NVIDIA just delivered their first Volta-enabled DGX-1 systems – great news for those who need the additional compute power of GV100 versus GP100:
|FP32 Compute||10.6 TFLOPS||15.0 TFLOPS|
|FP64 Compute||5.30 TFLOPS||7.50 TFLOPS|
|Memory Bandwidth||720 GB/s||900 GB/s|
Wait, you say, that’s an interesting qualifier. Who doesn’t “need the additional compute power…?” Did someone hack into Nick’s blog account and post on his behalf? Or has he become a Luddite in his dotage?
Nope, no, I still think more compute is generally better; but it is past time to question the architecture of these systems with huge, discrete GPUs connected to the world by buses. The problem with DGX-1 is that those GPUs are hungry! They need to be fed! And they can only sip data through the tiny soda straw known as the PCI Express bus.
For perspective, let’s compare these chips to G80, the first CUDA-capable GPU. Let’s set the stage by observing that G80 was the largest ASIC NVIDIA could feasibly design and fabricate in 2006, straining the limits of contemporary fabrication technology – a classic “win” chip. It had 684M transistors, a theoretical maximum performance of 384GFLOPS for single precision, and no support at all for double precision. GP100 and GV100 respectively have 22x and 31x more transistors, and 27x and 39x more single precision performance than G80. But the bandwidth to deliver data to and from these GPUs has not been increasing commensurately with that performance.
Here’s a table for all 3 GPUs – G80, GP100 and GV100 – that highlights the FLOPS/byte of bandwidth for device memory (attached to the GPU), NVLINK (NVIDIA’s property GPU-GPU interconnect), and PCI Express:
|GPU↔GPU memory||84 GB/s||720 GB/s||900 GB/s|
|GPU↔GPU||n/a||20 GB/s||20 GB/s|
|CPU↔GPU||3.1 GB/s||3.1 GB/s||3.1 GB/s|
The 3.1GB/s figure comes from dividing the available PCIe bandwidth by the number of GPUs in the system. Two 16-lane PCIe 3.0 connections are about 25 GB/s observed, and there are 8 GPUs.
As the number of FLOPS per byte of I/O diverges, the number of workloads that benefit from more FLOPS diminishes. Googling around for literature on FLOPS/byte, I ran across this 2011 presentation by Peter Kogge entitled “Hardware Evolution Trends of Extreme Scale Computing.” For anyone in the GPU business, the first sign that something’s amiss crops up in Slide 3, which cites “1 byte/FLOP” as the “classical goal.” Even G80’s device memory fell well short of that goal with 4.5B/FLOP (here, higher numbers are worse). I prefer this framing because it adopts the viewpoint of scarcity (bytes/FLOP – getting data in and out for processing) rather than abundance (FLOPS/byte – having lots of processing power to bring to bear on data once it is in hand).
The presentation is from 2011, but still very relevant: after reviewing Moore’s Law and the rise and fall of Dennard scaling, and the preeminent importance of power dissipation in modern computing, the concluding slide reads in part:
- World has gone to multi-core to continue Moore’s Law
- Pushing performance another 1000X will be tough
- The major problem is in energy
- And that energy is in memory & interconnect
- We need to begin rearchitecting to reflect this …
- DON’T MOVE THE DATA!
“DON’T MOVE THE DATA” has been good advice to everyone who’s had the data for decades (in 1992 I wrote a Dr. Dobb’s Journal article that focused on hand-coding x87 assembly to keep intermediate results in registers)… but the advice has more currency now.
Moving The Data on CPUs
The data/compute conundrum finds expression on modern multi-core CPUs, too. Each core on a modern x86 CPU has ILP (instruction level parallelism) of 5, meaning it can detect parallelism opportunities between non-dependent instructions and execute up to 5 instructions in a single clock cycle. Latency to the L3 cache is about 50 clock cycles. So a CPU core can perform dozens of FLOPS on data in registers during the time it takes for the L3 to service a load (conservatively – 2 of the 5 pipelines can do 8 FLOPS per instruction via AVX). And that’s assuming the data was in cache!
As an aside, this observation helps explain why “optimized” numerical Python code is still dead slow. Python is interpreted, so has a library called Numpy that wraps vectorized implementations of operations that do things like element-wise addition or multiplication between arrays. But for arrays that don’t fit in cache (and to some extent, even arrays that do fit in cache), it is very inefficient to do multiple passes over the data if the computation could have been fused into a single pass. The code spends all of its time moving data, and very little time processing it.
DON’T MOVE THE DATA!
