RTX 6000 ADA is now available!

That core count can give an idea of what the 4090 TI will look like I imagine, 18,176 is getting quite close to a full18,432 silicon
 
What do workstation models do differently than consumer models these days, again?

Back in the day the only real difference I was aware of was that consumer models disabled hardware antialiased lines in the drivers (or firmware, not sure which).

These days I don't have a clue what the workstation models can do that the consumer models can't.

At $7,350 for a model that is on paper ~11% faster than a 4090 (albeit with double the VRAM) that really should be spelled out.

Maybe that's all there is? Double the RAM and ECC instead of non-ECC? That's a steep price to pay for that small benefit.

Almost makes you wonder why Enterprise users bother.

How it maintains the same 2.5Ghz clock with 11% more cores, and 150 fewer watts than the 4090 is a mystery. Better binning? Or does it just throttle the hell out of the thing? :p

LukeTbk said:
That core count can give an idea of what the 4090 TI will look like I imagine, 18,176 is getting quite close to a full18,432 silicon
Click to expand...

Yeah, makes you wonder if some idiot Youtuber will drop $7,350 as some stupid flex so they can claim "I am the fastest". At least for a few months until the 4090 TI (which likely also will be unobtainable) drops.
 
DPI said:
New Quadro. A hardware dongle to run professional nvidia drivers required by many pro applications.
Click to expand...
Naming scheme is absolutely horrid. and i thought they dropped quadro as well. which makes it even more horrible.
 
Zarathustra[H] said:
What do workstation models do differently than consumer models these days, again?

Back in the day the only real difference I was aware of was that consumer models disabled hardware antialiased lines in the drivers (or firmware, not sure which).

These days I don't have a clue what the workstation models can do that the consumer models can't.

At $7,350 for a model that is on paper ~11% faster than a 4090 (albeit with double the VRAM) that really should be spelled out.

Maybe that's all there is? Double the RAM and ECC instead of non-ECC? That's a steep price to pay for that small benefit.

Almost makes you wonder why Enterprise users bother.

How it maintains the same 2.5Ghz clock with 11% more cores, and 150 fewer watts than the 4090 is a mystery. Better binning? Or does it just throttle the hell out of the thing? :p



Yeah, makes you wonder if some idiot Youtuber will drop $7,350 as some stupid flex so they can claim "I am the fastest". At least for a few months until the 4090 TI (which likely also will be unobtainable) drops.
Click to expand...
I have a box with a pair of RTX A6000s and they game like absolute crap, strangely it gets better when I subdivide them a little, but better is a highly relative term because I would still choose to game on the old ASUS sitting behind me running the 1650 than I would the server running the A6000s.
But really the difference is memory, drivers, and PCB components.
 
Zarathustra[H] said:
What do workstation models do differently than consumer models these days, again?

Back in the day the only real difference I was aware of was that consumer models disabled hardware antialiased lines in the drivers (or firmware, not sure which).
Click to expand...
OUtside VRAM amount it is still usually under those application, maybe less about wire rendering than before:

https://techgage.com/article/best-g...-of-catia-solidworks-siemens-nx-blender-more/

In some stuff in Catia the older Titan was still beating a 3090, despite being usually much much slower:
A-1080p-Viewport-Performance-February-2021-680x370.jpg


X-1080p-Viewport-Performance-February-2021-680x370.jpg


o-1080p-Viewport-Performance-February-2021-680x370.jpg


Here the ability to stack them in a server rack/workstation out of the box in a signed workstation could be part of the appeal has well.
 
Thunderdolt said:
No NVLink on it. Hmmmm.
Click to expand...
No, they told us that a while ago.
https://www.techpowerup.com/299107/...n Huang in,over a separate physical interface).
PCIe Gen4 is fast enough if you need to link a few of them together that way.
Yeah the dedicated NVlinks are faster but the Threadripper and Xeon workstation platforms have other bottlenecks that you just end up running against instead so the NVlinks there don't get to fully stretch their legs.
Once you throw them in a full server it changes but with how things have changed in the latest CUDA builds the cards work more independently decreasing the need for NVLink until you get into the stupid fast numbers the OAM solutions provide and that has its own NVLink fabric and blah blah blah.
I get why they removed it from the cards, and it's not a big deal.
 
