Mate, we didn't ask you to give us a comparison video on different GPU's, we simply wanted to know what CUDA cores are and what they do, instead im left slightly confused and annoyed. Thanks.
I completely agree with your comment. He didn't explain what CUDA cores are at all. He compared a few GPUs and told us how Nvidia tweaked their GPUs. Good video, but not really related to it's title.
@@Andreas_Mann CUDA cores are in fact not cores. They are just FPUs (floating point units) which are only a component of what makes a core. GN has some more technical details here: ruclips.net/video/x-N6pjBbyY0/видео.html I might be able to answer questions if you have any left of what CUDA cores are.
They allow a person to compute vast amounts of data in as little as a day or less vs yesterday's technology which used to take several days. In short, CUDA Cores deal in computational power, efficiency and speed.
I'm going to guess you were the guy in class that sat through an entire lecture and then said "This teachers an idiot... He didn't teach me anything..." and then went back to eating cheetos and mountain dew.... He literally answered the question 3 times in this video, if you are too distracted to catch it that's your problem.
You are correct sir. Explaining what something does is not the same as explaining what something actually is. The sun warms the Earth. That does not explain what the Sun is.....son.
The whole key behind CUDA is that you can program your GPU to run high process calculations in parallel. This offers a higher performance over the CPU. This can be useful in LIDAR and camera filtering, or other large data calculations.
That really has nothing to do with CUDA, you can do the same thing with any GPU and always could have. CUDA was just a nice C++ interface that was easy to use.
One caveat with cuda is the thread warp. Each SM can handle many blocks of threads, but there is this issue where there is only one control unit for every 32 cores. This means if there is a branch in the instructions, unless all 32 cores in the warp follow the same execution path, the differing execution branches must be serialized. In other words, decreasing the warp size by adding more control units would contribute greatly to the performance of a gpu even of the number of SMs and cores remained the same.
he duz neigh unnerstand,by writing sums you wit out crumbs too do too that you imagined as combe fly pi ano two mix linear sum popular based old cold miff thats meant thinking mans sum,bit yer dont do it,yer try spell agaimst machine,crazy
It doesn't really matter, all you need to know is what the best performing card for your budget is. How it works isn't that important and even if you cannot be bothered reading a single article, there are a million people who will be happy to give you advice on which card to buy....
just what im looking for - a YT channel with a very intelligent man who is very astute and knowledgeable on the technical aspects of the topics as i am - looking forward to more videos - thanks
So Greg. Ever taught of being an University Computer Engineer Teacher? Because the things i learn on your channel i somewhat wish i could learn at my universitity but all they do is hand paperwork and be quiet. The way you explained the cuda cores are exellent. Good Job!
05:44 According to the Nvidia's GTX1080 whitepaper, the GP104 used in the GTX1080 actually has 128 CUDA cores per SM. It's the GP100 (eg. as used in Tesla cards) which has 64 cores per SM.
980ti does not have 8.1 billion transistors, titan X (2015) does. 980ti is technically the faulty titan X dies which have minor flaws in them, so its just easier and cheaper to turn off a few malfunctioning cores and sell it off as 980ti rather than throw them off as garbage. Both titan x and 980 ti have same die size and same transistors which should prove my point.
GM200 is simply a core architecture, just like Broadwell cores architecture, 980ti uses 2816 of those gm200 cores while titan uses 3072 of those. 980ti sure does have all the 8 billion transistors of titan x but a few of them have been disabled to accommodate for those missing 256 cores. The lithography process isn't perfect, and especially not if the die size is as big as 610mm2. 980ti dies are simply the titan x which had minor faults in them You can observe the same phenomenon in intel core i7, core i5 as well
Having transistors and having working transistors are two different arguments. I don't think you are in a correct mindset to understand it right now, probably read a bit more on the topic of wafer yield and die yield, and Optical Proximity Correction(OPC) , you'll get it though. BTW awesome video, just thought to address that one calculation. Sorry to bother, Peace! :) Thanks for "congratulations"
Dunno why he is so rude. BTW you are right. A titan x should approximatly have 26.367.187 tpc, so could look like the 1080 does have more tpc. Actually it does since they doubled the cache for cuda core and added some components to help improve async compute
I would question one thing said there. The name "CUDA" may have been coined on 2005 or so, I don't entirely remember, but when I was studying computers at a university level, in 1995 ~ 2000, nVidea was very much talking about the concept, and making GPUs that would work on a desktop, but where the exact same chip would be a compute unit to go in a supercomputer, (because multi core / multi thread / extreme parallel processing) and I'm almost sure I heard the term CUDA right back then. The term I don't quite know, but the concept was most certainly there.
