Slashdot Mirror
AMD Catalyst Driver To Enable Mantle, Fix Frame Pacing, Support HSA For Kaveri
MojoKid writes "AMD has a new set of drivers coming in a couple of days that are poised to resolve a number of longstanding issues and enable a handful of new features as well, most notably support for Mantle. AMD's new Catalyst 14.1 beta driver is going to be the first publicly available driver from AMD that will support Mantle, AMD's "close to the metal" API that will let developers wring additional performance from GCN-based GPUs. However, the new drivers will also add support for the HSA-related features introduced with the recently released Kaveri APU, and will reportedly fix the frame pacing issues associated with Radeon HD 7000 series CrossFire configurations. A patch for Battlefield 4 is due to arrive soon as well and AMD is claiming performance gains in excess of 40 percent in CPU limited scenarios but smaller gains in GPU-limited conditions, with average gains of 11 — 13 percent over all."
First time accepted submitter Spottywot adds some details about the Battlefield 4 improvements, writing that Johan Andersson, one of the Technical Directors in the Frostbite team, says that the best performance gains are observed when a game is bottlenecked by the CPU, "which can be quite common even on high-end machines." "With an AMD A10-7850K 'Kaveri' APU Mantle provides a 14 per cent improvement, on a system with an AMD FX-8350 and Radeon 7970 Mantle provides a 25 per cent boost, while on an Intel Core i7-3970x Extreme system with 2x AMD Radeon R9 290x cards a huge 58 per cent performance increase was observed."
MaximumPC paints this a little bit different. Where only lower end cpu's get a big boost in conjecture with higher end AMD cards.
I guess we will wait and see with benchmarks later today when 14.1 is released.
This is great news for those like me on older Phenom II 2.6 ghz systems who can afford to upgrade the ram, video card, and to an ssd but not the cpu without a whole damn new system. I use VMWare and this obsolete system has a 6 core cpu and hardware virtualization support. Otherwise I would upgrade but only an icore7 or higher end AMD FX-8350s have the same features for non gaming tasks. I can play Battlefiend 4 on this soon with high settings at 1080p would be great!
http://saveie6.com/
We are talking about a real time application, so even without 100% load over a relatively large sampling interval, performance can be degraded.
Let's assume that you have 2 sequential things that cannot be overlapped. CPU setup and GPU processing. You cannot begin CPU setup of next frame until GPU is done with current frame (gross oversimplification, but there are sequencies that bear some resemblence to this).
So a hypothetical CPU takes 1 ms to setup a frame, and then the hypothetical GPU then takes 4 ms to render it. There's 20% usage CPU wise reported on average over as small a sampling interval as 5 ms, and a throughput of 200 frames per second.
Let's say that the CPU now takes 10 times less time, and takes 100 us to setup a frame, and then the GPU still takes 4 ms to render it. You have gotten about a 20% speedup. Even though the CPU was not fully utililized, it did represent a throttling factor in FPS.
XML is like violence. If it doesn't solve the problem, use more.
So gamers get a small boost to their gaming rigs, but that's not *really* the goal for AMD.
The real goal is that AMD demonstrably lags Intel in *CPU* performance, but not GPU. OpenGL/Direct3D implementations cause that to matter, meaning AMD's cpu business gets dinged as a valid component in a configuration that will do some gaming. Mantel diminishes the importance of the CPU to most gaming, therefore their weak CPU offering is made workable to sell their APU based systems. It can do so cheaper than Intel+Discrete GPU while still reaping a tidy profit.
XML is like violence. If it doesn't solve the problem, use more.
On a related note, Microsoft are working on an update to Direct3D to provide a "light weight" runtime similiar to the XBone. Presumably, this will solve the same draw call issues that Mantle deals with.
Unfortunately, it doesn't sound like the update will happen anytime soon - maybe for Windows 9?
Also, it's unclear whether they will back port the update to Windows 7.
https://blogs.windows.com/wind...
