Back in the day, when the AI revolution was young, Nvidia co-founder and chief executive officer used to make the rounds with each new generation of datacenter GPUs and provide the first DGX systems based on any particular GPU to marquee customers. Sam Altman and Elon Musk, co-founders of OpenAI, famously were donated the first DGX-1 system to come off the production line in August 2016.
Now, it takes a cluster with thousands of GPU accelerators to tackle any significant AI or HPC work and the marquee customers are much larger – and they are not always traditional HPC centers, who used to get Nvidia’s technology first because they were the pioneers in the early days of accelerated computing between 2008 and 2014.
In the “Blackwell” generation, a much more established customer that is central to the consumer and commercial IT sector and that absolutely not coincidentally is based on Taiwan, is the marquee customer: Hon Hai Technology Group, colloquially known as Foxconn.
Foxconn, which is based in Taipei and which operates factories in Taiwan and in mainland China, has been the largest electronics manufacturer in the world for the past decade and a half. Foxconn’s factories in Zhengzhou in Henan province, which have a quarter million employees, are where most of Apple’s iPhones are manufactured and are also where Foxconn has facilities to make electric vehicles.
Foxconn generated NT$6,162 billion in revenues in 2023 (about $192 billion at prevailing exchange rates at the end of last year) and NT$142.1 billion (about $4.4 billion), which works out to a paltry 2.3 percent of revenues. Contract manufacturing is a cut-throat business and like being an OEM, you really need to want to do it to stay in the game because you are going to work hard and there is not going to be much in the way of profits.
Which is why Foxconn – and not the National Center for High-Performance Computing (NCHC) in Hsinchu City – is getting a cluster of Blackwell B200 NVL72 systems.
The NCHC, as the center is known, was founded in 1993 and made it to the bigtime in November 2018 with its “Taiwania 2” system, an NT$430 million ($13 million) hybrid CPU-GPU machine that entered the Top 500 supercomputer rankings at number 20. Taiwania 2, which was built by Taiwanese contract manufacturer Quanta, was based on clones of the “Volta” generation of DGX systems, and had two 18-core Intel Xeon Gold processors and eight V100 GPU accelerators in each of its 252 nodes. Those 2,106 GPUs delivered 15.3 petaflops of peak FP64 oomph. The follow-on Taiwania 3 machine was an all-CPU cluster, also built by Quanta, that delivered 4.35 petaflops peak for 64-bit floating point math. The original 2 petaflops Taiwania 1 machine from 2017 was built by Fujitsu and is slated to be replaced with Taiwania 4, but as far as we know that has not happened as yet.
You can sure bet that the NCHC would love to have Taiwania 4 look like the system that Foxconn has scored as one of the first customers for the Blackwell GB200 NVL72 rackscale nodes from Nvidia.
To put it bluntly, Foxconn has factories and an EV business as well, and it needs bigtime AI power to create virtual factories in the metaverse and it needs big iron to create ADAS systems for self-driving cars. And that, along with being in Taiwan, makes it an ideal customer for Nvidia.
The neat bit about this is that Ingrasys, the datacenter equipment manufacturing arm of Foxconn, can build its own Blackwell systems and use the learnings to grow its OEM and ODM business selling AI clusters to other enterprises. This is something that a national HPC center cannot do, and that is another reason Nvidia is touting the deal and putting Foxconn at the front of the Blackwell line.
Foxconn and Nvidia actually worked out the deal for the supercomputer back at the Ingrasys booth at the Computex trade show in Taipei back in June, according to the Foxconn announcement, and the feature image above is from that meeting. The announcement was made at Hon Hai Tech Day 2024 this week.
The supercomputer that Foxconn will build will go into the Foxconn computing center in Kaohsiung, and it will consist of 2,304 of the GB200 superchips, which put one “Grace” CG100 Arm server processor and a pair of “Blackwell” B100 GPU accelerators, in a single node. That yields a total of 4,608 Blackwell GPUs, which will be configured as GB200 NVL72 racks with 72 GPUs per rack for a total of 64 racks. At FP4 precision on the tensor cores with sparsity support on, the peak theoretical performance of this AI system is 92.16 exaflops; maximum FP64 performance on this machine will be 207.4 petaflops.
The system will also have 165,888 “Demeter” V2 Arm cores across those Grace chips, and they have a certain amount of vector oomph to add to the mix as well as acting like a scratch memory tier for the Blackwell GPUs across NVLink ports on the Grace chip.
The system, which has not been nicknamed as yet, will be the most powerful supercomputer in Taiwan, according to Nvidia. The machine is being built in phases for the Hon Hai Kaohsiung Super Computing Center. The first phase of the machine will be operational by 2025, with full deployment completed sometime in 2026.
Which tells us that everyone wants Blackwell GPUs and even Foxconn cannot spare the full complement of 4,608 Blackwells for itself.
The Foxconn AI supercomputer will be used to do cancer research as well as large language model development and other AI tasks relating to smart cities and smart manufacturing. While that is all well and good, NCHC is getting by with one-tenth the FP64 performance that Foxconn will be able to bring to bear, and has orders of magnitude less AI computing power, which will hamper its research until the Taiwanese government shells out funds to upgrade its Taiwania family of machines.
Neither Nvidia nor Foxconn talked about what this unnamed AI supercomputer in Kaohsiung might cost, but at an estimated $3.4 million per rack at street price (that’s our estimate), that comes in at $220 million to buy 64 racks of the GB200 NVL72. The Taiwanese government would probably have a hard time coming up with that money.
Using an Rpeak to GDP ratio of 0.2, which is the average of what we calculated for the United States and for China for exascale-class machinery and which we talked about when arguing that the United Kingdom deserves its own exascale-class machine back in August, then at a GDP of $792 billion (in US dollars) expected for 2024, then Taiwan should have at least 160 petaflops of FP64 oomph, which is eight times what NCHC has. The question is: Does Foxconn’s future AI supercomputer count, or not? We don’t think so, any more than the vast fleets of GPUs at the hyperscalers and cloud builders in the US and China count when we are considering investment in HPC by public research institutions that are trying to advance science and health and the art of computing.
Foxconn is sure getting itself a nice present for its 50th birthday! ( https://www.foxconn.com/en-us/press-center/events/foxconn-events/1284 )
And if it’s anything like last time (2018), when NCHC had the 2 PF Taiwana 1, and Foxconn leapfrogged it with a 6 PF machine in Kaohsiung, taking the Taiwanese HPC lead, only to be hopscotched by NCHC with the 15 PF Taiwana 2 ( https://www.top500.org/news/foxconn-builds-taiwans-largest-supercomputer/ ), then we can expect that Foxconn’s brand new 207 PF machine, will soon be proportionately pole-vaulted by a 500 PF machine at NCHC!
(yes, 160 PF is nice wrt GDP, but proportions are everything! eh-eh-eh!)