It's fast approaching a decade since I wrote an article for Scientific Computing World, back in December 2016! Where does the time go? Did we enter a new age?
|
At the top of the article I mention IBM, sadly, IBM are no longer prevalent in the HPC CPU space (however, they remain active in the HPC storage space), after their withdrawal from the marketplace, having discontinued their relationship with NVIDIA that stopped with NVIDIA V100. That’s a shame, as competition breeds better products. However, in the meantime, we have had the unparalleled technology march from NVIDIA, through their HGX platform, which goes someways to negating the USPs that IBM had with their Minsky platform. Essentially, the HGX platform allows multiple NVIDIA GPUs to be connected via NVSwitches, to create multi-GPU systems, connected via NVLink and NVSwitches providing four-way and eight-way systems. |
|
Over this last decade, we have seen not just IBM, but also Intel fall away and not just in their accelerator platform, Knights Landing(KNL) but also, losing ground in their processor business for HPC. For those of you with long memories, this is a reversal of the Opteron -> Nehalem scenario that we saw back in 2008. Intel cancelling the accelerators platform in 2018 KNL and is successor Knights Hill led to the loss of a number of US Government contracts, but really wasn’t that unexpected given their reticence in the accelerator market space, anyone remember Larrabee? In recent press reports, Intel CEO reportedly admits “it is too late for us to catch AI leaders like NVIDIA”[1] More concerning to Intel, is the processor business, and gradual emergence of AMD as the processor of choice for HPC.
AMD have gone from strength to strength, from the first Zen 1 – Naples cores back in 2017 to the present Zen 5 – Turin architecture, AMD have consistently leap-frogged Intel and have been at the top of the tree for some time now, in terms of processors for HPC.
The current Top 10 from Top 500 list as of June 2025 currently has four AMD systems (including #1 and #2), three Intel, three ARM based system. So, Intel are still managing to hold their own. Compare this to June 2016, when there were five Intel, one Chinese processor, two BlueGenes, one AMD, and one ARM. This comparison, is roughly inline with the emergence of AMD and the decline of IBM, with the blip of the Chinese processor in the Sunway TaihuLight (#1) system built using ShenWei SW26010 256 core processors, where we see the emergence of ARM. However, for the UK and for smaller HPC systems, AMD is far more prevalent in the HPC space than Intel, partly because of the greater number of cores in the AMD processors, which leads to a better price performance and also reduction in costs for infrastructure.
What about Field Programmable Gate Arrays (FPGAs)? Since 2016, both of the main FPGA players were both bought up by the two main chipmakers, Intel and AMD. The market is worth to be an estimated $11.14 billion in 2025[2]. Undoubtedly, their flexibility remains, however, we haven’t really seen the uptake in HPC, rather, I would imagine, the main markets, are within the AI, embedded and autonomous driving and the like. The same drawbacks remain, in terms of programming, these other applications that are the primary purpose of the embedded system, are ideal for FPGS, where they can be programmed to provide specific applications and functionality, rather than being a general-purpose system as per a GPU.
NVIDIA has gone from strength to strength, becoming the #1 company in the world by market capitalization, eclipsing both stalwarts Microsoft and Apple due to the emergence of AI since 2016. The worldwide frenzy that is AI has driven NVIDIA’s stock price higher and higher – is there a limit? As the world settles down to the AI revolution, NVIDIA has positioned itself extremely well in terms of both HPC and AI, since it ultimately shares a common infrastructure platform, whilst not exact, very similar requirements.
Taking a little step back, we have seen the emergence of NVIDIA's own HGX platform, superchips, Grace -Hopper, Blackwell, Ada, etc. and the acquisition of Mellanox by NVIDIA, has made NVIDIA a dominant force in the HPC and indeed the datacentre space. However, AMD actually holds positions 1 & 2 in the Top 500 slots, in terms of GPU and processor, in the form of “El Capitan” and “Frontier”. Indeed, out of the top 10 systems, AMD GPUs power 4 out of 10 systems. NVIDIA has 5 (with the 10th system being the Fugaku non-GPU based system).
Does this show hope for AMD, and will it herald a new dawn of Aquarius?
ARM has also gone from strength to strength over the past decade, not just in mobile phones. NVIDIA has embraced the ARM processor (having nearly bought the company back in 2020, which however failed due to regulatory approval, or rather non-regulatory approval). NVIDIA continued development with ARM processors, creating the Grace family of superchips in 2022. Consisting of a Grace-Grace (ARM processor – ARM processor) and a Grace-Hopper (ARM processor – Hopper GPU) tightly integrated superchips with high speed NVLink between them. They have continued to develop the ARM processor symbiotic GPU system, culminating in a GB200 NVL4, essentially two Grace CPUs connected directly to four Blackwell GPUs on a single board, Blackwell being the latest generation of NVIDIA GPU.
It is not just NVIDIA that has taken to ARM, both Google and AWS have designed their own, ARM variants, Axion and Graviton respectively. These aren’t the only companies to do so, given the licensing arrangements of ARM and the extreme power efficiency of the processor. ARM features more and more in the datacentre, It's hard to actually gauge how many workloads are running on ARM in Cloud Providers, but it must be significant. In the early days, it was lack of applications and operating systems, however, these issues seem to have gone away, as we now see more mainstream operating systems supporting ARM, and even Microsoft Windows runs on ARM. ARM is making significant in roads, not just in HPC, but in the wider computing field.
What is driving the uptake in ARM? Well, one contributing factor is surely the increase in power requirements, not just for CPUs but also GPUs. Back in 2016, the current Intel CPU was Skylake, the maximum TDP was 205W with up to 28 cores. Contrast this with today's Intel Granite Rapids processor, with up to 500W and as many as 128 cores. The NVIDIA datacentre GPU of 2016 was the P100 which is a 300W part, today the B200 is the current datacentre GPU rated at 1000W. Datacentre power is another story for another day.
Competition is always good for driving innovation and reducing costs. As I said back in 2016, NVIDIA has a head start in terms of CUDA adoption and continues to lead. However, as the world leans towards AI, it's the frameworks that have gained more traction, the PyTorches, Tensorflows of this world. AMD make great performant GPUs, which is seen in the performance specifications, and indeed is mirrored in the Top500 list. In the UK, will users now realise that the top two supercomputers in the world, can’t be mistakes?
IBM have just recently released Power 11, with the introduction of the Spyre AI chip. Is this a new lease of life for IBM within the AI space? Spyre has been designed very differently to a traditional GPU - it has 128GB of memory, but with a fraction of the computing power and memory bandwidth found in other GPUs. Critically, only consuming 75W of power! As many as eight Spyre cards can be linked together, providing 1000GB of memory in 600W. IBM have tested a system with 96 cards.
Not all is lost with Intel! Granite Rapids has recently been announced and has started to show signs of marking a beginning of a recovery for Intel, with up to 128 cores. It is becoming more attractive and with Multiplexer Combined Ranks (MCR) memory DIMMs. In my opinion, Intel Granite Rapids has the potential to become the Go-To product for memory bandwidth bound applications.
Only time will tell whether we have reached the new age of Aquarius. I look forward to revisiting this once again in a couple of years time.
----------------------------------------------------------
[1] https://www.techradar.com/computing/cpu/intel-ceo-reportedly-admits-it-is-too-late-for-us-to-catch-ai-leaders-like-nvidia-but-heres-how-it-could-still-recover
[2] https://www.logic-fruit.com/blog/trends-and-reports/future-of-fpga/
