The NVIDIA Titan V Preview - Titanomachy: War of the Titansby Ryan Smith & Nate Oh on December 20, 2017 11:30 AM EST
The Volta Architecture: In Brief
Before we dive into our benchmark results, I want to spend a bit of time discussing the Volta architecture and the GV100 GPU in particular. Understanding what Volta brings to the table is critical for understanding the performance possibilities of the card, and understanding the GV100 GPU is similarly important for understanding the practical performance of the card.
Volta is a brand new architecture for NVIDIA in almost every sense of the word. While the logical organization is the same much of the time, it's not Pascal at 12nm with Tensor Cores. Rather it's a significantly different architecture in terms of thread execution, thread scheduling, core layout, memory controllers, ISA, and more. And these are just the things NVIDIA is willing to talk about right now, never mind the ample secrets they still keep.
Of the many architectural changes that Volta makes, there are four items in particular that I feel really set it apart from Pascal and help shape its capabilities.
- New tensor cores
- Removing the second warp scheduler dispatch unit & eliminating superscalar execution
- Separating the Integer cores
- Finer-grained thread scheduling
NVIDIA's Big Bet: Tensor Cores
The big story here of course is the new tensor cores. While the Volta architecture has all the makings of a strong HPC architecture even in a traditional context, the specific massive performance numbers that NVIDIA quotes for the Titan V and other Volta cards has come from the use of these tensor cores.
Tensor Cores are a new type of core for Volta that can, at a high level, be thought of as a more rigid, less flexible (but still programmable) core geared specifically for tensor math operations. These cores are essentially a mass collection of ALUs for performing 4x4 Matrix operations; specifically a fused multiply add (A*B+C), multiplying two 4x4 FP16 matrices together, and then adding that result to an FP16 or FP32 4x4 matrix to generate a final 4x4 FP32 matrix. Tensor operations are common in certain types of workloads, but in particular neural networking training and execution (inferencing).
The significance of these cores are that by performing a massive matrix multiplication operation in one unit, NVIDIA can achieve a much higher number of FLOPS for this one operation. A single tensor core performs the equivalent of 64 FMA operations per clock (for 128 FLOPS total), and with GV100 packing 8 such cores per SM, results in 1024 FLOPS per clock per SM. By comparison, even with pure FP16 operations, the standard CUDA cores in a GV100 SM only generate 256 FLOPS per clock. So in scenarios where these cores can be used, we’re looking at the ability to hit 4x the performance versus Pascal.
The flip side to all of this is that the tensor cores are relatively rigid cores. They really aren’t very good for anything except tensor operations, so they are only applicable to certain classes of compute tasks, and right now I don’t know of any graphics tasks that would really benefit from the cores. The benefit to NVIDIA of doing this is that this allows the tensor cores to be very performance dense, both in total performance and in actual die space usage; by lumping together so many ALUs within a single core and without duplicating their control logic or other supporting hardware, the percentage of transistors in a core dedicated to ALUs is higher than on a standard CUDA core. The cost is flexibility, as the hardware to enable flexibility takes up space. So this is a very conscious tradeoff on NVIDIA’s part between flexibility and total throughput.
Meanwhile because tensor cores are brand-new to NVIDIA’s GPUs and because they need to be explicitly called upon, NVIDIA has a bit of a chicken & egg problem here. To make the most of the new hardware, NVIDIA needs developers to write software that taps into the new tensor cores, and in a roundabout way this is one of several roles the Titan V is designed to fill. The $3000 card is still expensive, but it’s the first workstation-class card incorporating these cores, vastly increasing the accessibility of the hardware to software developers. So whereas the first wave of Volta-optimized software has been specific big-ticket items like NVIDIA’s libraries and deep learning frameworks like Caffe2, Titan V will help developers put together the second wave of software.
Scheduling, ILP, & Integers
The second big change that Volta brings to the table is that, at least for GV100, the second warp scheduler dispatch port has been eliminated. Ever since GF104 in 2011, NVIDIA’s architectures have featured two dispatch ports per warp scheduler, allowing for superscalar execution. In other words, their architecture has relied on a degree of instruction level parallelism, requiring the ability to execute a second, non-dependent instruction from a thread in order to get the most out of the hardware.
Volta/GV100, by contrast, is no longer superscalar. Each partition within an SM is now feed by a single dispatch unit warp scheduler, with no opportunity to extract ILP. This means that Volta is a pure thread level parallelism (TLP) design: max utilization comes from maximizing the number of threads active at any given time.
