NVIDIA's 1.4 Billion Transistor GPU: GT200 Arrives as the GeForce GTX 280 & 260
by Anand Lal Shimpi & Derek Wilson on June 16, 2008 9:00 AM EST- Posted in
- GPUs
Derek Gets Technical: 15th Century Loom Technology Makes a Comeback
Because it's multithreaded...
Yes I know it's horrible, but NVIDIA has gone a bit deeper in explaining their architecture to us and they thought borrowing terminology from weaving was clever. But as much as that might make you want to roll your eyes, the explanation of how things work that is enabled is worth it.
In cloth weaving, a warp is the vertical group of parallel threads that are held taught while the weft are the threads passed through these. I suppose it makes sense, then, that NVIDIA decided to call their grouping of parallel threads to be executed on an SM a warp.
See the group of threads that hang from the top of this loom? That's called a warp.
With each SM having 8 SPs and G80 having two SMs per TPC (for a total of 16 SPs) it looked like a natural fit to execute quads across these 16 scalar units. We learned that this was not the case, but seeing the grouping of three SMs per TPC in GT200 still looked a little funny until we really learned more about how things are scheduled on NVIDIA's unified architecture. Each SM's scheduler picks a new warp to work on every clock cycle, and since the scheduler runs at half speed this ends up being every other clock cycle from the perspective of the rest of the SM. Each warp is made up of a group of 32 threads (in pixel shaders this is a group of 8 quads) that share an instruction stream (a shader program or a kernel if you're talking about GPU computing).
32 threads in a warp, issued in two groups of 16 threads - one such group is depicted above
Different warps built from threads executing the same shader program can follow completely independant paths down the code, but every thread within a warp must be exectuing the same instruction. This means that our branch granularity is 32 threads: every block of 32 threads (every warp) can branch independantly of all others, but if one or more threads within a warp branch in a different direction than the rest then every single thread in that warp must execute both code paths. Each threads only retains the result from the path it was supposed to follow probably by using the branch result to dynamically predicate the apporpriate path per thread. Each SM can have 32 (up from 24 in G80) warps in flight at the same time (for a total of 1024 threads in flight per SM). With 10 TPCs each containing 3 SMs, that's up to 30720 threads in flight on a GTX 280.
Each SM with it's 8 SPs and 2 SFUs is capable of processing up to 8 ops per clock (MAD (a fused floating point multiply and add) being the most complex) in the SPs and 8 ops per clock on the SFUs (which are made up of 4 floating point MUL units and other logic). Warps are executed on SMs in 2 groups of 16 threads (probably four quads for pixels) over four clock cycles.
This is where it gets interesting.
The SPs and SFUs are scheduled on alternating clock cycles from the perspective of the SM scheduler. They execute independent warps. SPs are scheduled on one clock cycle and SFUs are scheduled the next. Each is able to complete the processing of an entire warp before the next time it is scheduled, which happens to be four clocks (relative to the SPs and SFUs) after they it was scheduled with the last warp.
But wait ... aren't SMs supposed to be able to support dual-issue MAD+MUL? Well, back when G80 launched, Beyond3d did an excellent job of exploring the reality of the situation. With the fact that the SFU handles the "dual-issued" MUL and shares its time with other responsibilities like transcendental calculation and attribute interpolation, the ability of the hardware to actually accelerate cases that could take advantage of dual-issue was significantly reduced. But this scheduling revelation makes it clear that the problem is much more complex.
In order for code that could benefit from dual-issue hardware to take advantage of G80 or GT200, a warp must be scheduled on an SP for the MAD and then it must be re-scheduled on the SFU before it hits the SPs again. If the SFU is too busy handling attribute interpolation or transcendental math, then our warp will be rescheduled on the SPs to calculate the MUL and we will have lost our potential dual-issue speed up.
NVIDIA tells us that they "had some scheduling/issue problems with getting the MUL in the SFU to work consistently per clock. We fixed that in GT200." This makes perfect sense in light of what we now know about scheduling on G80/GT200. What they needed to do to improve the ability of their hardware to speed up cases where a MAD+MUL dual issue would help was to make sure that these cases were properly prioritized to execute on the SFU. We don't see MAD+MUL in every line of code, so giving these cases higher priority on the SFU than its other duties should help to optimize the utilization of the hardware. They do have to be careful not to start just running random MULs on the SFU because those special functions do still need to get done, but pushing a select subset of MULs onto the SFU (where they directly follow MADs that have just been executed on SPs) should definitely help in certain cases.
