PBO
Cinebench
7980X PBO
W9-3495X Falcon Northwest OC
For the PBO testing, The Intel system was consuming ~1250w while the AMD 64c system was using ~851 watts.
Non-PBO
HWInfo64
7970X 32C
7980 64C
13 Likes
Seen Jay and Gamers Nexus have benchmark vids out already, been checking back here about ever 10 minutes waiting for Wendel to pass judgement.
1 Like
minor snafu, my fault, about to be live. video be up soon. linux channel video live now.
8 Likes
I skipped over most of parts. Cinebench, Tomb Raider and Photoshop just aren’t my kind of stuff for considering Threadripper 
Wendell has MBW, kcbench and all the Science stuff too.
4 Likes
so much power. i’d love to run large FEA analysis on one of these.
modelling and meshing benefits from single thread performance and the simulation itself can run on a lot of threads.
The price of these is definitely outside of my realm for how rarely i run them.
1 Like
The stock performance is very impressive. My tuned 3495x only got around ~73k CB23 at 450W or so.
Now I wish the 7980X were an 8-channel system so having more than 256 GB could be less painful… (especially since 2DPR doesn’t seem to be a thing with TR7000, though I guess you could do 96 GB×4 = 384 GB).
Definitely looking forward to the Threadripper Pro.
@wendell would love to see updated openembedded/yocto build perf results from threadripper and recent epyc.
The entire 4-channel thing seems to be market segmentation to protect EPYC and TR Pro from people getting TR for their server stuff. EPYC F-SKUs aren’t cheap. I like RDIMMs on TR. Was UDIMMs last time we got TR non-pro.
96GB RDIMMs aren’t cheap. Might as well pay the hike on TR Pro. 3DS DIMMs listed on TR is pure marketing to get that 1TB number.
Behold! My graphs 
improved design
12 Likes
Numbers are looking good!
Now begins the waiting for availability…
All I want is PCIe lanes! 
3 Likes
Im in for the “pages and pages of forum posts” discussing bench results when these monstros get let loose.
Welcome to the Zen4c waiting room. I’m waiting for my 96 lanes of Siena efficiency. But I guess TR boards, being consumer boards more than niche server boards, will arrive earlier.
Wouldn’t be L1T without some large Workstation thread on the latest generation. 2024 will be glorious.
3 Likes
Certainly.
But I am happy that I don’t have to pay for all these 8 channel logic and lanes. I personally am very stoked about the additional PCIe4 lanes.
Mobo + CPU are already expensive enough. 4 memory sticks are much cheaper than 8 (assuming all sticks have the same capacity). And for me the main limiting factor of the desktop class is number of PCIe lanes.
I think my main concern at this point is the power consumption. I have not yet heard anything about idle power consumption. Building a HEDT system with 800W+ peak consumption is already nuts in my book. At the rate power consumption of computers are increasing online retailer will offer addon power panels and circuits with CPU purchase by next year.
1 Like
But instead of e.g. 32GB DIMM x4 you can get 8x16GB. Which is comparable in price (more or less). And you get all the bandwidth. There are a lot of options, even more now with non-binary DIMMs.
LTT showed Bergamo 128 core idling at 75W. And that SKU has all 8xCCDs and all cores enabled. Threadripper is Zen4c too. So might not be that high. Or might be tweaked for more power in general because TR.
Doesn’t cost much by itself. You can get 96 lane Siena for $430 with 8 cores and 70W TDP. If you are ok with running the waste production of Zen4c, it’s rather cheap.
And server boards don’t cost much more than an AsRock Taichi.
The combination of clock speed and lanes is the shit everyone wants and why it’s so expensive
1 Like
@wendell Someone posted on the Linux Threadripper video about benching Large Language Models (i.e Llama2 70B)
For some simple benchmarking i would recommend using llama.cpp (github. xxx /ggerganov/llama.cpp)
It’s currently the best all around CPU/GPU inference engine.
They even have a simple benchmarking tool (llama-bench) that will export in multiple different formats <csv|json|md|sql>.
It will need to be build from source, but on linux its pretty simple.
for CPU tests, use OpenBLAS, for NVIDIA GPU (I use it on a Tesla P40), CuBLAS is used (need nvidia toolkit installed) I havent done much with the Radeon side of things, but hipBLAS works for most on Linux.
Once dependancies are installed, its as simple as make LLAMA_OPENBLAS=1
As for a suitable model, just grab a GGUF converted model from huggingface. the user 'The Bloke" does this and updates them on a regular basis. I’d suggest something like huggingface. co/TheBloke/Llama-2-70B-GGUF and the “quants” are bit reduced files. Q4_0 is 4bits, Q4_K_S is mostly 4bits with some tensors at higher bitrates. Non-k are less computationally intensive.
As for usage - once built from source, you can run a simple benchmark using ./llama-bench -m <filename> -t <num threads> -o <output format>
Inferencing on LLMs is mostly memory bound, thus the interesting test for the new platform.
Hope that helps.
-long time viewer, first time responder.
3 Likes
Based on the benchmarks for the non-pro models, I got my order in for same-day-delivery of the 7975WX, but none of the WRX90 motherboards seem to have materialized. Has anyone heard anything about their availability?
I would love to see a benchmark made with AWS palace (you can find it on github or build it automatically with the spack package manager)
Ansys would be great too but if they are being dumb and not providing licenses thats their loss in free advertising.
The built in examples from palace might not be heavy enough to load the huge core count so let me know if you have time to give it a try and I could look at expanding that. It also lets you import meshes from COMSOL so it might work well to build large problems.
It’s concerning microcenter has stock of chips but no motherboards
3 Likes
Same with EPYC Siena. Chips but no boards to be seen in retail. Zen4c board shortage.
Neither 7-series threadripper nor motherboards from the reputable retailers in Germany yet.