You've got to love a site that keeps hosting their reviews, images and all, in a perfectly readable format and without major errors 21 years after publication.
Gamers are obsessed with overhyping product launches so it makes sense to push product out without accurately assessing demand, since people will line up to buy it anyway.
Customers looking at quadro are much less likely to impulse buy, especially enterprise that need large orders, and relationships are critical to maintaining long term sales. If anything quadros will be delayed until they have enough volume to meet demand while 3080/3090 will continue to drip in dozens at a time.
A6000 cards use full featured GA102, while 3080 and 3090 do not. So the number of ex-Quadro card that can be build depends only on the capacity of Samsung to bake perfect chips. Seen that it seems it has problem doing any kind of big chip (full, less good and even-less-than-good) shortage is not determined by the existence of this ex-Quadro cards that are usually sold in small percentages with respect to the gaming ones.
Considering the 3090 does not get quadro validated drivers and is considered "titan class" but not titan, it seems like nothing has changed from that standpoint, so no quadro optimizations on gaming cards for software that requires vendor validation.
AFAIK that is valid only for OpenCL accelerated applications, while for the rest (compute and DX) the support should be the same as there are lot of those cards sold for 3D rendering.
Support for many opencl and opengl software does not require vendor certification, which was already the case for every gaming GPU even before turing.
Anything that requires VENDOR certs (i.e.Siemens NX) do not get accelerated drivers on gaming GPUs, even the 3090 with studio drivers. Confirmed to reviewers by nvidia themselves.
I was speaking about acceleration options and support, not certification.
Certification is needed for certain works or certain working environments, but if work for such markets you can surely afford a certified piece of HW (that has also better HW and controls, like ECC and not only ad hoc drivers).
While, if you use the GPU for doing usual work (be it modeling, rendering or calculations) you can use a 3090 (or even a couple of them) to accelerate all you want, but, as said, few OpenGL applications (but again, if you use them regularly with a paid license the cost of a Quadro, not necessarily the latest most powerful one to just accelerate them, is not a problem).
If you actually read nvidia's comment on the matter, it is 100% driver related, not ECC related. 3080/3090 do not ship with driver side optimizations for software that requires vendor certs. Software like blender does not require certs and will run normally regardless of gaming or quadro hardware, but you can see in the specviewperf test that NX simulation runs 21x worse on a 3090 than a titan, which hardware wise makes absolutely no sense and is entirely due to its driver.
In case it wasn't clear, there is no piece of hardware on the RTX titan that performs 21x faster than the 3090, including the tensor cores and error correction, yet the titan is indeed 21x faster in NX sim and slightly faster in solidworks even with slower hardware.
Nvidia releases "Studio" drivers now that contain those acceleration unlocks. And as CiccioB mentioned, this has only ever really applied to certain OpenGL applications. The DirectX and OpenCL interfaces are completely wide open on the gaming cards, there are no artificial limitations put in place there because many games are now using the GPU for compute so Nvidia and AMD had to give in to studio demand.
How is removing a brand from a name but keeping it targeted to the same market makes the product cheaper? AMD has nothing in this market that can compete. RDNA2 lacks most of Nvidia advanced features (Tensor for instance) and cannot be used for a bunch of works these cards are meant for (IA, DeepLearning and 3D rendering where tensor helps with filtering).
It's a fanboy hope for a "wait and see" suggestion when RDNA2 won't have any tensor acceleration for sure (the die is already big the way it is now with hybrid faked RT accelerators). You may be speaking about CDNA, but that is N/A for now.
Machine learning using GEMM, 3D, etc., with AMD Radeon Pro Graphics and AMD RDNA Architecture, Vega, GCN even APUs & mobile
CDNA is CPU/GPU on what appears to be Eypc, a fabric arch and unified memory(?) ___ does not mean that the above will not to continue work just dandy in the pro environment INCLUDING RDNA2 (► What Spunjji said)
We will find out in 9 days or so ___ and I am feeling a sense of desperation from the other green team.
Stating that compute cards can be provisioned for proviz isn't exactly truthful. For example the A100 doesn't even have NVENC and thus would be terrible for VDI.
But that wasn't what the article said. They said some of those functions are overlapping. Obviously, an A100, which has literally no display port would be a hard sell for local visualization.
The point is that the functions performed by those products are increasingly overlapping, hence the end of the distinct brand names (quadro and tesla). Each model will still have a set of capabilities that will make it fit or unfit for your purpose.
Seems to me the user now has to assess the new cards on their own merit and decide which one is best, rather than sticking to a known brand name. For some people that's going to make it harder.
True, but if you're shopping for a multi thousand $ card and you buy the wrong product cause you were not willing to do your homework, the fault lies in you not NVIDIA. I'm more interested in seeing if this decision will lead to a lesser amount of "artificial segmentation" in the sector, and if that will come or not with an increased price tag attached.
To me, all this means is that everything will just cost more.
