I will probably end up frying these guys.
But from a PCB/Chip temp perspective, I am 100% confident that there’s enough static pressure with two of these quiet fans. I’ve been running a decent QTY of Sophon SC-1 cards since February of this year, and have had no problems running the chips up to 80C while pushing a constant data stream through the API.
From the tear downs posted on other threads, we know these miners are indeed powered by the exact same BM1680 as the SC1s. So Im just working from that frame of reference.
An SC1 running a full and sustained workload will pull about 85W on the 12V rail of our servers.
A B3 is 12 BM1680 chips. We know they’re not clocking in at 1020W, so the TPU ASICs are likely not running full clip.
They’re likely low corner ASICs with too many dead cores to run in a Sophon product.
There’s also no external DDR in this configuration. So there’s a few watts shaved off there.
Given that bitmain usually has a habit of running about 80-135W per PCIe connector, it sure looks like they were expecting it push about 600-750W. The heat sink is a beefy structure. They also used thermal pads between the hashboards (making contact with both). This is a first for Bitmain if I’m not mistaken.
All signs point to a device that either didn’t reach its potential after development, or one that has a LOT of headroom to grow as a more functions and op codes are added to the BYTOM (or other Tensority) blockchains.
Hi,
i was several times looking at Bitmains SC-1 cards, could you explain a bit about what you use them for? Are they ok to mine as well? Price?
I'am really interested to learn more about these, so every information you can provide would be appreciated.
We have a few products for the medical and security industries that use Cafe and Tensor pre-trained CNNs. (Convolutional Neural Networks)
The SOPHON Cards outperform NVIDIA GPUs by an exponential factor on certain workloads.
Think OpenCV using GoogleNet, AlexNet, GenderNet, AgeNet, in an automated threat detection application. Or using pixel differentiation to locate tumors and cancerous growths in X-Rays, MRIs, and CTs. They also have a few unique blocks that enable other types of hardware acceleration.
Sophon cards really reduce the amount of guided training necessary and subsequently lowers the time to completion for our pipeline.
Regarding mining, we were able to produce a solver based on Tensority, but never fully developed that into a miner.
Really never seemed like there was a compelling reason to do so.