NVIDIA Ampere (RTX 30) Architecture Deep-Dive: RT Cores, GDDR6X vs. GDDR6, & More
video description
Date: 2020-09-16
Related videos
Comments and reviews: 10
Ingwie
NVIDIA could've seen this coming. When the ideal scenario is that both triangle and bounded compute units are loaded simultaneously - then they probably never realized that the real world is never ideal :P . To be fair, RTX on the 2000 series was a Public Beta , well, it still is to some degree. It was the first set of GPUs to feature this and only now are we getting the actual second iteration which will thus contain fixes and updates stemming from real world experience. It still feels a bit of a beta gimmick , but it also feels more refined now. I am expecting the next upgrade to outright parallelize the two modules to a point where timing could become flexible enough for no bottleneck to be present in this tiny unit alone. But, as seen in this video and countless examples therein, any small improvement that is made is effectively repeated all over the die, making it a bigger, if not even the biggest improvement.
Really enjoyed this video! I am not a low-level graphics expert - more of a CPU guy myself and because I am visually impaired and lack depth perception AND imagination - but I can still take a lot away from this. If I should upgrade during this generation, I'd go from 2080ti to 3070 and sit out the 4000 launch and THEN buy the highest end one. At this point, RTX has seen a multitude of years in actual use and several of the techs used in CUDA- and Tensor-Cores should be at a point where RTX within them is just _there_ . =)
reply
NVIDIA could've seen this coming. When the ideal scenario is that both triangle and bounded compute units are loaded simultaneously - then they probably never realized that the real world is never ideal :P . To be fair, RTX on the 2000 series was a Public Beta , well, it still is to some degree. It was the first set of GPUs to feature this and only now are we getting the actual second iteration which will thus contain fixes and updates stemming from real world experience. It still feels a bit of a beta gimmick , but it also feels more refined now. I am expecting the next upgrade to outright parallelize the two modules to a point where timing could become flexible enough for no bottleneck to be present in this tiny unit alone. But, as seen in this video and countless examples therein, any small improvement that is made is effectively repeated all over the die, making it a bigger, if not even the biggest improvement.
Really enjoyed this video! I am not a low-level graphics expert - more of a CPU guy myself and because I am visually impaired and lack depth perception AND imagination - but I can still take a lot away from this. If I should upgrade during this generation, I'd go from 2080ti to 3070 and sit out the 4000 launch and THEN buy the highest end one. At this point, RTX has seen a multitude of years in actual use and several of the techs used in CUDA- and Tensor-Cores should be at a point where RTX within them is just _there_ . =)
reply
krr
No offense guys, but you're not really qualified to do a video like this, so you just end up regurgitating what nvidia told you. Nvidia is notorious for deceptive marketing so this is really not the best idea. If you want to do this kind of stuff you should involve someone who worked in semiconductors and undertands these various process nodes, a graphics programmer, someone like AdoredTV whose job it is to cut through nvidia marketing bs. You missed a lot of very interesting detail on Samsung 8nm vs Tsmc 16 and 7, the disadvantages or compromises that came with nvidia's choices (obviously they are not going to talk about those so you need another source of information), etc.
reply
No offense guys, but you're not really qualified to do a video like this, so you just end up regurgitating what nvidia told you. Nvidia is notorious for deceptive marketing so this is really not the best idea. If you want to do this kind of stuff you should involve someone who worked in semiconductors and undertands these various process nodes, a graphics programmer, someone like AdoredTV whose job it is to cut through nvidia marketing bs. You missed a lot of very interesting detail on Samsung 8nm vs Tsmc 16 and 7, the disadvantages or compromises that came with nvidia's choices (obviously they are not going to talk about those so you need another source of information), etc.
reply
Hero
I'm a technology teacher looking to raise money for a rolling desk and portable monitor, which will help with social distancing when we return to campus next week. If you would like to help support my classroom (which will consist of this because students will be in their rooms all day this year and I'll go to them instead of them coming to the computer lab), please consider helping. My school is a title 1 school, and unfortunately we've been told that we don't have the funds to purchase extras like this this year:
t
reply
I'm a technology teacher looking to raise money for a rolling desk and portable monitor, which will help with social distancing when we return to campus next week. If you would like to help support my classroom (which will consist of this because students will be in their rooms all day this year and I'll go to them instead of them coming to the computer lab), please consider helping. My school is a title 1 school, and unfortunately we've been told that we don't have the funds to purchase extras like this this year:
t
reply
TechyBen
Nobody in the entire world likes motion blur.
