Recent content by Patriot

need large bar support >4g decoding on. We need pictures...and details as to what external hackjob you are using.

I think picking Intel after epyc's 4004 launch should call into question calling this a power efficient server. Imo, get a 4004 board, and the base chip or 5600x. https://www.phoronix.com/review/amd-epyc-4124p/8

You can physically change the setting, but your ram wont make it that far stably. You will most likely fail to post, have missing dimms and be throwing a metric F-ton of ecc errors. Some dimms wont handle 1 notch, I've not seen any handle 2.

Not beyond cpu support, but if you have lower than 3200 you can manually set different frequencies. I have overdriven 2400 to 2666 and 2666 to 2933. You can also flash ram's spd chips across the spi bus if you have tested the ram to be stable at lower cas or higher freq. I am not aware of a...

Pretty straightforward naming scheme. M Maxwell, K Kepler, P pascal, V Volta, T Turing, ... they get a little screwy with Ampere and Ada Lovelace, Hopper, and Blackwell in 6mo. Volta is first gen matrix, Turing 2nd (rtx2000 gen), Ampere desktop and server resync'd A100 and RTX3000, split Ada...

Good price.

Efficiency is all about power/watt, Many people have waaay more compute than they need, myself included. My expertise is in accelerated workflow scaling. My homelab is pretty heavy iron. My nas, is weak and old because it doesn't need more... it is slated for replacement but is handling its...

I will have to check with mi100 and cx5, what is the kernel and amdgpu driver version you are using?

No, I don't see anything in there that relates to TDP or power limits. but if I did, I might start there.

I wonder if it would do better if you used a redriver board from an X16 slot, or shorter cables.

link to what you purchased? I don't know if they support binding those ports together. You might have more luck if you convert an X16 to SlimSAS via adapter rather than use those ports individually.

I wish I had noticed this thread, but my december was quite busy. W7900 should in theory outperform the mi100 in LLMs depending on memory bandwidth dependancy. Before AMD started publicing numbers I took their blogs flops/clock numbers and did some simple math. I also took what public numbers...

https://enterprise-support.nvidia.com/s/article/mellanox-adapters---comparison-table

you have to contact them directly to get blowers... they wont sell retail

Please don't, get 4090 blowers or RTX6000 Adas... don't chop it.