The RTX generation of cards is great for single precision and half precision computing, but it SUCKS for double precision computing.
I'm still waiting to see what the official numbers are going to be, but I would not be surprised if FP64 is going to be 1/32 FP32 like it typically is for consumer-grade cards which means that my Quadro 6000 is still faster than most cards that are out and availble now in terms of FP64 performance, which is just sad.
I rather they keep it simple. Geforce => Consumer Market (Games, Content Creation, and Desktop Compute scenarios.) Quadro => Business Market (Pro-visual, Production, and Server Compute scenarios.)
We’ve updated our terms. By continuing to use the site and/or by logging into your account, you agree to the Site’s updated Terms of Use and Privacy Policy.
34 Comments
Back to Article
oRAirwolf - Tuesday, October 13, 2020 - link
:|Smell This - Tuesday, October 13, 2020 - link
'Quadro' comes from the initial collaboration/litigation between SGI and nVidia 20+ years ago ___ may have had some Elsa in there, too.3Dlabs Oxygen GVX1 PCI
by Anand Lal Shimpi on December 1, 1999 3:26 AM EST
https://www.anandtech.com/show/418
Samus - Wednesday, October 14, 2020 - link
I remember when a Quadro was nothing more than soldering some via's on the videocard, tricking the drivers into thinking you had a Quadro ;)Valantar - Thursday, October 15, 2020 - link
You've got to love a site that keeps hosting their reviews, images and all, in a perfectly readable format and without major errors 21 years after publication.domboy - Thursday, October 15, 2020 - link
Agreed! That was a fun trip down memory lane.DiHydro - Thursday, October 15, 2020 - link
Big shout out to the team here at Anandtech for doing this!zamroni - Tuesday, October 13, 2020 - link
So the already scarce ga102 will be prioritized for expensive a6000 cards rather than "cheap" 3080 and 3090.Gamers should expect more shortage
whatthe123 - Tuesday, October 13, 2020 - link
Gamers are obsessed with overhyping product launches so it makes sense to push product out without accurately assessing demand, since people will line up to buy it anyway.Customers looking at quadro are much less likely to impulse buy, especially enterprise that need large orders, and relationships are critical to maintaining long term sales. If anything quadros will be delayed until they have enough volume to meet demand while 3080/3090 will continue to drip in dozens at a time.
CiccioB - Wednesday, October 14, 2020 - link
A6000 cards use full featured GA102, while 3080 and 3090 do not.So the number of ex-Quadro card that can be build depends only on the capacity of Samsung to bake perfect chips.
Seen that it seems it has problem doing any kind of big chip (full, less good and even-less-than-good) shortage is not determined by the existence of this ex-Quadro cards that are usually sold in small percentages with respect to the gaming ones.
Gigaplex - Tuesday, October 13, 2020 - link
What does this mean for the "gaming vs business" driver validation segmentation?whatthe123 - Tuesday, October 13, 2020 - link
Considering the 3090 does not get quadro validated drivers and is considered "titan class" but not titan, it seems like nothing has changed from that standpoint, so no quadro optimizations on gaming cards for software that requires vendor validation.CiccioB - Wednesday, October 14, 2020 - link
AFAIK that is valid only for OpenCL accelerated applications, while for the rest (compute and DX) the support should be the same as there are lot of those cards sold for 3D rendering.CiccioB - Wednesday, October 14, 2020 - link
I meant OpenGL, not OpenCL, sorrywhatthe123 - Wednesday, October 14, 2020 - link
Support for many opencl and opengl software does not require vendor certification, which was already the case for every gaming GPU even before turing.Anything that requires VENDOR certs (i.e.Siemens NX) do not get accelerated drivers on gaming GPUs, even the 3090 with studio drivers. Confirmed to reviewers by nvidia themselves.
https://i.imgur.com/9Lk1z50.png
https://i.imgur.com/1Xxrzg3.png
CiccioB - Wednesday, October 14, 2020 - link
I was speaking about acceleration options and support, not certification.Certification is needed for certain works or certain working environments, but if work for such markets you can surely afford a certified piece of HW (that has also better HW and controls, like ECC and not only ad hoc drivers).
While, if you use the GPU for doing usual work (be it modeling, rendering or calculations) you can use a 3090 (or even a couple of them) to accelerate all you want, but, as said, few OpenGL applications (but again, if you use them regularly with a paid license the cost of a Quadro, not necessarily the latest most powerful one to just accelerate them, is not a problem).