Really the main focus for NVIDIA is proper utilization and prioritization. Everything for any given frame needs to get done at some point or other, so simulating "dual-issue" shouldn't take priority over other things that need to get done that might be more important to completing the next frame. But organizing how to handle it is a big deal. Real time compilers are a big part of that, but internal thread management and scheduling in an architechure this wide also cannot be ignored. NVIDIA says they can now get 93-94% efficiency from their "dual-issue" implementation in directed tests and that this is significantly higher than on G80. Real world results will be lower, but the thing to remember is that the goal is simply to make maximally efficient use of the hardware availalbe. Just because the SFU isn't assisting in a MAD+MUL every clock doesn't mean it isn't doing something important.
This whole situation leaves us with mixed feelings. The hardware itself is not capable of "dual-issue" as it is understood in an architectural sense. The obfuscation of graphics hardware technology in a competitive industry has been the norm for the past decade, and we can accept this. We would prefer to know what the hardware is actually doing, but we are more than happy to have an explanation of something as a hardware feature where the hardware merely simulates the effect a specific architectural design has. But in the case where the hardware doesn't perform in nearly the same way as it would if the feature had actually been implemented as a hardware feature, we just can't help but be a little disappointed.
And we are conflicted about this because NVIDIA's design is actually very elegant. Attribute interpolation will always need to be done, and having hardware set aside for complex math is also very useful. But rather than making a dedicated fixed function interpolator or doing taylor expansion of complex math on SPs, NVIDIA built hardware that could serve both purposes and that had time left over to help offload some well placed MULs within the instruction stream of running programs.
If NVIDIA had been as open about their architecture as Intel is about their CPU designs, we could not have helped but to be impressed by this. The "missing MUL" wouldn't have been seen as a problem with NVIDIA's dual-issue "hardware"; we would have been praising NVIDIA's ability to schedule and multitask the different units within their SMs in order to improve utilization.
Facebook
Twitter
RSS
108 Comments
View All Comments
skiboysteve - Tuesday, June 17, 2008 - link
FANTASTIC write up on fine-grained TMT. I was unaware about this threading technique and was always thinking of this in class or whenever someone would talk about hyperthreading. this technique was literaly in my head for well over a year and I didn't know what it was called or that it even had a name. I always thought there had to be a more elegant way than hyperthreading to do multithreading down at the chip level without doing the OS style time slicing.i was sitting there wondering how the hell the schedule and run these SPs and then bam whole page about it
really appreciate the effort that goes into researching the core of these chips. i know not everyone likes it but for guys that are educated and work in the field its really interesting
DerekWilson - Tuesday, June 17, 2008 - link
remember though that this type of fine-grained TMT only has payoffs in systems running millions of threads concurrently.on an OS you'll see hundreds or even thousands of threads on heavily used systems, but there still wouldn't be enough concurrent action to justify this type of architecture for general purpose computing.
of course, as developers push towards an effort to thread their code as much as possible, who knows what architectures might be worth exploring on the desktop ...
coder0000 - Tuesday, June 17, 2008 - link
Very well written! A couple of points:1) Last week at WWDC Apple announced OpenCL as an alternative to CUDA. It's a C99 based HLL for creating compute kernels that can be deployed to GPU's and CPU's. Today Khronos officially announced a working group for this, and NV is a part of the committee. As such, your wish for an industry standardized compute language similar to CUDA that runs on all platforms and vendors HW may not be so far off.
2) I believe your interpretation of how multiple threads simultaneously execute in an SM is incorrect. Per thread context switching is not free, and you would never be able to execute a different thread every cycle in the manner described. There is far too much context that needs to be swapped out, and there would be significant power implications for doing that, in addition to the latency. Instead, I believe what NV is claiming is that any given SP executes a single thread. All threads in the SM can all be a single warp, but you can also have multiple threads (one per SP) all executing simultaneously in an SM.
DerekWilson - Tuesday, June 17, 2008 - link
1) I haven't had a good chance to look at OpenCL, but I certainly hope that if it's everything everyone is saying it is in the comments here that it takes off in a bigger way than CUDA :-)2) it does not context switch per thread -- warps define a context, and you have 32 threads grouped together. these threads all share the same instruction stream, which is why if threads in a warp take different directions on a branch all 32 threds must follow both paths.
NVIDIA has flat out stated that every schedule clock a new warp is scheduled and that it takes 4 clock cycles to process one warp on an SM. For both of these to be true, we conclude that the scheduler alternates scheduling SPs and SFUs on altenating clocks which means the SPs would be scheduled every 4 clocks relative to itself.