LukeTbk said:
That core count can give an idea of what the 4090 TI will look like I imagine, 18,176 is getting quite close to a full18,432 silicon
Click to expand...

I'm actually wondering if there will be a 4090TI at all this generation.

They have leapfrogged AMD so far with the 4090 that they really don't need to come out with a higher tier to stay ahead, and I have a feeling AMD won't catch back up for a while, unless they do something drastic crazy high power with those chiplets.

Might as well save the top bin of fully enabled ADA chips for professional GPU's that they can sell for over $7k than put them in consumer GPU's that will have to sell for no more than $2k.
 
Last edited:
Zarathustra[H] said:
I'm actually wondering if there will be a 4090TI at all this generation.

They have leapfrogged AMD so far with the 4090 that they really don't need to come out with a higher tier to stay ahead, and I have a feeling AMD won't catch back up for a while, unless they do something drastic xrazy high power with those chiplets.

Might as well save the top bin of fully enabled ADA chips for professional GPU's that they can sell for over $7k than put them in consumer GPU's that will have to sell for no more than $2k.
Click to expand...
Even if they did offer a 4090TI the only significant improvement they could really add would be DP 2.1 for the yet-to-exist screens, because any real performance improvement they could bring to the card for gaming doesn't really add much given the current displays out there and you could say well 11% is 11% and that is true and not statistically insignificant but it may not be something noticed outside of a bar graph I mean going from 60 to 66 fps isn't something you are going to really notice. And really by the time the card launches its intended audience is probably looking more towards next-gen than they are a minor incremental upgrade, the 4090 has an extremely small target audience and it's nailed every metric it needs to for that audience but it's still a small audience.
 
Lakados said:
Even if they did offer a 4090TI the only significant improvement they could really add would be DP 2.1 for the yet-to-exist screens, because any real performance improvement they could bring to the card for gaming doesn't really add much given the current displays out there and you could say well 11% is 11% and that is true and not statistically insignificant but it may not be something noticed outside of a bar graph I mean going from 60 to 66 fps isn't something you are going to really notice. And really by the time the card launches its intended audience is probably looking more towards next-gen than they are a minor incremental upgrade, the 4090 has an extremely small target audience and it's nailed every metric it needs to for that audience but it's still a small audience.
Click to expand...

Well, it seems like it is a lot larger of an audience than there is supply, judging by how difficult it is to get your hands on one, even at the outrageous $1,600 MSRP
 
Lakados said:
Even if they did offer a 4090TI the only significant improvement they could really add would be DP 2.1 for the yet-to-exist screens,
Click to expand...
Maybe keeping the same bus but gaining bandwith going with GDDR7 ?

https://www.tweaktown.com/news/8883...emory-with-insane-36gbps-bandwidth/index.html

That could be a "free" 70% memory bandwith boost, no idea how possible it is, apparently:
In fact, earlier this year I reported on leaks from Moore's Law is Dead, who said that NVIDIA's next-gen AD102-based GeForce RTX 4090 could... not would, but could use GDDR7 memory. MLID explained that the memory controllers in Ada Lovelace dies are "vastly improved over Ampere, with AD102 being built to support GDDR7 memory speeds, and possibly GDDR7 itself".

Same for a RDNA 3 2023 refresh, little +12 core, +60/70 on bandwith and at 4k/max RT heavy memory workload maybe you can get a 20% gain ?

Looking at the price they sell those A6000, it would be after they are not selling all of them too I imagine, if ever.
 