whatthe123 - Wednesday, October 14, 2020 - link
If you actually read nvidia's comment on the matter, it is 100% driver related, not ECC related. 3080/3090 do not ship with driver side optimizations for software that requires vendor certs. Software like blender does not require certs and will run normally regardless of gaming or quadro hardware, but you can see in the specviewperf test that NX simulation runs 21x worse on a 3090 than a titan, which hardware wise makes absolutely no sense and is entirely due to its driver.whatthe123 - Wednesday, October 14, 2020 - link
In case it wasn't clear, there is no piece of hardware on the RTX titan that performs 21x faster than the 3090, including the tensor cores and error correction, yet the titan is indeed 21x faster in NX sim and slightly faster in solidworks even with slower hardware.Kakkoii - Wednesday, October 14, 2020 - link
Nvidia releases "Studio" drivers now that contain those acceleration unlocks.And as CiccioB mentioned, this has only ever really applied to certain OpenGL applications. The DirectX and OpenCL interfaces are completely wide open on the gaming cards, there are no artificial limitations put in place there because many games are now using the GPU for compute so Nvidia and AMD had to give in to studio demand.
boeush - Tuesday, October 13, 2020 - link
"can be provisioned as a virtual ProViz"- with the proviso that it's done by a pro vis a vis the provisioning for the ProViz
wolfesteinabhi - Wednesday, October 14, 2020 - link
it may also allow them to price them lower or compete with AMD on pricing .. incase their new RDNA/CDNA cards are any good.(to me this move by nvidia looks similar to how Intel was moving its "platinum" cpus to "gold" tier with new name ..to price them lower,etc)
CiccioB - Wednesday, October 14, 2020 - link
How is removing a brand from a name but keeping it targeted to the same market makes the product cheaper?AMD has nothing in this market that can compete. RDNA2 lacks most of Nvidia advanced features (Tensor for instance) and cannot be used for a bunch of works these cards are meant for (IA, DeepLearning and 3D rendering where tensor helps with filtering).
Spunjji - Wednesday, October 14, 2020 - link
Bold statement to make when we don't actually know any of this :|CiccioB - Wednesday, October 14, 2020 - link
It's a fanboy hope for a "wait and see" suggestion when RDNA2 won't have any tensor acceleration for sure (the die is already big the way it is now with hybrid faked RT accelerators).You may be speaking about CDNA, but that is N/A for now.
Smell This - Wednesday, October 14, 2020 - link
Your pants are on fire in so many different ways . . .AMD Radeon Pro Graphics Supported on AMD Radeon Software
https://www.amd.com/en/technologies/radeon-pro-sof...
CiccioB - Thursday, October 15, 2020 - link
What is that link related to what I have stated?Smell This - Monday, October 19, 2020 - link
high performance computing on graphics processing units: hgpu.orghttps://hgpu.org/?s=amd+gemm
Machine learning using GEMM, 3D, etc., with AMD Radeon Pro Graphics and AMD RDNA Architecture, Vega, GCN even APUs & mobile
CDNA is CPU/GPU on what appears to be Eypc, a fabric arch and unified memory(?) ___ does not mean that the above will not to continue work just dandy in the pro environment INCLUDING RDNA2 (► What Spunjji said)
We will find out in 9 days or so ___ and I am feeling a sense of desperation from the other green team.
evilspoons - Wednesday, October 14, 2020 - link
The expression is "one and the same", not "one in the same".Jbon - Wednesday, October 14, 2020 - link
Stating that compute cards can be provisioned for proviz isn't exactly truthful. For example the A100 doesn't even have NVENC and thus would be terrible for VDI.frenchy_2001 - Wednesday, October 14, 2020 - link
But that wasn't what the article said. They said some of those functions are overlapping.Obviously, an A100, which has literally no display port would be a hard sell for local visualization.
The point is that the functions performed by those products are increasingly overlapping, hence the end of the distinct brand names (quadro and tesla). Each model will still have a set of capabilities that will make it fit or unfit for your purpose.
Carl Bicknell - Wednesday, October 14, 2020 - link
Seems to me the user now has to assess the new cards on their own merit and decide which one is best, rather than sticking to a known brand name. For some people that's going to make it harder.El_Rizzo - Thursday, October 15, 2020 - link
True, but if you're shopping for a multi thousand $ card and you buy the wrong product cause you were not willing to do your homework, the fault lies in you not NVIDIA.I'm more interested in seeing if this decision will lead to a lesser amount of "artificial segmentation" in the sector, and if that will come or not with an increased price tag attached.
PaulHoule - Wednesday, October 14, 2020 - link
For me I had enough bad experiences with the Quadro brand that I am glad to see it go.alpha754293 - Thursday, October 15, 2020 - link
To me, all this means is that everything will just cost more.The RTX generation of cards is great for single precision and half precision computing, but it SUCKS for double precision computing.
I'm still waiting to see what the official numbers are going to be, but I would not be surprised if FP64 is going to be 1/32 FP32 like it typically is for consumer-grade cards which means that my Quadro 6000 is still faster than most cards that are out and availble now in terms of FP64 performance, which is just sad.
bottle23 - Thursday, October 15, 2020 - link
I rather they keep it simple.Geforce => Consumer Market (Games, Content Creation, and Desktop Compute scenarios.)
Quadro => Business Market (Pro-visual, Production, and Server Compute scenarios.)