Item 1 of 1
forum threadDr.W posted Jul 10, 2026 2:04 PM
Micro Center
Deal Details
Available In-Store Only; In stock at all stores when posting.
SPECS:
PNY NVIDIA GeForce RTX Pro 4500 Blackwell Single-Fan AI & Workstation Graphics Card; 32GB GDDR7 Memory; PCIe 5.0 x16 - Micro Center
https://www.microcenter
SPECS:
- 32GB GDDR7 256-bit Memory
- 7680 x 4320 Maximum Resolution
- PCIe 5.0
- Full Height, Dual Slot
- DisplayPort 2.1b
PNY NVIDIA GeForce RTX Pro 4500 Blackwell Single-Fan AI & Workstation Graphics Card; 32GB GDDR7 Memory; PCIe 5.0 x16 - Micro Center
https://www.microcenter
Community Notes
About the Poster
Deal Details
Community Notes
About the Poster
Available In-Store Only; In stock at all stores when posting.
SPECS:
PNY NVIDIA GeForce RTX Pro 4500 Blackwell Single-Fan AI & Workstation Graphics Card; 32GB GDDR7 Memory; PCIe 5.0 x16 - Micro Center
https://www.microcenter
SPECS:
- 32GB GDDR7 256-bit Memory
- 7680 x 4320 Maximum Resolution
- PCIe 5.0
- Full Height, Dual Slot
- DisplayPort 2.1b
PNY NVIDIA GeForce RTX Pro 4500 Blackwell Single-Fan AI & Workstation Graphics Card; 32GB GDDR7 Memory; PCIe 5.0 x16 - Micro Center
https://www.microcenter
Community Voting
Deal Score
+3
Good Deal
Bad Deal
Found by Bojjihuntindeals
Forum Thread
77% off
Found by Eragorn
Forum Thread
34% off
Found by phoinix
Forum Thread
50% off
Found by gaamn114
Forum Thread
64% off
Found by Eragorn
Forum Thread
80% off
Found by neomatrixviper
Forum Thread
Found by Red_Liz
Forum Thread
50% off
Found by suihuamo
Forum Thread
50% off
Found by babgaly
Forum Thread
50% off
Found by ItsSoCheap
Forum Thread
50% off
Join The Conversation
Share information with the community. Please follow our Community Guidelines and be kind!
16 Comments
Sign up for a Slickdeals account to remove this ad.
The NVIDIA GeForce RTX 5090 offers massive core-count superiority, featuring 21,760 CUDA cores compared to the 10,496 CUDA cores found on the NVIDIA RTX Pro 4500 Blackwell. This represents a 107% advantage in raw parallel processing hardware for the RTX 5090, putting it in an entirely different tier for raw compute
Super low TDP of just 200W though, can run 3 with the same power of a single 5090.
Pro has ECC too, idk how often memory related crashes happen though but this is for enterprise use.
Memory bandwidth is 1792 GBs vs 896 GB/s
But I can also make the argument that if someone needs a Pro card, they're already getting it paid for somewhere already and not waiting for a deal in SD.
But for Local AI crowd, they are comparing many diff cards to what suits them best. There are people getting 2 3060 TI cards 12GB x 2 to get to 24GB. or old Tesla V100 hacks to get the 32GB. VRAM. All interesting and I think people could consider them as they find what works. This would be good for someone who does need 32GB on a single card but can't stand the 5090 power draw and noise. or want to pack 64GB in a single PC without crazy power draw and did not need 96GB.
I considered RT6000 Pro but then it's actually slower than 5090 for what I use it for so it was pointless since I only do inference anyway but I did a research it.
I saw it and wanted to see what this Pro 4500 can do to see if I should use it but realized it's not for me and pasted what I learned.
Sign up for a Slickdeals account to remove this ad.
RTX Pro cards are warrantied through the seller so in the case of Dell you have a Dell warranty.
If money is not a concern, this card's advantages over 5090 for LLM inference is 1) nvlink capable so you can pair these up for tensor parallelism and pooled memory, the combined memory lets you run unquantized model for much higher quality LLM output and bigger context window. 2) lower power draw and physical profile.
