Well the last Titan was a gamer card w/HBM2 instead. Now the rumor is that it'll essentially just be a fully unlocked card with 24GB of consumer RAM (GDDR6x). Which I'd be totally fine with, at least then you could run some of the bigger deep learning models without cashing out $2500. When AMD can sell 8GB of RAM on the RX 570 for less than $200 then you know 24GB worth of RAM shouldn't cost that much. They just lack competition.
Source? Because I really doubt that. They also sell the RX 5500 xt with 8GB and I highly doubt they would lose money on three generations of mid range cards.
Source? AMD quarters results. See the gain of the GPU+CPU division. They are so low for a reason. And it is certainly not due to CPU over selling.
GPU division has been in red much before AMD hid the revenue under the same CPU invoice in the reports. And now, after years of combiner red, the GPU+CPU division positive. Starting from the large acceptance of Zen architecture on the market. While GPU numbers monitored by market share are still bad if not worse. With great peaces of market not or badly covered (top, server and professional). Why should anyone believe it is now suddenly gaining money?
I don't know if AMD lost money on the 470/480, but I'd venture Global Foundries did on every 14/12nm die sold. They never explicitly pointed out AMD, but they always claimed it their leading edge nodes were not profitable and they really only had a single leading edge node customer.
The Titan V wasn't a gaming card it was a card aimed at scientific computing with its 7.450 TFLOPS of FP64 power. While the Titan RTX was aimed at deep learning with 24GB of DDR6 memory.
Of course it doesn't have CUDA because that's nVidia's proprietary platform. AMD made a tool called HIPIFY which automatically converts source from CUDA to HIP. Tensorflow-ROCM works very well, I have been using it for over a year. The Radeon VII was the best GPU for full precision training. Outperformed the RTX 2080 Ti and had more memory while being almost half the price.
The Titan V was lacking memory for training SOTA models which the Titan RTX filled the position. Turing also had second generation tensor cores with a few benefits over Volta's first generation tensor cores.
I'm having a bit of trouble finding a source on this, but I thought Jensen said, in the GTC 2020 keynote, that they would continue to support all workloads on all GPUs, or something to that effect.
I believe this was intended to draw a contrast between them and AMD's rumored compute-only Arcturus chip.
I think the confusion arises from the fact that they refer to their tensor accelerator products as "GPU"; e.g. "NVIDIA V100 TENSOR CORE GPU", "NVIDIA A100 Tensor Core GPU".
In other words, when they say "A100 GPU", they mean the card, not the actual GPU chip.
The same error was made in your first ampere/A100 article:
"Officially, this is the name of both the GPU and the accelerator incorporating it; and at least for the moment they’re both one in the same, since there is only the single accelerator using the GPU."
So, is most of the power reduction actually coming from cutting the NVLinks down? I don't know about the current NVLink, but the first two versions, Nvidia claimed they didn't have to do external retimers or redrivers, which I can only assume drove up the power consumption, especially with the distances some of the NVLink signals have to travel on the HGX/DGX system vs a NVLink bridge.
Thanks Ryan! Also, just to add that cold shower that reality provides so well, I suggest to add the MSRP for those cards. Yes, they'll all run Crysis, but... (And the accelerators don't even have a video signal out)
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.
Facebook
Twitter
RSS
29 Comments
Back to Article
mode_13h - Monday, June 22, 2020 - link
I hope they release a Titan A version.Kjella - Monday, June 22, 2020 - link
Well the last Titan was a gamer card w/HBM2 instead. Now the rumor is that it'll essentially just be a fully unlocked card with 24GB of consumer RAM (GDDR6x). Which I'd be totally fine with, at least then you could run some of the bigger deep learning models without cashing out $2500. When AMD can sell 8GB of RAM on the RX 570 for less than $200 then you know 24GB worth of RAM shouldn't cost that much. They just lack competition.CiccioB - Monday, June 22, 2020 - link
AMD lose money on the 480/470, so why should you expect a healthy company like Nvidia to do the same?ragenalien - Monday, June 22, 2020 - link
Source? Because I really doubt that. They also sell the RX 5500 xt with 8GB and I highly doubt they would lose money on three generations of mid range cards.CiccioB - Tuesday, June 23, 2020 - link
Source? AMD quarters results.See the gain of the GPU+CPU division. They are so low for a reason. And it is certainly not due to CPU over selling.
GPU division has been in red much before AMD hid the revenue under the same CPU invoice in the reports. And now, after years of combiner red, the GPU+CPU division positive. Starting from the large acceptance of Zen architecture on the market.
While GPU numbers monitored by market share are still bad if not worse. With great peaces of market not or badly covered (top, server and professional).
