News Nvidia Blackwell RTX Pro with up to 96GB of VRAM — even more demand for the limited supply of GPUs

[Admin]

Nvidia announced the upcoming Blackwell RTX Pro GPUs, which will power upcoming desktop and mobile workstations as well as data center inferencing platforms. The top solution will leverage the GB202 chip with up to 96GB of GDDR7 memory.

Nvidia Blackwell RTX Pro with up to 96GB of VRAM — even more demand for the limited supply of GPUs : Read more

 

[Li Ken-un]

So uh… what organ are y’all selling for this?

 

[hannibal]

So uh… what organ are y’all selling for this?

You have slaves to sell for these! Or their organ parts if needed.
😉

 

[Joseph_138]

I saw on wccftech that it's only 5% faster than a 5090. I saw somewhere else that it's supposed to sell for $10-15k. 5x-7.5x the price, for a 5% gain. No, thanks.

 

[Notton]

These are the poor man's B100/B200/B300, not the rich man's RTX 5090.
Presumably for AI workloads where your model requires more than 32GB but less than 96GB.

 

[A Stoner]

I saw on wccftech that it's only 5% faster than a 5090. I saw somewhere else that it's supposed to sell for $10-15k. 5x-7.5x the price, for a 5% gain. No, thanks.

Professional cards come with professional services and abilities that consumer cards do not come with. They are priced based on those extras that businesses demand and consumers do not. It is OK for a consumer card to flip a bit every now and again, not so for a professional one.

 

[Dr3ams]

What game needs a card like this?

 

[Elrabin]

I saw on wccftech that it's only 5% faster than a 5090. I saw somewhere else that it's supposed to sell for $10-15k. 5x-7.5x the price, for a 5% gain. No, thanks.

It'll be 5% faster until you fill the 5090s VRAM, then the RTX Pro will smoke it. We're not talking gaming here sir, we're talking massive models that even 32gb of VRAM is insufficient.

 

[Elrabin]

What game needs a card like this?

What part of "workstation/datacenter" card did you miss?

 

[Dr3ams]

What part of "workstation/datacenter" card did you miss?

Hmmm...doesn't say anywhere in the article that it is exclusive to anything.

But it does say this...

These will target both laptops and desktops, as well as standalone PCs and data center products.

 

[spongiemaster]

Hmmm...doesn't say anywhere in the article that it is exclusive to anything.

But it does say this...

It does.

Nvidia has also announced a change in branding, if you missed that. Where the previous professional and data center solutions were sold under various RTX names (RTX A6000/A5500/A5000/etc. for Ampere, then RTX 6000/5000/4500/etc. for Ada), the new Blackwell generation of professional and data center GPUs will use RTX Pro nomenclature.

 

[Dr3ams]

It does.

The article seems to contradict itself...

These will target both laptops and desktops, as well as standalone PCs and data center products.

It doesn't say: "These will target both laptops and desktops, as well as standalone PCs and other data center products".

 

[Elrabin]

The article seems to contradict itself...



It doesn't say: "These will target both laptops and desktops, as well as standalone PCs and other data center products".

Bro. Not every single desktop and laptop is consumer grade.

I'm typing to you right now from a "desktop" that is a workstation and have a laptop workstation 5 feet away.

Know what the difference is? ECC memory and professional tier graphic cards and all the software benefits that brings from a support/warranty perspective.


It was much simpler to explain this back when there was a clear separation between Geforce and Quadro/Tesla cards

I'm just glad they're splitting the lines more clearly again with the RTX PRO branding

 

[williamjeremiah]

Bro. Not every single desktop and laptop is consumer grade.

I'm typing to you right now from a "desktop" that is a workstation and have a laptop workstation 5 feet away.

Know what the difference is? ECC memory and professional tier graphic cards and all the software benefits that brings from a support/warranty perspective.


It was much simpler to explain this back when there was a clear separation between Geforce and Quadro/Tesla cards

I'm just glad they're splitting the lines more clearly again with the RTX PRO branding

But they'll likely ALL play Crysis.

 

[valthuer]

@JarredWaltonGPU just out of sheer curiosity; how is this card expected to perform in gaming, compared to 5090?

 

[JarredWaltonGPU]

@JarredWaltonGPU just out of sheer curiosity; how is this card expected to perform in gaming, compared to 5090?

In theory, it should be basically just as fast if not slightly faster. In practice? That's a more difficult question. I know you can install Studio drivers on a GeForce card, and I think you can also install Game Ready drivers on a professional card. If the latter is true, then you should get all the game optimizations and so performance should be at least equal to the 5090.