well, here are something I want to mention after watching the video: 1. Transistors on the GPU are not for just the shader processors (cuda core or stream processor). There are also ROPs which have much higher transistor count than single shader, as well as the L2 cache. 2. The GM200 chip contains 8.1B transistors, which are fully utilized in Titan X. But 980ti only uses 2816 shaders among 3072 of them. As a result, the total active cuda cores might have 7.4B transistor count. While the 384 bit memory controller takes some of them, I don't think Nvidia has any essential modification in Pascal compared with Maxwell. 3. Nvidia has changed the way cuda cores are counted since Kepler. GTX 580 (Fermi arch) has only 512 Cuda cores run at twice as the core clock speed(1.5GHz on 40nm fab). And they changed this in Kepler.
I have a somewhat off-topic question... how did you learn/how do you know so much IT technical information? (very clear in your presentation & other videos). Just a lot of personal research? College? And what recommendation would you have for someone to get on your "level"? Ha ha.
I live your videos, a major factor is that is that you actually use 1080p60. by far for me the best science and technology channel. Proud to be here from your sciencey talks!
4:40 um akshully, this is not a good comparison because the reason an f1 car’s 1.6L engine revs up to 18,000rpm is because the engines are literally worth millions of dollars and are expected to last only a few races. They idle at an insane 9,000rpm (about 3,000 more than the redline of me mum’s civic) and if you tried to drive one on the road like a viper you’d probably explode the engine because you stood still for too long at a stoplight. You might be right about piston size vs rpm, it is likely it’s a factor, but I’ve never heard of it. Usually the limiting factor in rpm is valve float, reliability and overall stability/engineering of the engine. Valve float is where the driveshaft (and therefore the cams that push open the exhaust and intake valves, which are held closed by springs) is spinning so fast that the springs can’t push the valves closed as fast as the cam curves away (causing the valve to “float”). This means the valve is still partly open for a millisecond when it’s not meant to, leading to the piston ramming itself into the valve, damaging the valve and possibly the piston beyond repair. Keep in mind this happens at speeds greater than 6000rpm (100 revolutions per second or 1 rev every 10ms). It takes some expensive engineering to get any engine to not tear itself to shreds with its own force at high revs. Balance, timing and lubricant have to be perfect and this compounds in complexity more with the number and position of pistons than with displacement itself. One last “um akshully”, you put an inline 4 above the 1.6L F1 when F1s all have turbocharged V6s by regulation. I do highly recommend you look into the insane engineering of F1 cars. Those things are so incredibly smart and wasteful at the same time. It’s like “how fast could we possibly make this go given basically unlimited resources and a some of the smartest engineers in the world”.
You're hilarious man! The vids with you and McLovin I got to see a whole different side of you. You seem easygoing and chill. Always informative and you don't seem arrogant when you help people. Thanks bruh.
There are a couple of things you could mention. The difference in transistors is to do with the cores and memory controllers too. So a cuda core between maxwell and pascal is the same but the SM core config isnt not only in the number of cores per SM but also in other units that are present. Theres also the ROP and TMUs that also account for the transistors and are evenly distributed to every SM.
Was the 1080 the first to break to 2.0ghz barrier? Excluding crazy OCing rigs with liquid nitrogen cooling or whatever. I've always been so used to cards being 900-1300 at stock speed, that I was pretty blown away during the livestream when it got clocked passed 2.0
Your are amazing!! Just sat here watching your videos for hours before I had to login to sub to your channel. Will tell my family and friends abt this :) They love this kinda stuff too.
Before I even watch the vid, I just want to say that I love your videos. All of your videos are very informative, and still easy to understand. Keep up the good work!