Why should anyone believe it is now suddenly gaining money?
jeremyshaw - Monday, June 22, 2020 - link
I don't know if AMD lost money on the 470/480, but I'd venture Global Foundries did on every 14/12nm die sold. They never explicitly pointed out AMD, but they always claimed it their leading edge nodes were not profitable and they really only had a single leading edge node customer.Irata - Monday, June 22, 2020 - link
I very much doubt they do.That said, those are different cards for a very different audience.
quorm - Monday, June 22, 2020 - link
What you're saying about the rtx titan is incorrect. It has 24gb gddr6 ram. The new titan will probably cost just as much.WannaBeOCer - Monday, June 22, 2020 - link
The Titan V wasn't a gaming card it was a card aimed at scientific computing with its 7.450 TFLOPS of FP64 power. While the Titan RTX was aimed at deep learning with 24GB of DDR6 memory.mode_13h - Monday, June 22, 2020 - link
The cool thing about the Titan V was that it could do it all. Deep learning, GPU-compute, graphics, and gaming.We got one for deep learning, where I work, but it was idle more than enough for me to fool around with graphics on it.
mode_13h - Monday, June 22, 2020 - link
Same is true of the Radeon VII, I should add. And it's much cheaper, although not as well-suited to deep learning.WannaBeOCer - Monday, June 22, 2020 - link
The Radeon VII is fine for full precision training. The RTX 2080 Ti and Titan RTX are 3x faster when using half precision.firewrath9 - Tuesday, June 23, 2020 - link
Except of course, it lacks CUDA, and the current ROCm implementation in Tensorflow is frankly, trash.WannaBeOCer - Wednesday, June 24, 2020 - link
Of course it doesn't have CUDA because that's nVidia's proprietary platform. AMD made a tool called HIPIFY which automatically converts source from CUDA to HIP. Tensorflow-ROCM works very well, I have been using it for over a year. The Radeon VII was the best GPU for full precision training. Outperformed the RTX 2080 Ti and had more memory while being almost half the price.mode_13h - Thursday, June 25, 2020 - link
Thanks for sharing your experience. It's good to hear.WannaBeOCer - Monday, June 22, 2020 - link
The Titan V was lacking memory for training SOTA models which the Titan RTX filled the position. Turing also had second generation tensor cores with a few benefits over Volta's first generation tensor cores.mode_13h - Tuesday, June 23, 2020 - link
Yeah, the newer GPU is better. Nothing new or surprising about that.Still, for its day, the Titan V was sweet if you could afford it.
ImSpartacus - Monday, June 22, 2020 - link
Nvidia can't.I believe the GA100 GPU lacks a couple things necessary to directly output graphics. They were removed to save die space.
So it can't easily be used in a graphics card with display outputs out of the back of the card.
I would expect a Titan based on GA102 though.
mode_13h - Monday, June 22, 2020 - link
I'm having a bit of trouble finding a source on this, but I thought Jensen said, in the GTC 2020 keynote, that they would continue to support all workloads on all GPUs, or something to that effect.I believe this was intended to draw a contrast between them and AMD's rumored compute-only Arcturus chip.
CiccioB - Tuesday, June 23, 2020 - link
The GA100 comes with TMUs and ROPs, quite useless if they where not intended for graphics workload, don't you think?mode_13h - Wednesday, June 24, 2020 - link
Thanks. I was looking for that, as well, but didn't find any such specs.eddman - Monday, June 22, 2020 - link
Are you sure the GPU chip is also called A100? Even nvidia calls it GA100:https://devblogs.nvidia.com/nvidia-ampere-architec...
I think the confusion arises from the fact that they refer to their tensor accelerator products as "GPU"; e.g. "NVIDIA V100 TENSOR CORE GPU", "NVIDIA A100 Tensor Core GPU".
In other words, when they say "A100 GPU", they mean the card, not the actual GPU chip.
eddman - Monday, June 22, 2020 - link
https://www.nvidia.com/content/dam/en-zz/Solutions...Page 37
Ryan Smith - Monday, June 22, 2020 - link
Ahh yep, there's a typo. Thanks!eddman - Monday, June 22, 2020 - link
The same error was made in your first ampere/A100 article:"Officially, this is the name of both the GPU and the accelerator incorporating it; and at least for the moment they’re both one in the same, since there is only the single accelerator using the GPU."
CiccioB - Monday, June 22, 2020 - link
The same mistake was done with Volta launch, where the V100 accelerator name was commonly exchanged with GV100 when talking about the GPU by itself.jeremyshaw - Monday, June 22, 2020 - link
So, is most of the power reduction actually coming from cutting the NVLinks down? I don't know about the current NVLink, but the first two versions, Nvidia claimed they didn't have to do external retimers or redrivers, which I can only assume drove up the power consumption, especially with the distances some of the NVLink signals have to travel on the HGX/DGX system vs a NVLink bridge.eastcoast_pete - Monday, June 22, 2020 - link
Thanks Ryan! Also, just to add that cold shower that reality provides so well, I suggest to add the MSRP for those cards. Yes, they'll all run Crysis, but... (And the accelerators don't even have a video signal out)mode_13h - Tuesday, June 23, 2020 - link
Did they even announce a list price? I don't see it in the press release.I'm betting the street price will be in excess of $10k.