Trying Out a $20K Machine Learning Workstation, The Lenovo ThinkStation P920 

How do you get around the new tech that is Nvidia RTX cards? I am running into so many problems because the new cards can't run older projects using cuda 10 and tensorflow 1.x. The new RTX cards seem to only run Cuda 11.1 and tensorflow 2.x

Nice GPU, I will have to make the hop to Ampere at some point.

@@HeatonResearch The Nvidia RTX A6000 for 45 hundred us dollars is an approprate card for a professor: ru-vid.com/video/%D0%B2%D0%B8%D0%B4%D0%B5%D0%BE-kw78MUnOqIs.html

oh yes, for sure.

@@HeatonResearch SUBSCRIBED!

@@shadowshadow2724 Thanks! and welcome to the channel.

That is actually next week's video.

Hah, indeed. That last 512GB of RAM is REALLY expensive, as it means you need a higher grade Xeon CPU.

@@HeatonResearch That is true. I also noticed that on the configuration page,

@@HeatonResearch forget xeon. Threadripper is where all the cool kids are

Perhaps you found what you needed already. Most High End workstations from Dell, Lenovo and HP would work. They have multiple PCIe x 16 slots and tons of RAM sockets. Most of them that came out in the past 10 years or so supports dual CPUs as well. Look into Dell 7910/20/60, HP Z840, HP Z8/Z6 G4

more like 2030

I do not currently have the system, I returned it many months back. The system did have standard 8-pin connectors TO the graphics card. The other end of the cable, that connected into the power supply, MIGHT be more proprietary. I do not recall if there were extra connections, I know it was apparently possible to put a 3rd GPU on that system somehow. Mine was only equipped with 2. The entire power supply system in that machine is very modular, things snap in and out quite easily. BUT, that also means it is very Lenovo specific/proprietary.

Define what “better” means to you.

@@monkyspnk777 better training yield

@@seanhdr4k629 well that would more depend upon the model or with nvidia cards you need to google how many cuda cores they have. More cores doesn’t imply a better model, more cores allow you to do more training in the same amount of time.

@@monkyspnk777 There are many factors to be considered when measuring GPU performance for deep learning training. For example, memory bandwidth, Not only CUDA cores, memory. I already knew about things you mentioned. I was asking for general opinion or remarks in real life. I wasn't being serious.

I have Ryzen, Core, Xeon, and Epyc systems. You don't understand the limitations of Ryzen. The point of running Epyc or Xeon is for compute scalability, memory scalability, and expansion capability. Ryzen has very few PCIe lanes, and even with HD-DIMMs, it caps at 192GB RAM. Whereas, Epycs and Xeons support terabytes of RAM, and have close to or over 100 PCIe lanes available. Xeon systems make great VM servers as the DCPMM memory can now be had for reasonably 29 cents per gigabyte. There are also vendor-specific CPUs such as the 8259CL that require a voltage regulator mod but are quite capable in 2S at $260 for two. Your comment was untrue 3 years ago and it's untrue now.

The machine was decently loud, running full power, I could not really use anything but a lav mic. It also really made my office hot if I had the door closed, and it was the dead of winter!