the older driver. Perhaps that doesn’t sound too bad, but even with an SSD-based system installing AMD’s drivers (or uninstalling them) typically takes 5-10 minutes, so the whole process of reverting to a new driver requires around 30 minutes or as much as 60+ on an HDD. [Yes, I am bitter about the number of times I’ve done this in the past few months, thank you very much!] What AMD needs is an option to do a “clean install” that takes care of all of the above for you, preferably without more than a single reboot.

One last point on the drivers and then I'll move on. There was a time (last year) when AMD stated that their goal was to get the Enduro drivers working for all Dynamic Switchable Graphics and Enduro platforms. DSG is what AMD called Enduro before it was Enduro, more or less, so mostly that applies to pre-Ivy Bridge and pre-Trinity laptops. Their current drivers page states, "Please note, that the AMD Catalyst Mobility driver package can only be installed on specific AMD Enduro platforms, that are second generation AMD A-series APU, or third generation Intel Core family based." At some point you have to cut your losses and move on; unfortunately, any owners of Llano or Sandy Bridge laptops with AMD switchable graphics are basically stuck with old drivers or trying to get hacked/modded drivers like those from LeshCatLabs to work.

Okay, I’m done beating on AMD’s driver team. Let me beat up on AMD’s CPU performance and MSI for a moment. Given these results, the GX60 notebook we used for testing at best appears to be severely over-equipped in the GPU realm. It seems MSI could have gotten pretty much similar results by going with a 7870M instead of the 7970M, at least in less strenuous GPU workloads. However, let's return again to the subject of bottlenecks.

Ian has done some testing of gaming performance on the desktop with a large selection of CPUs and a few GPUs, but there’s a lot more going on in the laptop realm that can muddy the waters. On paper, I don’t think a Trinity APU running at 2.4-2.7GHz is too slow for gaming, especially when we’re looking at 1080p and a 7970M (aka desktop 7870). Sure, the A10-5800K is clocked 50% higher than the A10-4600M, but even the older Llano-based A6-3650 tends to provide a reasonable gaming experience and that’s only clocked at 2.6GHz. The only place where Ian’s testing sees a similar 2X increase over the A6-3650 is in triple-GPU configurations, where there’s a lot more going on than with a single GPU.

If we just look at the numbers in Sleeping Dogs, A6-3650 hits 49.2FPS with two 7970 GPUs, which appears to be where the CPU bottleneck becomes the limit. A single 7970M at 1080p can hit 60FPS (in the M17x), so why then is the 7970M with A10-4600M, which has a 2.3-3.2GHz range but usually runs at 2.7GHz during game benchmarking, sitting down at 31.7FPS? Or if you want another point of reference, look at Anand’s CPU results with Skyrim—a 2.66GHz Core i7-920 still manages 182.1FPS at 1680x1050 medium detail; the GX60 even at 1366x768 Medium still couldn’t get above 40FPS!

I can’t point to the CPU as the sole bottleneck, and monitoring the CPU clock speeds indicates there’s something else holding back the GX60. I don’t know if it’s drivers and Enduro again, something with the mobile AMD platform, or if there’s something MSI failed to get right that’s limiting performance. I did update to the latest MSI BIOS at one point and saw 3-5% performance increases in most of the gaming benchmarks, but we’re still way off of where I would expect performance to be. It will be interesting to see if the Richland-based GX60 can manage to do better.

The good news is that when everything works with the HD 7970M (with or without Enduro—though the latter only appears to be an option on the M17x R4, and I’m not sure if Alienware will be keeping the switching option with their new Alienware 17 or not), it’s still a potent mobile gaming GPU. NVIDIA’s GTX 680M is more or less on equal footing these days, the 8970M will improve performance around 5%, and even NVIDIA’s latest GTX 780M isn’t so far ahead to be untouchable—especially when we factor in pricing. Take AVADirect’s Clevo P170SM offering for instance; the new 8970M is a $91 upgrade over the GTX 770M (which should be around 20% slower than GTX 680M), while the new GTX 780M adds another $222. If you need the last 10-25% performance increase (or if you simply prefer NVIDIA’s drivers), GTX 780M is the better buy, but you’ll pay for the privilege.

As for pricing of the MSI GX60, yes, it costs a lot less than an Intel notebook with 7970M/8970M, but it's also severely underperforming in far too many games. MSI's own GE40 (in hand for review as we speak) generally bests the GX60 in gaming benchmarks, and it's priced at $1200 with quad-core Haswell and GTX 760M. (Not to spoil the review, but the 1600x900 LCD is unfortunately not a high quality LCD, sadly.) The MSI GE60 increases the GPU to the GTX 765M and adds what should hopefully be a better quality 15.6" 1080p display for the same $1200 asking price. Or if you want to give Intel's Iris Pro a shot, the ASUS G750JW-DB71 has an i7-4700HQ and GTX 765M for $1400. If you're after a gaming notebook that can run any game you might throw at it, unfortunately the Trinity-based GX60 just falls short in my testing.

And on that note, we should have a true battle royal pitting the GTX 780M in a Clevo chassis against the 8970M in the near future. Dustin also has the new GX60 with Richland APU in hand, so hopefully his results are better than the earlier GX60. We’re probably still a year or more away from getting Titan levels of performance in a single notebook GPU, but rest assured that time will come. With the new consoles targeting GPU performance that’s already below the level of the 7970M, gaming on a notebook is already reasonably easy to achieve if a bit expensive. Give us another process shrink or two and even the next generation of qHD and above laptop LCD resolutions won’t be out of reach.