When it comes to application acceleration, having a deep understanding of the complex algorithms that underly most HPC and AI applications is perhaps the most important thing that can be brought to bear. And the relative handful of experts in coding field programmable gate arrays – and the hybrid compute and networking complexes that surround them these days – that have unique industry and algorithm knowledge are a valuable commodity.
Witness the $16.7 billion that Intel paid to acquire the Altera FPGA business in 2015 and the $49 billion that AMD just shelled out to take over Xilinx. Those deals were not just about buying FPGA revenue streams and profit pools, but in gaining that expertise in dataflow processing and in the mathematical algorithms that underpin industry-specific applications in both HPC and AI. That is the real reason that memory maker Micron Technology bought Convey Computer, a creator of hybrid supercomputers based on X86 CPUs and FPGA accelerators, back in April 2015, and it is also why AI chip startup Groq, one of the bigger players among a crowded field of datacenter AI accelerator makers, has just acquired Maxeler Technologies for an undisclosed sum.
“Maxeler knows the dataflow side of HPC and machine learning really better than just about anyone in the industry,” Adam Tachner, vice president of corporate development at Groq, tells The Next Platform. “They have seen it in oil and gas discovery, financial risk analysis, climate modeling, quantum computer simulation, and so many fascinating areas where their level of insight and understanding of the customer problem was really what they were selling as a consulting service.”
We have always admired the very interesting clustered FPGA systems that Maxeler put together and the various supercomputer products it has been involved with since it was founded in 2003 by Oskar Mencer –notably the EuroEXA project to create a petascale prototype of an exascale system based on Arm CPUs, FPGAs, and other ASICs. Maxeler’s key insight was to customize the FPGA to meet the needs of the algorithms and their dataflow, rather than try to bend the algorithms to so they would spread across racks of CPUs and now GPUs.
“There’s an old saying: The key to happiness is low expectations, and no one has lower expectations than a developer is accustomed to working with FPGAs,” Tachner quips. “And the Maxeler team came to us, and they saw the Groq flow and the compiler and other tools, and the ability to just shove these massive amounts of compute right through the chip with low latency. And everything that they have been breaking their heads against the wall to try to solve – and now it’s just there for them. And the Groq platform is scalable, and it is super low latency – and they can just solve the problem.”
And thus, the Maxeler team, which is based in London, can be a catalyst to bring more new workloads to the Groq platform, given the vast connections that Mencer has in academic, government, and enterprise computing where acceleration has been necessary.
Don’t jump to the wrong conclusions about the acquisition, though. Tachner was very clear that the acquisition of Maxeler will not change Groq’s ASIC and system roadmap. So don’t expect hybrid systems using the GroqChip and, say, AMD Xilinx FPGAs. That is probably not going to happen – and if it does, FPGA will probably be some sort of network or offload engine like a DPU, if there is ever a time when a DPU is necessary in the GroqNode systems, which cram eight GroqChips into a single chassis and can scale to 64 chips all interconnected in a GroqRack rated at 12 petaflops at FP16 half precision.
There are around two dozen people from Maxeler who will be joining the London Groq office, and everyone, including Mencer, is staying on. With the addition of the Maxeler team, Groq is now at around 225 employees and is climbing that hockey stick curve. In April last year, when it completed its Series C round to bring it to a total of $367 million in venture funding, Groq broker through a $1 billion valuation to reach “unicorn” status, for whatever that is worth.
We will be catching up with GroqDay ’22 event soon to talk architecture and markets and get the lowdown on what the roadmap looks like and what customers are up to.