In the last 2-3 days we have gotten over a dozen folks writing in about the Cheyenne Super Computer listing on GSA Auctions. What is inside that listing is really neat since it gives us some insight into the SGI system and the value of the components therein. At the time of writing this, the auction has just crested $51,000, which is not bad for thousands of CPUs and tens of thousands of DDR4 DIMMs. On the other hand, it would require a lot of work to extract.
Leaky HPE SGI Cheyenne Supercomputer For Auction
Of course, we need some stats on Cheyenne. It is an Intel Xeon E5-2697 V4 system from the 14nm “Broadwell” era that has been operating since 2017. That was an era just as the Intel Xeon “Skylake” and first-gen AMD EPYC “Naples” were hitting the market. This was being installed as HPE acquired SGI for $275M, a deal that closed in November 2016. With a 2017 go-live date it was a HPE system at that point.
Quick specs on the systems were that there were 4032 dual CPU nodes for 8064 CPUs total. Each processor has 18 cores for 145,152 cores total. If you were thinking of valuing those CPUs, we just checked ebay and they are about $45-50 each today. If you could sell all of the CPUs at $45, that would be over $360K of CPUs, excluding transaction costs.
We do not have details on the DIMMs, but the GSA auction has a few tidbits. There is a total of 313,344GB of memory. Assuming each CPU had 4-channels of DDR4 populated for memory bandwidth purposes, that gives us an average of about 9.7GB of memory per DDR4 DIMM channel. The auction says there are a few high-memory nodes as well. We also know that there are about 32,256 DDR4-2400 memory modules, although that could be a bit different depending on how the high memory nodes are configured. At $10 or so each these days, the 8GB memory modules are not expensive, and that would be another $320K or so of memory, if not a bit more, accounting for the extra memory nodes.
Of course, there are challenges. About 1% of the nodes have failed recently with ECC memory errors being the chief culprit. In addition to that “the system is currently experiencing maintenance limitations due to faulty quick disconnects causing water spray.” (Source: GSA Austions) This is the number 1 concern of liquid cooling in the data center, and it sounds like Cheyenne is leaky.
The nodes themselves are probably not able to be re-purposed, but the CPUs and memory can be. This has become a trend beyond just supercomputers, but also with hyper-scale data centers. Many older-generation CPUs end up becoming much less expensive than the older servers they go into these days.
We did spot some 1U Supermicro systems with 3.5″ drive trays in the auction, however, so those should be standard systems. SGI often used Supermicro hardware for standard nodes even as we saw in the firstĀ HPE Spaceborne Computer.
Aside from the compute, it seems like some of the networking cables along with a number of the switches, CDUs, and more will be included in the sale. The EDR Infiniband gear may still have some value as well.
Final Words
Of course, the idle power consumption of everything was 747.41kW, or around six and a quarter NVIDIA GB200 NVL72 racks. Max power consumption was rated for around 1.75MW. We do not see many sites with >1MW of power available looking to house old supercomputers instead of new AI servers these days. That leads us to parting out the system.
Extracting the value from this system is likely a challenge. First, one will need a moving crew with proper equipment to load the system into likely more than a single truck based on size and weight. The entire system weighs around 95,000lbs or around 43,000kg.
Even after that, one would likely still have to strip the components and then find buyers for them. The CPUs and memory alone would fetch close to $700K based on ebay pricing. There is some value in the networking side, and perhaps the CDUs. One also has to remember that there are likely transaction fees involved. Further, dumping thousands of CPUs or tens of thousands of DIMMs onto marketplaces like eBay will likely deflate pricing.
Still, it is always fun to imagine what could happen when buying and parting out a previous-generation Top500 system like this. Given the value of the parts, our sense is that it will close well above $51K. Still, it might be a more fun way to spend $51K than buying a NVIDIA H100 GPU and a system to house it in.
If you want to dream, head over to GSA Auctions.
As of this post, price is at $480k. So yeah, it went up a tiny bit