Lakados said:
No, they told us that a while ago.
https://www.techpowerup.com/299107/jensen-confirms-nvlink-support-in-ada-lovelace-is-gone#:~:text=NVIDIA CEO Jensen Huang in,over a separate physical interface).
PCIe Gen4 is fast enough if you need to link a few of them together that way.
Yeah the dedicated NVlinks are faster but the Threadripper and Xeon workstation platforms have other bottlenecks that you just end up running against instead so the NVlinks there don't get to fully stretch their legs.
Once you throw them in a full server it changes but with how things have changed in the latest CUDA builds the cards work more independently decreasing the need for NVLink until you get into the stupid fast numbers the OAM solutions provide and that has its own NVLink fabric and blah blah blah.
I get why they removed it from the cards, and it's not a big deal.
Click to expand...
I guess my concern is as much from the software side as the hardware. How much re-development is required in order to replace NVLink usage with PCIe Gen 5 Peer-to-Peer usage?

The removal of NVLink from the workstation cards has nothing to do with speed or utility. It is entirely about forcing differentiation between workstation and server cards. IMO, Nvidia is pushing hard to play the long game of getting their customers off of workstations and into simple machines connected to dedicated servers.
 
Thunderdolt said:
I guess my concern is as much from the software side as the hardware. How much re-development is required in order to replace NVLink usage with PCIe Gen 5 Peer-to-Peer usage?

The removal of NVLink from the workstation cards has nothing to do with speed or utility. It is entirely about forcing differentiation between workstation and server cards. IMO, Nvidia is pushing hard to play the long game of getting their customers off of workstations and into simple machines connected to dedicated servers.
Click to expand...
From the bits of documentation I have read so far, there is no programming change needed just some CUDA library updates.
But looking at Nvidia's own press releases it's not to push dedicated servers but Nvidia's cloud solutions, it ups the cost of entry for upper-end AI where PCIe4 speeds won't cut it.
Their moderators are responding to questions on the subject with things like.
"buying a full-fledged DGX H100 is beyond the budget of smaller companies or educational institutions. For that there are alternatives like Cloud service providers, for example the NVIDIA partner Cyxtera. Check the DGX page for DGX as a service." and "you also have the possibility to test drive AI development on H100 in our Launchpads"
I'll need to see benchmarks of course, but supposedly the speed decreases of the PCIe4, as opposed to the NVlink, won't change much unless you are working with training sets larger than the combined cards' memory pool which would require both cards to dump data from VRam and start coordinating and swapping with system ram but even then the developer forums are saying that you would still see an increase in performance over the Ampere cards because of just how much faster the Ada cores are.
 
Last edited:
The A6000 is for us research types. We need stable development drivers, extra support from nvidia when drivers and software clashes, and application support which may not exist for normal consumer level cards. These cards are run at full load doing machine learning tasks, or running in clusters dealing with things like password cracking and decryption.
 
Slade said:
The A6000 is for us research types. We need stable development drivers, extra support from nvidia when drivers and software clashes, and application support which may not exist for normal consumer level cards. These cards are run at full load doing machine learning tasks, or running in clusters dealing with things like password cracking and decryption.
Click to expand...
Yeah, I buggered up my vGPU drivers last week during a routine update, even versions (even the most recent) don't work properly with the cloud license servers. had to go back to an odd version, apparently, the odd numbers are their LTS service branch of driver support. Good to know.
But Nvidia support had me patched up, licenses intact, and back in operation in under 20 minutes.
 
Zarathustra[H] said:
I'm actually wondering if there will be a 4090TI at all this generation.

They have leapfrogged AMD so far with the 4090 that they really don't need to come out with a higher tier to stay ahead, and I have a feeling AMD won't catch back up for a while, unless they do something drastic crazy high power with those chiplets.

Might as well save the top bin of fully enabled ADA chips for professional GPU's that they can sell for over $7k than put them in consumer GPU's that will have to sell for no more than $2k.
Click to expand...
There's a rumor that AMD was aiming for 3GHz on Navi 31 and a hardware "bug" prevented it. Supposedly they have it figured out and could hit it with a refresh. We'll see...
 
The term “workstation” for these cards is used very loosely. These cards mostly get used for vGPU in rack setups.
 
Mchart said:
The term “workstation” for these cards is used very loosely. These cards mostly get used for vGPU in rack setups.
Click to expand...
That's where mine mostly reside now, works really well with VMWare Horizon
 
You must log in or register to reply here.
Back
Top