I had a question , you are a brilliant super nerd and I admire you , one of the true RUclips greats. Question : Someone said that the CUDA cores nowadays are similar to Sony Cell processor SPE's but are more parallel and and easier to develop for and numerous, but the concept is similar and more evolved ? . However there are now thousands of CUDA cores in GPUs, as the IBM Cell only had 8 SPE's , but they ran at a very very fast 3.2GHZ .So Cuda is limited to the GPUs Speed which are currently under 2 ghz. Had IBM Sony commercially succeeded in Cell and the next Playstation were to say have 128 or 256 SPE's running at 4.0GHZ would this had been able to match current GPU power as well as work well with a CUDA based CO-GPU setup similar to the PS3. With a 300+ GBper sec fast BUS connecting the two? Would this be a monster gaming system or a flop? Because the exclusive games on PS3 we phenomenal .
i was thinking you should talk about what makes a pc really a pc and how it varies from a laptop or a workstation, phones and tablet please that would be great
I was just thinking, now we have stream processors and Cuda cores to look at when checking out relative performance. Back then it was pixel pipelines. Can you do a vid explaining the differences?
I think you're massively playing down the importance of memory speeds. DDR5X can clock much higher speeds than DDR5 which come with 980ti which overclocked normally tops out at 8gbps where as the 1080 OC reaches 11gbps
btw you kinda off on the f-1 engine reasoning.... the engine pistons are smaller and short stroke to lower the rotational and inertial mass of the moving parts and why they use lightweight materials. if the moving parts had no mass this wouldnt even be a consideration anymore as stress issues would be eliminated.
As a critique, I would have like to learn more about the inner workings if the cuda core. Similar to some of the videos out there about the x86 and 680x0 processors and how they throughput data and instructions. I am trying to understand why these devices are so well suited for graphics and large array calcs.
5:50 THIS IS INCORRECT!!! Both Maxwell and Pascal use 128 CUDA per SM! The only EXCEPTION is the Pascal Tesla! As that Compute Card has 64 FP32 Single Precision CUDA, and then a mixture of FP16 Half Precision and FP64 Double Precision CUDA per SM that adds up to DRUMROLL 128 CUDA per SM! All GPU Pascal cards have the EXACT SAME 128 CUDA per SM! This includes the NEW Titan X!
the problem is the ridiculously high prices we are made to pay for new gpus every 3 months with only minute incremential base clock frequency increases, 3-400 extra cuda cores and the "zomg new architecture" sellpoint nvidia keeps throwing on the market which still makes 3840x2160p@60fps on latest gen titles at max settings still hard to reach unless u have a $1000 to spend on 1080tis and titan x sli....more competition needed in the gpu market to lower prices
Could you make an explanation video about Quadro and GTX? For CG artists there not too much help around to decide which GPUs are most effective, for different tasks.
I guess u will have to explain how cuda and stream processor differ (nvidia switched from stream processors that already existed long before cuda). but on the matter of 980ti vs 1080 I didn't notice a very large fps difference essentially reducing cuda cores in 1080 made it almost equal in performance to 980ti (not everyone wants to overclock their gpu) which actually makes sense considering the flagship of a 10xx series should always be the one on top of tha last flagship.
Nice Marshall 4X cabinet...what head do u use and what is the black guitar swinging from the ceiling. Btw the way Ive read a lot about the performance with the 970 vs. the 1060 and the 970 (which I have) stays right with it...have a great day
Hey SS, i was planning to upgrade my pc: specs below i5 4570, asus b85m-g board, 1tb hdd, 8gb ram, amd hd 7870, 43w800c sony 1080p 3d tv. and i planned to buy rx 480 with 290$ in hand but due to overpricing(430$ in india) and power issue i afraid to buy. and later i just bought asus gtx 970 turbo for 290$ on jul 01 2016. later on jul 19 gtx 1060 released. before buying 970 i checked the specs between 970 and 1060 970 - 1664 cuda cores with 256 bit bus / 224.4gb bandwidth 1060 - 1280 cuda cores with 192 bit bus / 192.2gb bandwidth but in gaming performance 1060 beats 970 in all games.(now i learned less transistors more clock speed) my friend asks me the 970 for 270$. My question is, can i give the 270$ gtx 970 and go for the 370$ gtx 1060? Is it worth it? Pls check my specs once again before giving a suggestion.
Yeah, go for it. The 1060 is much better than the 970, especially in the VRAM department. As you might know, the 1060 has 6 GB vram, wheras the 970 only has 3.5 GB. This will help a lot in modern games. Sell the 970 to your friend, if you already have an interested "customer", it makes the whole thing much easier.