I am very happy with my investment money in purchasing the Mod Mat going towards such important statements!
PS, motion blur code could also possibly be used for temporal shift code? To avoid having to re-render every ray traced frame, as in VR temporal shift geometry, you could just adjust the existing rendered frame to a roughly new close position for the new one (on the raytraced layer only) and only render in new unrayraced areas of the frame.
reply
Nobody in the entire world likes motion blur.
I am very happy with my investment money in purchasing the Mod Mat going towards such important statements!
PS, motion blur code could also possibly be used for temporal shift code? To avoid having to re-render every ray traced frame, as in VR temporal shift geometry, you could just adjust the existing rendered frame to a roughly new close position for the new one (on the raytraced layer only) and only render in new unrayraced areas of the frame.
reply
Titanium
I heard that Nvidia worked with Micron on creating GDDR6X. Does that mean that 6X is only available to Nvidia or could this potentially be the memory on the new high end RDNA2 cards?
Buildzoid was concerned about memory bottlenecks for RDNA2 cards and judging be the 3D model that was shown in Fortnite he thinks that the card has GDDR memory and not HBM. Edit: The other way around, he thinks it's HBM and not GDDR ;)
reply
I heard that Nvidia worked with Micron on creating GDDR6X. Does that mean that 6X is only available to Nvidia or could this potentially be the memory on the new high end RDNA2 cards?
Buildzoid was concerned about memory bottlenecks for RDNA2 cards and judging be the 3D model that was shown in Fortnite he thinks that the card has GDDR memory and not HBM. Edit: The other way around, he thinks it's HBM and not GDDR ;)
reply
rberky01
Been hearing that the Founders might be worth buying if you can get your hands on it versus the AIBs this time round because of the big jump in performance between generations, and because Nvidia has put a decent amount of effort in its cooling. What are your thoughts? I m a bit fan of Gigabyte and ASUS but without seeing any tests , scared I might miss out on the limited stock haha
reply
Been hearing that the Founders might be worth buying if you can get your hands on it versus the AIBs this time round because of the big jump in performance between generations, and because Nvidia has put a decent amount of effort in its cooling. What are your thoughts? I m a bit fan of Gigabyte and ASUS but without seeing any tests , scared I might miss out on the limited stock haha
reply
lowzyyy
I personally think that this kind of information is useless because in the end of the day most of this audiance are gamers who want temp, power consumption, benchmarks and some sarcastic jokes.
More than half of this audiance dont give a f about sparse matrix and AI.
People who do this stuff will appreciate Ampere and they already know what to expect
reply
I personally think that this kind of information is useless because in the end of the day most of this audiance are gamers who want temp, power consumption, benchmarks and some sarcastic jokes.
More than half of this audiance dont give a f about sparse matrix and AI.
People who do this stuff will appreciate Ampere and they already know what to expect
reply
Zigmar
Many people think motion blur in modern games still blurs the whole image all the time or causes ghosting. That hasn't been the case for a while now. Yes, the old implementations were bad (and it was better to turn them off), but many games nowadays use proper per-object motion blur. DF also had a video on this subject.
reply
Many people think motion blur in modern games still blurs the whole image all the time or causes ghosting. That hasn't been the case for a while now. Yes, the old implementations were bad (and it was better to turn them off), but many games nowadays use proper per-object motion blur. DF also had a video on this subject.
reply
Joel
Sparse matrix is interesting. if I am understanding this right, it is basically doing what the JPEG format does to image information, hence less pure data, more volume through the same data pipe.
Edit: isn't the GDDR6X voltage stepping the same concept as quantum computing? Using voltage levels instead of bits.
reply
Sparse matrix is interesting. if I am understanding this right, it is basically doing what the JPEG format does to image information, hence less pure data, more volume through the same data pipe.
Edit: isn't the GDDR6X voltage stepping the same concept as quantum computing? Using voltage levels instead of bits.
reply
bean
My pants are getting tensor by the minute.
As a side observation, I'm pretty sure a tensor core/unit is identical in architecture to the typical CUDA cores but they are FP8/16 instead of FP32/64, purely because neural networks don't require such precision. They have no relation to a tensor in physics.
reply
My pants are getting tensor by the minute.
As a side observation, I'm pretty sure a tensor core/unit is identical in architecture to the typical CUDA cores but they are FP8/16 instead of FP32/64, purely because neural networks don't require such precision. They have no relation to a tensor in physics.
reply
Add a review, comment
Other channel videos
