Got a tip for us?
Let us know
Become a MacRumors Supporter for $50/year with no ads, ability to filter front page stories, and private forums.
Nvidia Volta.
- Thread starter AidenShaw
- Start date
- Sort by reaction score
- Status
Clean interior. I like that a lot. No doubt the mMP folks have seen this.Its got 3x DisplayPort.
NVIDIA Advances AI Computing Revolution with New Volta-Based DGX Systems
Meanwhile ATI fans are waiting for news of the new Vega.
![1972_Chevrolet_Vega_Hatcback[1].jpg](https://forums.macrumors.com/data/attachments/614/614381-c0c2302dd0170167bb1e7e8f911b9a91.jpg "1972_Chevrolet_Vega_Hatcback[1].jpg")
Meanwhile ATI fans are waiting for news of the new Vega.
At least for those who prefer low price for modest performance over modest price for good performance.
Or did you mean GTO cars?
![1964-pontiac-gto-specs[1].jpg](https://forums.macrumors.com/data/attachments/614/614382-1d0eb289ee2e5f91a71d98b45a367e3e.jpg "1964-pontiac-gto-specs[1].jpg")
Looking at the pictures, it does look like these are generic PCIe form factor cards, probably with NVlink bridges above the cards. Three Tesla GV100 (no video ports) and one Quadro GV100 (with video ports).
It will probably be some time before the Volta PCIe Teslas and Quadros are available as add-in cards, though.
Last edited:
Winners:
GTX 1030 vs RX 550: possibly RX
GTX 1050 vs RX 560: possibly RX
GTX 1050 Ti vs RX 560 4GiB: ?
GTX 1060 3GiB vs RX 570: RX (not enough VRAM)
GTX 1060 6GiB vs RX 570: GTX
GTX 1070 vs RX 580: GTX
Vega:
GTX 1070 vs RX Vega 1: apparently Vega
GTX 1080 vs RX Vega 2: apparently Vega
GTX 1080 Ti vs RX Vega 3: apparently Vega
GTX Titan Xp vs RX Vega 3: probably GTX
FP64 ignored.
Links? Especially since Vega hasn't been released.
And please do post the comparisons of OpenCL (or whatever has replace OpenCL this week) performance for Vega vs. CUDA on Pascal.
Last edited:
Yeah, at least wait until Vega is announced before making any claims on relative performance.
Next year for Vega?
It looks like there's a fair chance the GV100 will ship before any Vega.
Aiden it appears you do have not done any analysis of the architecture.
Nvidia focused on Deep Learning capabilities of Volta, for very good reason. FP32 performance will be the same as with GP100 chip, clock for clock, core for core.
Kepler - 192 cores/256 KB Registry File Size.
Maxwell - 128 cores/256 KB Registry File size.
Pascal(Consumer) - 128 cores/256 KB Registry File size.
GP100 - 64 cores/256 KB Registry File size.
GV100 - 64 cores/256 KB Registry file size.
This shows you what differs the performance with each generation of GPUs, and why Nvidia was able to achieve increase with Maxwell over Kepler, and why there is no difference in FP32 performance workloads in Pascal(consumer) vs Maxwell, core for core, clock for clock. Consumer Pascal was just Maxwell architecture on 16 nm process.
Nvidia with release of Maxwell have said that 128 cores have 90% of performance of 192 cores from Kepler, and that is true. We have seen what you get with GTX 750 Ti, at the release.
But this is not important. Nvidia has very strong brand appeal. Yesterday they have talked about DL, and everybody jumped to conclusion that it is complete revolution, etc, etc, whereas there are workloads in which the GPU will not be faster core for core, clock for clock than GP100 chip. Of course, it will be 100% sure core for core, clock for clock faster than GP102 chips, for example, but it is not as big revolution as I have initially thought. Its the same field again, I have again been deceived by Nvidia's Reality Distortion Field.
Do not get me wrong, tho. In DL this is by far the most advanced GPU.
But FP32 is the workload that is most important to Apple, right now. In FP32 maybe there is after all a point in AMD's marketing gimmick: "Poor Volta"
.
Nvidia focused on Deep Learning capabilities of Volta, for very good reason. FP32 performance will be the same as with GP100 chip, clock for clock, core for core.
Kepler - 192 cores/256 KB Registry File Size.
Maxwell - 128 cores/256 KB Registry File size.
Pascal(Consumer) - 128 cores/256 KB Registry File size.
GP100 - 64 cores/256 KB Registry File size.
GV100 - 64 cores/256 KB Registry file size.