I have i5 4440 and B85M + 8GB RAM as well (with a 750Ti) but I found that getting a SSD (like Samsung EVO) is the best bang for the buck upgrade for a long amount of time (if you don't have one already) - the latest gen GPUs are overpriced like hell in India anyway, and the prices have increased a lot ever since I bought my PC. And I don't know about you but the PPI for a 43" 1080p is too low for me to use for normal stuff (like browsing)
Thanks for the video dude. I was so tried of people saying that the GTX 980ti was more powerful than the GTX 1070/1080. Even though they have 2 more GBs in the GPU, newer architecture and also more Teraflops.
Really interesting. I knew some of the stuff that you talked about but there are sure things that I did not know before and I am more than happy to learn. Thanks for sharing with us :)
I think CUDA are more like ARM. They are the architecture who design the individual core (vertex shader & fragment shader) and not physically manufacture them (That thing goes to TSMC or Samsung FinFET etc). They have their own sets of instruction and compatibility within their design. Btw this is just my assumption. I'm not familiar with gpu industry though.
Well, since you talked about CUDA cores, you should also do a video about STREAM processors from AMD so that people can understand the difference between them. It is also very useful when trying to compare green vs red. Cheers!
So if Nvidia can not actually do async compute do you think they will still be able to perform with DX12? If they cannot support it at a hardware level then wouldn't their cards eventually become obsolete?
Wow, good video and good job explaining! The only thing I have to criticize is this weird comparison with engines pistons :D... This really didn't made it any easier to me to understand but rather more abstract and difficult... Anyways, other people maybe see it different...
Well informed. I have a msi 8400 gs which only has 8 cuda cores and I have it as a dedicated physx card. seems pointless but without it the system crawls. paired with a GeForce 630 gts.
Mate, we didn't ask you to give us a comparison video on different GPU's, we simply wanted to know what CUDA cores are and what they do, instead im left slightly confused and annoyed. Thanks.
QueueDoor
I completely agree with your comment.
He didn't explain what CUDA cores are at all.
He compared a few GPUs and told us how Nvidia tweaked their GPUs.
Good video, but not really related to it's title.
Exacty!
@@Andreas_Mann CUDA cores are in fact not cores. They are just FPUs (floating point units) which are only a component of what makes a core.
GN has some more technical details here: ruclips.net/video/x-N6pjBbyY0/видео.html
I might be able to answer questions if you have any left of what CUDA cores are.
They allow a person to compute vast amounts of data in as little as a day or less vs yesterday's technology which used to take several days. In short, CUDA Cores deal in computational power, efficiency and speed.
Nice but what are cuda cores?
Exactly :D
Kinda feel like that
I was left with the same question
A shit ton of cores like in your CPU.
P
This video completely fails to answer the question that is its title and rapidly devolves into a relative performance discussion.
yes, you are right
Cuda cores are a general marketing thing. The thing that always matters most is how close you can get the cache to the processors. Aka die size.
I'm going to guess you were the guy in class that sat through an entire lecture and then said "This teachers an idiot... He didn't teach me anything..." and then went back to eating cheetos and mountain dew.... He literally answered the question 3 times in this video, if you are too distracted to catch it that's your problem.
cuda is parallel calculations chips, every chip can calculate 100 strings per secod, good to draw perspective or try passwords for example
You are correct sir. Explaining what something does is not the same as explaining what something actually is. The sun warms the Earth. That does not explain what the Sun is.....son.
The whole key behind CUDA is that you can program your GPU to run high process calculations in parallel. This offers a higher performance over the CPU. This can be useful in LIDAR and camera filtering, or other large data calculations.
Thank you so much
That really has nothing to do with CUDA, you can do the same thing with any GPU and always could have. CUDA was just a nice C++ interface that was easy to use.
One caveat with cuda is the thread warp. Each SM can handle many blocks of threads, but there is this issue where there is only one control unit for every 32 cores. This means if there is a branch in the instructions, unless all 32 cores in the warp follow the same execution path, the differing execution branches must be serialized. In other words, decreasing the warp size by adding more control units would contribute greatly to the performance of a gpu even of the number of SMs and cores remained the same.
he duz neigh unnerstand,by writing sums you wit out crumbs too do too that you imagined as combe fly pi ano two mix linear sum popular based old cold miff thats meant thinking mans sum,bit yer dont do it,yer try spell agaimst machine,crazy
I commonly watch these without largely understanding most of it because I'm dumb brained. I continue to watch in hopes something might stick somehow.
youll eventually understand it, or just look it up
same xD
rewatch til you understand
Ashish Parmar
Did you only read the first sentence of my two?