(Edit: Actually there's no nvlink support for this card... that's made it worse. I thought this card is similar to my RTX A4500 pair which is nvlink bridged)
But if money is not a concern you should be looking at the RTX 6000 pro with 96gb VRAM - a single card solution
It does work and models will be loaded across the cards. Not as efficient as a single card with the larger VRAM but LLM inference engines like LM studio will split the requests across by default and you'll get better performance but not double performance.
AI Desktop Unified memory = useless for my use case. I got the equivalent M5 Max 128GB and I find a Mac more useful since it is my daily driver, larger models not that useful since the best open source ones are now 27B or so. If you really going to TRAIN then it may be DGX Spark but maybe you meant "using models" and not training models. but idk.
I don't think you'll run GLM 5.2 local, that you need about 300GB of RAM and a lot more $ and I personally think it's silly when online sub is super cheap and 99.99% has no need to run it locally other than thinking they have super secure stuff...
I ran 5080 16GB and 5070 TI 16GB. on the same computer (sold one oh well), and it definitely works this way
You can even mix 5080 16GB with 2070 8GB, I don't recall how it decides where it goes since it's asymmetrical but there are some settings.
Lots if youtube on this, One guy was running 4x3090 (96GB) vs. 1 DGX 128GB etc. lots of options for "cheaper hardware" for local AI
32GB VRAM with 2 5080/5070 is just $2000. vs $4000 with a single 5090 but it's a lot slower.
If you use 5060 TI 16GB then it's just $1200 to run 32GB VRAM and it runs inside of a SINGLE computer assuming you got 1200w PSU.
1000w will cut it too close, my testing ran up to 997w, good thing I had a 1200w in it at the time.
It does work and models will be loaded across the cards. Not as efficient as a single card with the larger VRAM but LLM inference engines like LM studio will split the requests across by default and you'll get better performance but not double performance.
AI Desktop Unified memory = useless for my use case. I got the equivalent M5 Max 128GB and I find a Mac more useful since it is my daily driver, larger models not that useful since the best open source ones are now 27B or so. If you really going to TRAIN then it may be DGX Spark but maybe you meant "using models" and not training models. but idk.
I don't think you'll run GLM 5.2 local, that you need about 300GB of RAM and a lot more $ and I personally think it's silly when online sub is super cheap and 99.99% has no need to run it locally other than thinking they have super secure stuff...
I ran 5080 16GB and 5070 TI 16GB. on the same computer (sold one oh well), and it definitely works this way
You can even mix 5080 16GB with 2070 8GB, I don't recall how it decides where it goes since it's asymmetrical but there are some settings.
Lots if youtube on this, One guy was running 4x3090 (96GB) vs. 1 DGX 128GB etc. lots of options for "cheaper hardware" for local AI
32GB VRAM with 2 5080/5070 is just $2000. vs $4000 with a single 5090 but it's a lot slower.
If you use 5060 TI 16GB then it's just $1200 to run 32GB VRAM and it runs inside of a SINGLE computer assuming you got 1200w PSU.
1000w will cut it too close, my testing ran up to 997w, good thing I had a 1200w in it at the time.
I am currently running 2x 5060ti 16Gb + an oculinked 3070 on an external power supply.
My internal psu is 750w. The 5060ti has a Tdp of 180w. How the heck were you pushing 1000w ?
If you read up where I actually wrote the config that I had running until I sold the 5080
5060 TI 16Gb was too slow so I got rid of the single 16GB I was just saying the cost to get 32GB is cheapest with that pairing at $1200
I don't like almost half the speed it was providing due to gimped memory bandwidth.
But I see where you could have thought I ran 997w with the 2 5060 TI's.
I did run 5080 and 5060 TI 16GB and 1000w was ok. I don't remember the peak because peak number got erased when I ran it with 5080/5070
Sign up for a Slickdeals account to remove this ad.
Join The Conversation
Share information with the community. Please follow our Community Guidelines and be kind!