ILP versus TLP is a constant balance, and it’s not unusual to see NVIDIA shifting between the two, especially for a compute-centric GPU like GV100. ILP is nice to have, but extracting it can be difficult. On the other hand while GPUs are meant for embarrassingly parallel tasks, it’s not always easy to generate more threads. So there’s a very real question over whether the performance gains from adding the hardware for ILP justifies the power and complexity costs of doing so.
Meanwhile NVIDIA has also made an interesting change to the Volta architecture with respect to how the integer ALUs are organized. Though not typically a subject of conversation in NVIDIA architecture design, the integer ALUs have traditionally been paired with the FP32 ALUs to make up a single CUDA core. So a black of CUDA cores could either execute integer or floating point operations, but not both at the same time.
Volta in turn separates these ALUs. The integer units have now graduated their own set of dedicates cores within the GPU design, meaning that they can be used alongside the FP32 cores much more freely. The specifics of this arrangement get a bit hairy in light of the fact that Volta isn’t superscalar – so you technically can’t issue INT and FP32 instructions at the same time regardless – but given the fact that most GPU operations take multiple clocks to execute, this allows for much more flexibility than before. NVIDIA notes that it’s especially useful for address generation and in FMA performance, the latter of which we’re taking a look at a bit later.
Finally, and admittedly getting into the even more esoteric aspects of GPU design, NVIDIA has reworked how SIMT works for Volta. The individual CUDA cores within a 32-thread warp now have a limited degree of autonomy; threads can now be synchronized at a fine-grain level, and while the SIMT paradigm is still alive and well, it means greater overall efficiency. Importantly, individual threads can now yield, and then be rescheduled together. This also means that a limited amount of scheduling hardware is back in NV’s GPUs.
This generally doesn’t mean anything for existing software. But for developers who really know their GPUs and threading, it gives them ways to extract performance that couldn’t be done under Pascal’s more rigid SIMT model.
Post Your CommentPlease log in or sign up to comment.
View All Comments
CiccioB - Thursday, December 21, 2017 - linkAbsolute boost frequency is meaningless as we have already seen that those values are not respected anywhere with the use of boost 3.0. You can be higher or lower.
What is important is the power draw and the temperatures. These are the limiting factors to reach and sustain the boost frequency and beyond.
With a 800+mm^2 die and 21 billions of transistor you may expect that the consumption is not that low as for a 14 billion die, and the frequencies cannot be sustained that much.
What is promising is that if these are the power drain of such a monster chip, the consumer grade chips made on the same PP will really be fresh and that the frequencies can be pushed really high to suck all the thermal and power drain gap.
Just imagine a Volta/Ampere GP104 consuming 300+W, same as the new Vega GPU based custom cards.
croc - Wednesday, December 20, 2017 - linkI can't take a titan-as-compute seriously if its double precision is disabled That, to me, makes it only aimed at graphics Yet this whole article is expressing the whole titan family as 'compute machines'...
MrSpadge - Thursday, December 21, 2017 - linkIt has all compute capability enabled: FP16, FP32, FP64 and the tensor cores. The double speed FP16 is just not (yet?) exposed to graphics applications.
CiccioB - Thursday, December 21, 2017 - linkIn fact this one has 1/2 FP64 computing capacity with respect to FP32.
At least read the first chapter of the review before commenting.
mode_13h - Wednesday, December 27, 2017 - linkThe original Titan/Black/Z and this Titan V have uncrippled fp64. It's only the middle two generations - Titan X and Titan Xp - that use consumer GPUs with a fraction of the fp64 units.
Zoolook13 - Thursday, December 21, 2017 - linkThe figure for tensor operations seems fishy, it's not based on tf.complex128 I guess more probably tf.uint8 or tf.int8 and then it's no longer FLOPS, maybe TOPS?
I hope you take a look at that when you flesh out the tensorflow part.
If they can do 110 TFLOPS of tf.float16, then it's very impressive but I doubt that.
Ryan Smith - Thursday, December 21, 2017 - linkIt's float 16. Specifically, CUDA_R_16F.
CheapSushi - Thursday, December 21, 2017 - linkWould be amazing if tensor core support was incorporated into game AI and also OS AI assistants, like Cortana.
edzieba - Thursday, December 21, 2017 - link"Each SM is, in turn, contains 64 FP32 CUDA cores, 64 INT32 CUDA cores, 32 FP64 CUDA cores, 8 tensor cores, and a significant quantity of cache at various levels."
IIRC, the '64 FP32 cores' and '32 FP64 cores' are one and the same: the FP64 cores can operate as a pair of FP32 cores (same as GP100 can do two FP16 operations with the FP32 cores).
Ryan Smith - Thursday, December 21, 2017 - linkAccording to NVIDIA, the FP64 CUDA cores are distinct silicon. They are not the FP32 cores.