It doesn't really matter, all you need to know is what the best performing card for your budget is. How it works isn't that important and even if you cannot be bothered reading a single article, there are a million people who will be happy to give you advice on which card to buy....
I really liked that RPM to Hz analogy; Nice
- Suggest future topics: GO!
Space is pretty cool, do something on black holes?
put all the gtx 1070 aftermarket coolers against each other
Mark Merwitzer Nah black holes are more interesting
What is love?
I've never done a 'love' before.
3Dmark scores
In my language CUDA means miracles :E
What the language it is?
Polish, translate.google.pl/#pl/en/cud
CUDA is plurar from cud -> 1 cud, multiple cuda ;D
Piona :D
:D
CUDA is english, and it doesn't mean anything. (abbr. is not a meaning)
just what im looking for - a YT channel with a very intelligent man who is very astute and knowledgeable on the technical aspects of the topics as i am - looking forward to more videos - thanks
So Greg. Ever taught of being an University Computer Engineer Teacher? Because the things i learn on your channel i somewhat wish i could learn at my universitity but all they do is hand paperwork and be quiet. The way you explained the cuda cores are exellent. Good Job!
Content, delivery, and style just keeps getting better. To think, I subscribed when there were less than 25k of us. Triple that now. Awesome.
why this guy only has 80k subs????
He's relatively newer :)
cuz youtube suks
227k now
05:44 According to the Nvidia's GTX1080 whitepaper, the GP104 used in the GTX1080 actually has 128 CUDA cores per SM. It's the GP100 (eg. as used in Tesla cards) which has 64 cores per SM.
Yes, this has been previously noted. My source was incorrect.
980ti does not have 8.1 billion transistors, titan X (2015) does. 980ti is technically the faulty titan X dies which have minor flaws in them, so its just easier and cheaper to turn off a few malfunctioning cores and sell it off as 980ti rather than throw them off as garbage. Both titan x and 980 ti have same die size and same transistors which should prove my point.
It's a GM200 GPU. 8.1 billion transistors. Not even an argument.
GM200 is simply a core architecture, just like Broadwell cores architecture, 980ti uses 2816 of those gm200 cores while titan uses 3072 of those. 980ti sure does have all the 8 billion transistors of titan x but a few of them have been disabled to accommodate for those missing 256 cores.
The lithography process isn't perfect, and especially not if the die size is as big as 610mm2. 980ti dies are simply the titan x which had minor faults in them
You can observe the same phenomenon in intel core i7, core i5 as well
I know exactly what GM200 is. Read your first sentence in your original comment. You just contradicted yourself. Congratulations.
Having transistors and having working transistors are two different arguments. I don't think you are in a correct mindset to understand it right now, probably read a bit more on the topic of wafer yield and die yield, and Optical Proximity Correction(OPC) , you'll get it though. BTW awesome video, just thought to address that one calculation. Sorry to bother, Peace! :)
Thanks for "congratulations"
Dunno why he is so rude. BTW you are right. A titan x should approximatly have 26.367.187 tpc, so could look like the 1080 does have more tpc. Actually it does since they doubled the cache for cuda core and added some components to help improve async compute
I would question one thing said there. The name "CUDA" may have been coined on 2005 or so, I don't entirely remember, but when I was studying computers at a university level, in 1995 ~ 2000, nVidea was very much talking about the concept, and making GPUs that would work on a desktop, but where the exact same chip would be a compute unit to go in a supercomputer, (because multi core / multi thread / extreme parallel processing) and I'm almost sure I heard the term CUDA right back then. The term I don't quite know, but the concept was most certainly there.
last time i've been this early greg was still making mclovins pc
1080ti is insane considering it has a loooot more cuda cores than the 1080 for a slightly higher price
3,584 vs 2,560
well, here are something I want to mention after watching the video:
1. Transistors on the GPU are not for just the shader processors (cuda core or stream processor). There are also ROPs which have much higher transistor count than single shader, as well as the L2 cache.
2. The GM200 chip contains 8.1B transistors, which are fully utilized in Titan X. But 980ti only uses 2816 shaders among 3072 of them. As a result, the total active cuda cores might have 7.4B transistor count. While the 384 bit memory controller takes some of them, I don't think Nvidia has any essential modification in Pascal compared with Maxwell.
3. Nvidia has changed the way cuda cores are counted since Kepler. GTX 580 (Fermi arch) has only 512 Cuda cores run at twice as the core clock speed(1.5GHz on 40nm fab). And they changed this in Kepler.