On 8 SPs per SM, you some how need to execute 32 threads in 4 clock cycles. This makes sense if you execute 4 threads per SP in some way. The details at this point are fuzzy though.
regardless, if an SP executes 4 different threads from the same warp, there is no need to context switch to execute any of these threads -- again, threads in the same warp share context.
skiboysteve - Tuesday, June 17, 2008 - link
could be a large explanation of the 2x register file size. and remember that the SP doesn't have to worry about the context switch, the SM handles having the data in the right placeanandtech02148 - Monday, June 16, 2008 - link
From this conclusion, Amd seems to be the shrewd player, let nvidia and intel duke it out in the high voltage, heat, meaningless speed gpu while Amd can pull something like its first dualcore or athlon64 for the win.this new beast from Nvidia will have how many developers making games for it right away? i'm guestimating maybe 2yrs-4yrs down the road we'll see a decent title that take full advantage of this hardware.
by then Amd will have something of a midrange that can more than handle the games.
2 things nvidia could work on that it already has, the ps3 market, and small graphic devices to improve profits. shrink the ps3 gpu further so Sony can shrink it's machinel and sell more.
PrinceGaz - Monday, June 16, 2008 - link
The GT200 core may be a technical masterpeice in terms of actually making something that big which is fully functional on GTX280 cards, but it seems to me the penalty of fabbing it at 65nm negates much of the benefits of such a wide GPU.They've had to drop the clock speeds throughout presumably because of the ridiculous amount of heat such a large core generates, which means the ~60% performance advantage in current games over the G80 core at similar clock-speeds is somewhat reduced.
Given that ATI are not producing their 55nm cores in AMD's fabs but instead are getting them churned out reliably elsewhere, nVidia have made a mistake this time around in having their high-end product rely on previous-generation fabrication as it makes it run too hot to allow the clock-speeds needed for it to be the product it should be. There is always a risk in transitioning to a smaller fab technology, and nVidia suffered badly in the past by doing so too early, but with a chip the size of the GT200, they really should have gone to 55nm even if it meant a delay of a month or three, whilst the smaller cut-down derivatives were rolled out first.
ekpyr - Monday, June 16, 2008 - link
Great article, but what about the microstuttering issues present in Nvidia's 9800GX2 cards (both SLI and Quad-SLI)? There is very little discussion on this, but I've seen some benchmarks where the FPS floor is 4fps with the 9800GX2s. Can you add a subjective review of whether or not the actual gameplay is smoother with the GTX280s across these games? Aggregate numbers may say one thing, but I've returned a 9800 GX2 Quad-SLI setup because it was unable to handle the incredible amount of texture loading that was done in Age of Conan (2560x1600 4xAA 'High' settings = 4fps). The 8800 GTX Tri-SLI configuration I am currently using is more resilient to microstuttering with its increased bus and memory capacities, but I'm very curious about the GTX280s and their increased memory and bus on texture-heavy games like Age of Conan.DerekWilson - Monday, June 16, 2008 - link
the only game that came close to having this issue with quad sli for us was oblivion.in that game at high res lag and stutter are unbearable and the game is unplayable.
we didn't notice any stuttering issues with a single GX2.
i'm working on some analysis tools to show details like this better in future articles.
TheJian - Monday, June 16, 2008 - link
I find it humorous that nobody discusses the fact that the shrink has already taped out and will likely be out in two months or just after. This humongous chip was only released so that when AMD releases in the next few weeks they will be behind still in single GPU cards. This is basically what Intel does to AMD every time AMD has a better chip. For all intents and purposes this is a PAPER release of what will come in 2-2.5 months (In Intel's case they just show you what will be out 6 months from now, and a large portion of people don't buy an AMD because Intel might be ahead by xmas...LOL - works like a charm every time AMD is ahead). THE DIE SHRUNK CHIP! Most likely with faster speeds. I suspect they'll come with "ULTRA" version first (and stick it on top of the price heap, so as to not kill all FAT cards in the channel already) and then filter down as these big suckers leave the channel. That's if they even plan to sell more than a few of these to begin withat 65nm. It's only out there so AMD won't look any good in two weeks.MIND SHARE is everything, which is why Intel's KING of the paper launch when behind strategy. They've even went to doing it for all chips no matter what now. Nehalem scores 6 months before availability. AMD's marketers have no clue an should be fired. You have to play the same DIRTY game as your enemy or you've already lost. If AMD had half a brain in their head they'd paper launch an ultra or 2x4870 version for the same reason...LOL. Then claim "our 4870x2 makes nvidia look like crap for $600"...ROFL. Who cares when it's available, just say it. Having said that, Nvidia will wipe the floor with them in 2 months anyway on a 2xGTX280 that's die shrunk. Which is all they are doing today...BUYING TIME!