And the price is now $125,186.99!
I am still thinking it might be worth picking one up (assuming I am able to purchase one as an individual and not have to go through my employer), but at this price you are starting to be comparable to a 4x H200 NVL system and that would give you a platform for developing NVLink-based software.
I do have some concerns about these units though:
Each OEM seems to ship the same spec 1600W PSU, and only one of them. The GB300 in the data center configs runs at 1300W - and in one of the recent YouTube videos there is a clip of a terminal running nvtop on a SuperMicro GB300, and it also shows 1300W. Is 1600W actually sufficient for powering these units? I also see some OEMs having a configuration option to add a GPU to drive a display, up to an RTX PRO 6000 Workstation card. I find it hard to believe these units can power the GB300 fully as well as a 600W GPU from the 1600W PSU. What is the story here? I suspect there are going to be serious pain points around powering these workstations (at least in <200V countries as they could theoretically swap for a larger PSU).
Also, GB300 in data center gets 288GB of HBM - and the DGX Station was advertised as “up to” that amount when it was announced last year. Now they are advertising 252GB of HBM. What happened to the other 36GB of RAM 
. I am hoping it is only a reduction in capacity and not a binned GB300 with fewer memory channels and maybe fewer SM cores, but it seems like the hardware was down specced since announcement.
In @wendell ‘s recent video where there was some discussion on these workstations - I hope you are able to cover them more in depth in a future video. If you do, it would be great to get some clarification on those issues (or if you already know the answer, please share!).
For others who, like me, are curious about switching from RTX GPUs to Hopper or Blackwell I have done some performance testing to compare different configurations. I was testing with a Wan 2.2 I2V (video generation) script that I have put together that will do the denoising steps over multiple GPUs using Ulysses for sequence parallelism with a couple of different CUDA configs.
For a 1248x832 resolution 81 frame generation using FP16 model weights with CFG != 1.0 (so both a conditional and unconditional pass per step) and PyTorch 2.11 provided SDPA kernels:
| GPU Type |
GPU Count |
Seconds per Step |
| RTX PRO 6000 (600W) |
1 |
50.9 seconds |
| RTX PRO 6000 (325W & +250 Clock Offset) |
1 |
60.5 seconds |
| RTX PRO 6000 (325W & +250 Clock Offset) |
4 (PCIe) |
16.4 seconds |
| H200 NVL (600W) |
2 (NVLink) |
17.0 seconds |
| H200 SXM (700W) |
2 (NVLink) |
15.6 seconds |
| H200 SXM (700W) |
8 (NVLink) |
3.9 seconds |
| B200 (1000W) |
1 |
12.2 seconds |
So, for Wan 2.2 denoising, a B200 @ 1000W is a bit over 4x faster than an RTX PRO 6000 @ 600W. According to spec sheets, at FP8, B200 is 9 petaFLOPs and GB300 is 10 petaFLOPs - so we might theoretically see a 11% improvement over the B200 numbers.
But the interesting numbers for me were that an H200 performs about 2x my RTX PRO 6000 Max-Q, and B200 is about 5x.
H200s scale remarkably better than RTX PROs thanks to NVLink, with very little lost to communication overhead compared to PCIe-based P2P communication. All of these numbers line up with what you would expect from the NVIDIA datasheets, but I found it interesting to be able to confirm it with a real-world use-case.