It has more compute, and 3X more memory, with the same bandwidth. There's no reason it should perform worse. That hasn't always been the case in the past where you would see professional cards get lower TDPs and slower VRAM + ECC. Actually, if you can't disable the ECC, that might drop performance a few percent as well. But otherwise the cards should be similar performance.

 

[helper800]

In theory, it should be basically just as fast if not slightly faster. In practice? That's a more difficult question. I know you can install Studio drivers on a GeForce card, and I think you can also install Game Ready drivers on a professional card. If the latter is true, then you should get all the game optimizations and so performance should be at least equal to the 5090.

It has more compute, and 3X more memory, with the same bandwidth. There's no reason it should perform worse. That hasn't always been the case in the past where you would see professional cards get lower TDPs and slower VRAM + ECC. Actually, if you can't disable the ECC, that might drop performance a few percent as well. But otherwise the cards should be similar performance.

The main difference is in the TMUs, ROPs, Tensor/RT cores and clock speeds compared to the 5090. Anything beyond 16-20GB of vRAM for a gaming card is basically useless assuming bandwidth parity. The 5090 is going to have a very significant clock speed advantage and available wattage GPU core , but will that make up for 22 less ROPs, less TMUs, less Tensor/RT cores? Who knows.

 

[JarredWaltonGPU]

The main difference is in the TMUs, ROPs, Tensor/RT cores and clock speeds compared to the 5090. Anything beyond 16-20GB of vRAM for a gaming card is basically useless assuming bandwidth parity. The 5090 is going to have a very significant clock speed advantage and available wattage GPU core , but will that make up for 22 less ROPs, less TMUs, less Tensor/RT cores? Who knows.

Not if you're looking at the RTX Pro 6000 Blackwell Workstation Edition card. That has a 600W TGP, more cores, more L2 cache, etc. It's the exact same design for the card itself as well, other than not having an HDMI output. Based on the posted specs, it has a boost clock of 2.6 GHz, which is 200 MHz higher than the RTX 5090's 2407 MHz boost clock.

What will the real-world clocks be? We know the 5090 can hit 2.75~2.85 GHz during gaming. If the RTX Pro 6000 has similar headroom (no reason it couldn't), it would still be faster. But maybe Nvidia clamps down on clocks harder for a workstation part, in order to guarantee stability and reliability. So then maybe it's a 150 MHz deficit. That's not enough to make up for the increased ROPS and other parts, though, meaning the professional card in theory still ends up faster.

Again, in practice it may be different, as we don't know for sure if full gaming driver optimizations work 100% of the time on the pro GPUs.

 

[helper800]

Not if you're looking at the RTX Pro 6000 Blackwell Workstation Edition card. That has a 600W TGP, more cores, more L2 cache, etc. It's the exact same design for the card itself as well, other than not having an HDMI output. Based on the posted specs, it has a boost clock of 2.6 GHz, which is 200 MHz higher than the RTX 5090's 2407 MHz boost clock.

What will the real-world clocks be? We know the 5090 can hit 2.75~2.85 GHz during gaming. If the RTX Pro 6000 has similar headroom (no reason it couldn't), it would still be faster. But maybe Nvidia clamps down on clocks harder for a workstation part, in order to guarantee stability and reliability. So then maybe it's a 150 MHz deficit. That's not enough to make up for the increased ROPS and other parts, though, meaning the professional card in theory still ends up faster.

Again, in practice it may be different, as we don't know for sure if full gaming driver optimizations work 100% of the time on the pro GPUs.

We will have to wait and see what happens with the boost clocks. I assume the 5090 will boost higher as it has more access to its 600w for GPU core. Don't forget that the pro card has to power 96gb of the same mem chips with ECC instead of just 32gb non-ECC. That's a few watts per extra chip and the pro card has double the amount of memory chips. GDDR6 uses about 2.8 watts per chip, Samsung says GDDR7 is 30% more efficient than GDDR6 so about 2w per chip x 16 chips = 32w

 

[JarredWaltonGPU]

We will have to wait and see what happens with the boost clocks. I assume the 5090 will boost higher as it has more access to its 600w for GPU core. Don't forget that the pro card has to power 96gb of the same mem chips with ECC instead of just 32gb non-ECC. That's a few watts per extra chip and the pro card has double the amount of memory chips. GDDR6 uses about 2.8 watts per chip, Samsung says GDDR7 is 30% more efficient than GDDR6 so about 2w per chip x 16 chips = 32w

True, but then that's part of the extra 25W of TGP. Also, there are a lot of gaming workloads where the 5090 doesn't get anywhere near its power limit, so there could be a lot of scenarios where it just boils down to core counts and clocks and power doesn't really factor into the equation.