I think that misses the point. CPUs aren't the limiting factor in part because game devs limit the number of draw calls they issue to avoid it being a limiting factor (because not everybody has a high end CPU). Mantle may not offer vastly more performance in the short term, but it will enable more in game engines in the long term if the claims DICE and AMD make are accurate. That doesn't get away from the cost of lock in, but like any new release of this sort Mantle may never catch on but it may push DX and GL to change in a mantle-like direction which does then benefit all developers.
NVidia driver quality vs AMD driver quality...
Chances are in the GP's case there's something else going on. Since I've got a 7950 and a FX-6300, and can run skyrim on the max settings while getting no frame hitching. Then again it could just be the 290, since there were TDP issues on some batch runs from what I've heard.
But driver quality? Nvidia's drivers are the reason I went to AMD, after nearly a year of them blaming the end user for constant TDR crashes, then deciding to man up and pay to have rigs in the US shipped to California for TDR testing, then releasing a driver which mostly fixed the TDR issues--where they were very quiet on revealing why it was crashing(all they said was "we fixed it in most cases"). Though a few intrepid people found it had to do with the drivers dropping the core and ram voltages so low that the cards became unresponsive and unstable. Then there was the 5-7 months where the 290-3xx series drivers were causing hardlocks across the board for 400,500,600 series owners.
I do remember when ATI's drivers were shit, I owned a couple of radeon cards during that time. But since AMD bought them out, their driver quality has been increasing quite a bit.
Om, nomnomnom...
Hell, I'm stuck on Nvidias 314.22 drivers because every driver from the 32x and 33x series causes my machine to freeze or restart the driver in a "safe mode". You can read the many links to this horror via https://www.google.com/search?q=nvidia+video5
"When life gives you lemons, don't make lemonade. Make life take the lemons back!" -- Cave Johnson
It's not just that, it's the locking behaviour needed for correct concurrency. There's no point in multi-threading a lot of cores that are going to spend most of their time waiting on another thread to release a lock. Even Mantle is single threaded. i.e. There's a queue for the GPU, one for Compute and one for DMA, on different threads. There aren't going to be 8 threads all using the GPU at once. It'll still be serialised by the driver, just more efficiently than you can do it with existing D3D or GL drivers.
Does it increase my mining hash speed? If you are buying AMD cards for gaming at today's prices, you are an idiot.
Well, assuming it takes off which I don't think it will. If this stuff is truly "close to the core" as the Mantle name and marketing hype claim, then it'll only work so long as they stick with the CGN architecture. It won't work with any large architecture changes. So that means that they either have to stick with GCN forever, which would probably cripple their ability to make competitive cards in the future as things change, or they'd have to abandon support for Mantle in newer cards, which wouldn't be that popular with the developers and users that had bought in. I suppose they also could provide some kind of abstraction/emulation layer but that rather defeats the purpose of a "bare metal" kind of API.
I just can't see this as being a good thing for AMD in the long run, presuming Mantle truly is what they claim. The whole reason for things like DirectX and OpenGL are to abstract the hardware so that you don't have to write a render for each and every kind of card architecture, which does get changed a lot. If Mantle is tightly tied to GCN then that screws all that over.
So either this is a rather bad desperation move from AMD to try and make up for the fact that their CPUs have been sucking lately, or this is a bunch of marketing BS and really Mantle is a high level API, but just a proprietary one to try and screw over nVidia.
Even Mantle is single threaded
Single threaded per queue and you can create as many as you want per application. During setup, you create all of the queues you want to use, then register them. Once registered, the GPU can directly read from the queue instead of making a system call. This pretty much removes those pesky 30,000 clock cycle system calls. Kaveri goes a step further and registers queues directly with the hardware, because each queue item is exactly one cache line in size and the APU shares the cache-coherency and protected memory, the latency between the CPU and GPU is about the latency of the cache, which is around 10ns.
Because prior to Mantle, system calls were used, a 10ns communications latency was completely dwarfed by the 30,000 cycle system call. Now we're getting along the lines of 0 call overhead and 10ns latency. Obviously the GPU must have free resources to start working and some code may need to be ran in order for the GPU to schedule, but the main point still remains.