How do you only have 80,000 subs?
He's a newer tech RUclipsr, but quickly growing.
+Justin van der Werf Haha yeah this guy should be hired by LTT
No clue, but he deserves more exposure.
He's going to petroleum engineering, he can't lol.
According to his about page he joined 29 Jun 2015. 100k subs in this timeframe is actually pretty incredible
I have a somewhat off-topic question... how did you learn/how do you know so much IT technical information? (very clear in your presentation & other videos). Just a lot of personal research? College? And what recommendation would you have for someone to get on your "level"? Ha ha.
I just searched to know what cuda and tensor cores are, and I found a pretty cool channel. Thanks Greg,
this guy is perhaps the most sane and objective hardware dude on the tube
I live your videos, a major factor is that is that you actually use 1080p60. by far for me the best science and technology channel. Proud to be here from your sciencey talks!
Honored to have you on board!
+Science Studio pleased to have your reply
Ive been following tech youtubers for a while now and i have to say ur videos are much more interesting and i learn more from them keep it up
This channel will reach tremendous success if you play your cards right tbh
4:40 um akshully, this is not a good comparison because the reason an f1 car’s 1.6L engine revs up to 18,000rpm is because the engines are literally worth millions of dollars and are expected to last only a few races. They idle at an insane 9,000rpm (about 3,000 more than the redline of me mum’s civic) and if you tried to drive one on the road like a viper you’d probably explode the engine because you stood still for too long at a stoplight.
You might be right about piston size vs rpm, it is likely it’s a factor, but I’ve never heard of it. Usually the limiting factor in rpm is valve float, reliability and overall stability/engineering of the engine. Valve float is where the driveshaft (and therefore the cams that push open the exhaust and intake valves, which are held closed by springs) is spinning so fast that the springs can’t push the valves closed as fast as the cam curves away (causing the valve to “float”). This means the valve is still partly open for a millisecond when it’s not meant to, leading to the piston ramming itself into the valve, damaging the valve and possibly the piston beyond repair. Keep in mind this happens at speeds greater than 6000rpm (100 revolutions per second or 1 rev every 10ms).
It takes some expensive engineering to get any engine to not tear itself to shreds with its own force at high revs. Balance, timing and lubricant have to be perfect and this compounds in complexity more with the number and position of pistons than with displacement itself.
One last “um akshully”, you put an inline 4 above the 1.6L F1 when F1s all have turbocharged V6s by regulation.
I do highly recommend you look into the insane engineering of F1 cars. Those things are so incredibly smart and wasteful at the same time. It’s like “how fast could we possibly make this go given basically unlimited resources and a some of the smartest engineers in the world”.
You're hilarious man! The vids with you and McLovin I got to see a whole different side of you. You seem easygoing and chill. Always informative and you don't seem arrogant when you help people. Thanks bruh.
Very informative, and well presented. That dude's leather jacket is so nice.
Gotta say that this channel is quickly becoming one of my favorites, nice work.
I appreciate it!
I wonder if "Super Cuda Cores" will be a thing.. Adding more features and such to the Original Cuda Core.
Nice vid greg, but like you said pleeassee do a crash course for Asynchronous compute, really would love some in depth explanations to that.
GREAT video. First one that covers this clearly IMO. I had a hard time finding info on what these were.
My God, your one smart cookie. Great informative video. Sadly I'm still as thick as two short planks lol. Keep up your great work. Regards.
NEED A VID ON BEST CARD FOR CONVERTING MOVIES..
There are a couple of things you could mention. The difference in transistors is to do with the cores and memory controllers too. So a cuda core between maxwell and pascal is the same but the SM core config isnt not only in the number of cores per SM but also in other units that are present. Theres also the ROP and TMUs that also account for the transistors and are evenly distributed to every SM.
Was the 1080 the first to break to 2.0ghz barrier? Excluding crazy OCing rigs with liquid nitrogen cooling or whatever. I've always been so used to cards being 900-1300 at stock speed, that I was pretty blown away during the livestream when it got clocked passed 2.0
Your are amazing!! Just sat here watching your videos for hours before I had to login to sub to your channel. Will tell my family and friends abt this :) They love this kinda stuff too.
Glad you're enjoying the channel! I appreciate it.
I love learning with you. Keep up the good work!
