Nvidia fell off a cliff last October from a high of $290 to a low of $130. Meanwhile, the challenger Xilinx remained unharmed by the tech rout, and despite unfavorable macro conditions. Nvidia popularized GPUs in 1999 and Xilinx invented FPGAs in 1985, and both are chips that will define the computationally-intensive future.
GPUs originated from the advanced computations required in gaming and FPGAs originated from electronics engineering. There are strengths and weaknesses to both, however, these are the two that will power the artificial intelligence and machine learning-driven economy. The size of this AI and ML economy is expected to reach $15 trillion by 2030 up from $2 trillion this year.
Keep in mind, that long before technologies go public, they are incubating across the startup ecosystem. By the time AI and ML companies reach the public markets, the technology powering and developing this wave of companies was already decided in the years prior. We are in those critical years where startups must quickly design and develop AI if they want to have the first-mover advantage. This is creating a battle between FPGAs and GPUs.
Below, I break down the differences between Xilinx’s FPGAs and Nvidia’s GPUs before analyzing the financials and theories on how the two will perform in the future.
Note: Previously, I discussed how Nvidia stock has two impenetrable moats: the developer ecosystem and GPU-powered cloud. This previous analysis was written during the height of the panic sell-off, which I negated as being overly-pessimistic due to Nvidia’s strong fundamentals.Nvidia stock has two impenetrable moats: the developer ecosystem and GPU-powered cloud. This previous analysis was written during the height of the panic sell-off, which I negated as being overly-pessimistic due to Nvidia’s strong fundamentals.
AI and Machine Learning
On many technical levels, FPGAs (Xilinx) are considered superior to GPUs (Nvidia). They offer a higher amount of on-chip cache memory to help reduce the bottlenecks from external memory, and are flexible enough to be reconfigured for various data types, such as binary, ternary, and custom data types, whereas GPUs must be modified at the vendor level.
FPGAs are also known for power efficiency, and often test at 10x better in power consumption than GPUs and also 4x better than GPUs for general purpose compute[1]. Reconfigurability for FPGAs also helps provide this efficiency beyond deep learning for a large number of end applications and workloads. The architecture of FPGAs is very adaptable as the chips allow a user to address all of the needs of a workload with the resources provided by FPGAs, such as reconfiguring the data path during run time and with partial reconfiguration. Meanwhile, GPUs are restricted as the architecture is a Single Instruction Multiple Thread (SIMT), which provides an advantage over CPUs but can result in lower performance efficiency when enough parallels cannot be found while mapping the workload.
As pointed out in my previous analysis on Nvidia, software developers prefer GPUs as their frameworks are easier to develop on. Nvidia’s CUDA architecture, for instance, does not require an in-depth understanding of underlying hardware. FPGAs require knowledge of machine learning algorithms at the hardware level, in addition to the software development, and this has been a barrier to entry for FPGAs. FPGAs are a reconfigurable integrated circuit (hence the strengths on being easily reconfigured), which requires specifying a hardware circuit, whereas GPUs are configured via software[2].
“Nvidia, thanks to the CUDA software stack (which AMD cannot match), has a much more unassailable position than does Intel with Xeon CPUs (where an X86 application just runs on either a Xeon or an Epyc).”
– software developer on Reddit
Section takeaway: FPGAs result in faster and more efficient compute but are harder to program due to hardware circuit configurations when compared to GPUs for machine learning, which are more universal and require less engineering resources.
Financials
Nvidia and Xilinx power more than data centers, of course. Nvidia’s top revenue segment is gaming, the origin of GPUs, and this drives about $1 billion per quarter in revenue. Xilinx’s top segment is Communications with many investors using Xilinx as a global bet on 5G with communications revenue increasing 41% year-over-year as reported in the most recent quarter. Xilinx also was not as affected by crypto as the Broadcast, Consumer & Automotive category was 17% of revenue compared to 15% of revenue in the same quarter YoY. (Xilinx classifies crypto as consumer in this 10-K).
Xilinx has a direct competitor with Intel, who acquired Alterra for $16.7 billion. Intel is keen to solve the development uptake issues with FPGAs with the release of Stratix 10 hardware, which has a software layer to simplify development. Microsoft Azure is partnered with both Xilinx and Intel/Alterra on FPGAs although there is some indication that MS is leaning more towards Xilinx in the near future after announcing they will replace Intel chips with Xilinx in over half of their servers.