A Gift From Heaven: Deep Learning
Which workloads, pray tell, require endless FLOPS per byte of I/O? Or turn it around and ask, which workloads still thrive when there is barely any I/O per FLOP? NVIDIA hasn’t been shy about trumpeting its solution to this problem: deep learning! Training a deep learning network entails refining floating point weights that roughly represent neurons that “learn” as they are trained on the data. As long as the weights can reside in device memory, only a modest amount of I/O is needed to keep the GPU busy. In retrospect, NVIDIA is extremely fortunate that deep learning cropped up. Without it, it’s not clear what workload could soak up all those FLOPS without the GPUs starving. The importance of machine learning as a workload helps explain why GV100 contains purpose-built hardware for machine learning, in the form of the TensorCore. But that hardware actually exacerbates the GPU starvation problem, by increasing FLOPS without increasing bandwidth.
NVIDIA probably isn’t comfortable betting the farm on a single workload – especially one where their main customers are enterprises that can invest in their own machine learning hardware and that is attracting VC money for application-specific hardware. How do you hedge? How can NVIDIA relieve the bottleneck? Unless some workload materializes that is as compute-intensive (per byte of I/O) as machine learning, NVIDIA must seek out ways to address their GPUs’ I/O bottleneck.
I/O: NVIDIA’s Strategic Landscape
The problem confronted by NVIDIA is that they are hindered by some business and legal challenges. According to the terms of their 2011 settlement with Intel, 1) They do not have a license to Intel’s industry-leading cache coherency protocol technology, and 2) they do not have a license to build x86 CPUs, or even x86 emulators.
NVIDIA has done what they can with the hand they were dealt – they built GPUDirect to enable fellow citizens of the bus (typically Infiniband controllers) to access GPU memory without CPU intervention; they built NVLINK, a proprietary cache coherency protocol. They have licensed NVLINK to IBM for the POWER architecture and signaled a willingness to license it to ARM licensees. The problem is that POWER and ARM64 are inferior to Intel’s x86, whose high-end CPU performance is unmatched and whose “uncore” enables fast, cache coherent access across sockets. NVIDIA itself, though an ARM licensee, has announced that they will not be building a server-class ARM chip.
I’m not sure why NVIDIA announced they would not be building their own ARM to drive their GPUs, because that seems like an obvious way for them to own their destiny. It may be that NVIDIA concluded that ARM64 cores simply will never deliver enough performance to drive their GPUs. That’s too bad, because there is a lot of low-hanging fruit in NVIDIA’s driver stack. If they made the software more efficient, it could either run faster on the same hardware or run at the same speed on lesser hardware – like ARM64 cores.
Not being able to coordinate with Intel on the cache coherency protocol has cost NVIDIA big-time in at least one area: peer-to-peer GPU traffic. Intel could, but chooses not to, service peer-to-peer traffic between NVIDIA GPUs at high performance (Intel and NVIDIA give different stories as to the reason, and these conversations happen indirectly because the two companies do not seem to have diplomatic relations). As things stand, if you have a dual-CPU server (such as NVIDIA’s own DGX-1) with cache coherency links between the CPUs, any peer-to-peer GPU traffic must be carefully routed past the CPUs, taking care not to cross the cache coherency link. If Intel could license QPI to Altera, they could license it to NVIDIA. Failing to do so is a matter of choice and a by-product of the two companies’ respective positions in the business and legal landscapes.
As things stand, NVIDIA is dependent on Intel to ship great CPUs with good bus integration, and peer-to-peer-capable GPU servers have to be designed to steer traffic around the QPI link. The announcement that NVIDIA would not build ARM64 SOCs was done in 2014, so now that the competitive landscape has evolved (and though I can remember when Intel’s market capitalization was 12x NVIDIA’s, it is now only about 1.7x), it would not surprise me if NVIDIA revisited that decision.
One Path Forward: SoCs
One partial solution to the interconnect problem is to build a System on a Chip (SoC): put the CPU and GPU on the same die. Intel and AMD have been building x86 SOCs for many years; it is Intel’s solution to the value PC market, and AMD has behaved like their life depended on it since 2006, when they acquired GPU vendor ATI. NVIDIA’s Tegra GPUs are all ARM SoCs. The biggest downside of SoCs is that the ratio of CPU/GPU performance is fixed years before the hardware becomes available, causing workloads to suffer if they are more CPU- or GPU-intensive than the SoC was designed to address. And if the device doesn’t have enough performance, scaling performance across multiple chips may be more difficult because GPUs require such high bandwidth. A conspicuous success story for big SoCs has been in the gaming console market, where the target workload is better-understood and, in any case, game developers will code against whatever hardware is in the console.
So I suspect that as workloads continue to tap out the FLOPS and balance out the bandwidth/FLOPS, big SoCs will start to make more sense. In sizing the CPU/GPU ratio, hardware designers can create a device with the biggest possible GPU that doesn’t starve with the available bandwidth.
SoCs are just a stopgap, though. As the laws of physics continue to lower the boom, the importance of system design will continue to increase, as Kogge pointed out in his 2011 presentation. The fundamental problem of the speed of light isn’t going away… ever.