I'm happy to see you're channel grow I cant remember when I subscribed on my other account but it was around 40k. You're one of my favorite youtubers,
Before I even watch the vid, I just want to say that I love your videos. All of your videos are very informative, and still easy to understand. Keep up the good work!
4 years later, but damn was this still informative & relevant
I've learned so much from this channel.
I had a question , you are a brilliant super nerd and I admire you , one of the true RUclips greats.
Question : Someone said that the CUDA cores nowadays are similar to Sony Cell processor SPE's but are more parallel and and easier to develop for and numerous, but the concept is similar and more evolved ? . However there are now thousands of CUDA cores in GPUs, as the IBM Cell only had 8 SPE's , but they ran at a very very fast 3.2GHZ .So Cuda is limited to the GPUs Speed which are currently under 2 ghz. Had IBM Sony commercially succeeded in Cell and the next Playstation were to say have 128 or 256 SPE's running at 4.0GHZ would this had been able to match current GPU power as well as work well with a CUDA based CO-GPU setup similar to the PS3. With a 300+ GBper sec fast BUS connecting the two? Would this be a monster gaming system or a flop? Because the exclusive games on PS3 we phenomenal .
Great video and on point. This channel is so much better than those annoying hacks at Linus tech tips
i was thinking you should talk about what makes a pc really a pc and how it varies from a laptop or a workstation, phones and tablet please that would be great
I understand you wanted to get to the point. And you did. But it was a little too fast for me. Well done video. Now to watch again.
Seems like you have adjusted your lighting......Thank Gosh....now it's like a thousand times better!!!! Nice vids mannn!
No, it's a new camera.
Awesome
This appeared top searches.... There are other primary informations I need to know before this video.
Man!!! You go too deep into a topic, Exactly, this is what I came for!! Exactly the thing i subscribed for!!! Keep it up man 👍
geez I just wanted to know what a Cuda core is
CUDA cores are that thing that make my game go pew pew.
Hey, what is that game that looks like Starcraft? Looks nice.
Ashes of the Singularity
Hope you still need an answer :-)
@@igorthelight i think he got it figured out after 2 years hahah.
@@metalvideos1961 I hope so :-)
@@metalvideos1961 lol
We asked what is CUDA cores, he answer why is CUDA cores
I was just thinking, now we have stream processors and Cuda cores to look at when checking out relative performance. Back then it was pixel pipelines. Can you do a vid explaining the differences?
Super intelligent Explanation. Im smart but I enjoy people that humble me. Thank you very much.
The funny thing is, pascal cores are the same as maxwell. They have the same ipc. Pascal just has high clocks.
I liked the piston analogy for the transistor, good job.
Are you talking about the area of the piston (bore) or the number of pistons?
both
you deserve waaay more subs, keep up the great videos!
your videos are really, really great. succinct, very informative, and very well produced. Great job, Greg.
is this ultimate sarcasm
Science studio I LOVE VIDEOS LIKE THESE PLEASE MAKE MORE I LOVE COMPUTER HARDWARE ENGINEERING. Maybe do something on clock speed and ipc?
I think you're massively playing down the importance of memory speeds. DDR5X can clock much higher speeds than DDR5 which come with 980ti which overclocked normally tops out at 8gbps where as the 1080 OC reaches 11gbps
Tip: Make a video on how to interpret the amount of "Teraflops", many people have no idea what this means.
There needs to be a single cuda core benchmarker.
btw you kinda off on the f-1 engine reasoning.... the engine pistons are smaller and short stroke to lower the rotational and inertial mass of the moving parts and why they use lightweight materials. if the moving parts had no mass this wouldnt even be a consideration anymore as stress issues would be eliminated.
As a critique, I would have like to learn more about the inner workings if the cuda core. Similar to some of the videos out there about the x86 and 680x0 processors and how they throughput data and instructions. I am trying to understand why these devices are so well suited for graphics and large array calcs.
omg thanks for making this video everything makes sense now
Love this stuff, I'm so glad I stumbled onto this channel. Great vid!
You should do a crash course on how they make the 14nm architecture for example on how they make them
5:50 THIS IS INCORRECT!!!
Both Maxwell and Pascal use 128 CUDA per SM! The only EXCEPTION is the Pascal Tesla! As that Compute Card has 64 FP32 Single Precision CUDA, and then a mixture of FP16 Half Precision and FP64 Double Precision CUDA per SM that adds up to DRUMROLL 128 CUDA per SM!
