British chip designer Graphcore recently unveiled the Colossus MK2, also known as the GC200 IPU (Intelligence Processing Unit), which it calls the world’s most complex chip for AI applications. The chip offers eight times the performance of its predecessor, the Colossus MK1, and is powered by 59.4 billion transistors, which exceeds 54 billion transistors in NVIDIA‘s (NASDAQ: NVDA) Latest A100 top-level data center GPU.
Graphcore plans to install four IPUs GC200 on a new machine called the M2000, which is about the size of a pizza box and offers a computer-powered petaflop. On its own, the system is slower than NVIDIA’s A100, which can handle five petaflops on its own.
But the Graphcore M2000 is a plug-and-play system that allows users to link up to 64,000 IPUs for 16 exaflops (each exaflop equals 1,000 petaflops) of processing power. To put this in perspective, a human would need to perform a single calculation every second for almost 31.7 billion years to match what an exaflop system can do in a single second.
The GC200 and A100 are clearly very powerful machines, but Graphcore enjoys three distinct advantages over NVIDIA in the growing AI market.
Image source: Getty Images.
1. Graphcore is developing custom chips for AI tasks
Unlike NVIDIA, which expanded its GPUs beyond gaming and professional display purposes in the AI marketplace, Graphcore designs custom IPUs, which differ from GPUs or CPUs, for machine learning tasks.
Image source: Getty Images.
On its website, Graphcore states: “The CPUs were designed for office applications, GPUs for graphics, and IPUs for machine intelligence.” He explains that CPUs are designed for “scalar” processing, which processes one data at a time, and GPUs are designed for “vector” processing, which processes a wide variety of integers and floating point numbers at once.
Graphcore’s IPU technology uses “graph” processing, which processes all assigned data through a single graph at a time. It claims that the IPU structure processes machine learning tasks more efficiently than CPUs and GPUs. Many machine learning frameworks, including TensorFlow, MXNet, and Caffe, already support graphics processing.
Graphcore claims that the vector processing model used by GPUs is “much more restrictive” than the graphics model, which may allow researchers to “explore new models or re-explore areas” in AI research.
2. The Graphcore GC200 offers cheaper petaflop processing power
NVIDIA’s A100 costs $ 199,000, which equals $ 39,800 per petaflop. Graphcore’s M2000 system offers a petaflop of processing power for $ 32,450. That difference of $ 7,350 per petaflop could generate millions of dollars in savings on multi-exaflop systems for data centers.
That could spell trouble for NVIDIA’s data center business, which increased revenue 80% annually to $ 1.14 billion last quarter and accounted for 37% of the chipmaker’s top line. NVIDIA recently acquired data center networking equipment maker Mellanox to strengthen that business, but that larger scale may not deter Graphcore’s disruptive efforts.
3. Graphcore is backed by venture capital
Unlike NVIDIA, a publicly traded chipmaker that regularly reviews its spending practices, Graphcore is a privately held company that can focus on research and development (R&D) and growth rather than its short-term gains.
Graphcore was founded just four years ago, but was already worth $ 1.95 billion after its last round of financing in February. Its sponsors include investment companies like Merian Chrysalis and Amadeus Capital Partners, as well as large companies like Microsoft (NASDAQ: MSFT). Microsoft already uses Graphcore IPUs to process machine learning workloads on its Azure cloud computing platform, and other cloud giants could follow that lead in the coming years.
Should NVIDIA investors be concerned?
NVIDIA enjoyed the advantage of being one of the first to move in data center GPUs, but faces a growing list of challenges, including source chips Amazon, Facebookand From the alphabet Google Graphcore represents another looming threat, and NVIDIA investors must be wary of its new chips, which appear to offer a cheaper, more agile, and more flexible approach to tackling machine learning and artificial intelligence tasks.