This shows you what differs the performance with each generation of GPUs, and why Nvidia was able to achieve increase with Maxwell over Kepler, and why there is no difference in FP32 performance workloads in Pascal(consumer) vs Maxwell, core for core, clock for clock. Consumer Pascal was just Maxwell architecture on 16 nm process.
Nvidia with release of Maxwell have said that 128 cores have 90% of performance of 192 cores from Kepler, and that is true. We have seen what you get with GTX 750 Ti, at the release.
But this is not important. Nvidia has very strong brand appeal. Yesterday they have talked about DL, and everybody jumped to conclusion that it is complete revolution, etc, etc, whereas there are workloads in which the GPU will not be faster core for core, clock for clock than GP100 chip. Of course, it will be 100% sure core for core, clock for clock faster than GP102 chips, for example, but it is not as big revolution as I have initially thought. Its the same field again, I have again been deceived by Nvidia's Reality Distortion Field.
Do not get me wrong, tho. In DL this is by far the most advanced GPU.
But FP32 is the workload that is most important to Apple, right now. In FP32 maybe there is after all a point in AMD's marketing gimmick: "Poor Volta"
.
You make it sound like a bad thing.Aiden it appears you do have not done any analysis of the architecture.
Nvidia focused on Deep Learning capabilities of Volta, for very good reason. FP32 performance will be the same as with GP100 chip, clock for clock, core for core.
But FP32 is the workload that is most important to Apple, right now. In FP32 maybe there is after all a point in AMD's marketing gimmick: "Poor Volta"
GV100 has almost 50% better FP32 than GP100. Significantly higher than the Vega leaks. Isn't that important for Apple?
And it's interesting that you're so dismissive of machine learning (while recognizing that Volta is a fantastic ML engine). If Apple wants to get into virtual reality (or self-driving cars), ML is an important part of that.
Adding Tensor cores to CUDA cores changes the game. Unless all you care about is making cat videos.
Last edited:
This is just what you have read in my posts.
All I am saying is that in FP32 the chip is not that revolutionary, compared to GP100. In DL? That is whole different story.
Yes, it has almost 50% higher FP32 performance. But it is achieved through massively increased die size of the GPU, also. We are talking about biggest GPU that has ever been made.
Lets not turn this into AMD vs Nvidia.
50% more FP32 in the same power envelope - isn't that somewhat revolutionary?
What happened to the Koyoot who for a while framed everything as "performance per watt"?
It is not that revolutionary. 12 nm FFN is delivering 25% increased performance in the same thermal envelope, however, if you will look closely at the characteristics of the GPU, we are talking about 300W TDP 5120 CUDA core chip that has 835 mm2 die size. If we will directly compare it to GP100, the efficiency size/size, clock/clock, core/core increased by 15%.
Believe me, I am not spinning this anyway. I am just talking about observable things. Maybe it has 50% higher FP32 performance in similar thermal envelope, but, core for core, clock for clock, it will have the same performance as GP100 chip in FP32.
To which it appears that Nvidia agrees with me:
Average increase of performance around 45% over GP100 with 45% higher compute performance.
Anyways, I wish you putting those GPUs into productive job
.
That's about as believable as an interview with Kellyanne Conway.
You are belittling Volta because its per core performance matches Pascal - while ignoring the fact that its FP32 performance and performance per watt is 50% better than the top Pascal.
All of your "Pascal is just Maxwell" and similar lines now that "Volta is just Pascal" are trying to spin Nvidia's advances as nothing. It's transparent that you're spinning this stuff.
Completely transparent spinning.
It's like a Trump press conference.
Last edited:
Truly one of the more ludicrous comments lately.... Since you're trying to discredit Volta.
If I will say to you: "You are either blinded or stupid to not see what I am writing", will it be true?
You are doing with my words, exactly the same thing, what you are accusing me, for them. Spinning them, the way you want them to be, not what they are.
Is there a merit in my first question?
You can ignore everything what I write about Volta, but it will not change anything about the architecture. Core for core it is Pascal GP100 architecture, with heavily increased core count, die size, and incredibly advanced Deep Learning capabilities. But FP32 performance does not bring anything new. I want only to get the power consumption results, because I do not 100% believe Nvidia marketing numbers. The power consumption of 835mm2 GPU will be extremely high, regardless of everything(if for anyone 300W power consumption is actually high...).
So you deliberately ignore the point which is made in that post, through that graph, and spin the situation to what you want it to be?
Is there a merit in my first question, from this post?
![giphy[1].gif](https://spy.macrumors.com/attachments/giphy-1-gif.699506/ "giphy[1].gif")
Counter point:
http://www.anandtech.com/show/11367...v100-gpu-and-tesla-v100-accelerator-announced
Emphasis mine. But sure, go ahead and pretend this is just a bigger Pascal and that Vega will win.
- Status