All GPU Pascal cards have the EXACT SAME 128 CUDA per SM! This includes the NEW Titan X!
the problem is the ridiculously high prices we are made to pay for new gpus every 3 months with only minute incremential base clock frequency increases, 3-400 extra cuda cores and the "zomg new architecture" sellpoint nvidia keeps throwing on the market which still makes 3840x2160p@60fps on latest gen titles at max settings still hard to reach unless u have a $1000 to spend on 1080tis and titan x sli....more competition needed in the gpu market to lower prices
"CUDA stands for..." lol nope, you won't actually find that anywhere on nvidia's website.
I found that odd.
Could you make an explanation video about Quadro and GTX? For CG artists there not too much help around to decide which GPUs are most effective, for different tasks.
can't wait for that build with bowe
I guess u will have to explain how cuda and stream processor differ (nvidia switched from stream processors that already existed long before cuda). but on the matter of 980ti vs 1080 I didn't notice a very large fps difference essentially reducing cuda cores in 1080 made it almost equal in performance to 980ti (not everyone wants to overclock their gpu) which actually makes sense considering the flagship of a 10xx series should always be the one on top of tha last flagship.
Great video as always! Really looking forward to that LGA 775 build because I have a Q6700 and a P5Q SE2 motherboard as a secondary rig :P
Nice Marshall 4X cabinet...what head do u use and what is the black guitar swinging from the ceiling. Btw the way Ive read a lot about the performance with the 970 vs. the 1060 and the 970 (which I have) stays right with it...have a great day
I finally know what cuda cores are thank you
not from this video
Great video! Every single piece of information was spot on :)
Dude, best explanation ever, thank you!
Appreciate this channel and the work you put into it! Thanks!
Hey SS, i was planning to upgrade my pc: specs below
i5 4570, asus b85m-g board, 1tb hdd, 8gb ram, amd hd 7870, 43w800c sony 1080p 3d tv.
and i planned to buy rx 480 with 290$ in hand but due to overpricing(430$ in india) and power issue i afraid to buy. and later i just bought asus gtx 970 turbo for 290$ on jul 01 2016. later on jul 19 gtx 1060 released. before buying 970 i checked the specs between 970 and 1060
970 - 1664 cuda cores with 256 bit bus / 224.4gb bandwidth
1060 - 1280 cuda cores with 192 bit bus / 192.2gb bandwidth
but in gaming performance 1060 beats 970 in all games.(now i learned less transistors more clock speed)
my friend asks me the 970 for 270$.
My question is, can i give the 270$ gtx 970 and go for the 370$ gtx 1060? Is it worth it?
Pls check my specs once again before giving a suggestion.
Yeah, go for it. The 1060 is much better than the 970, especially in the VRAM department.
As you might know, the 1060 has 6 GB vram, wheras the 970 only has 3.5 GB.
This will help a lot in modern games. Sell the 970 to your friend, if you already have an interested "customer", it makes the whole thing much easier.
I have i5 4440 and B85M + 8GB RAM as well (with a 750Ti) but I found that getting a SSD (like Samsung EVO) is the best bang for the buck upgrade for a long amount of time (if you don't have one already) - the latest gen GPUs are overpriced like hell in India anyway, and the prices have increased a lot ever since I bought my PC.
And I don't know about you but the PPI for a 43" 1080p is too low for me to use for normal stuff (like browsing)
Thanks for the video dude. I was so tried of people saying that the GTX 980ti was more powerful than the GTX 1070/1080. Even though they have 2 more GBs in the GPU, newer architecture and also more Teraflops.
Love these Greg. Great video as always.
nice report,maybe overviews without linear time transpect as omega,sigma,unit developments excel built by pascal because,etc etc,unknown delta??
all this stuff about more cuda cores doesn't help me understand what is a cuda core.
what is a core in this context anyway?
Really interesting. I knew some of the stuff that you talked about but there are sure things that I did not know before and I am more than happy to learn. Thanks for sharing with us :)
I think CUDA are more like ARM. They are the architecture who design the individual core (vertex shader & fragment shader) and not physically manufacture them (That thing goes to TSMC or Samsung FinFET etc). They have their own sets of instruction and compatibility within their design. Btw this is just my assumption. I'm not familiar with gpu industry though.
Well, since you talked about CUDA cores, you should also do a video about STREAM processors from AMD so that people can understand the difference between them. It is also very useful when trying to compare green vs red. Cheers!
my man. I see you have a guitar stack and nice guitar. you should do video about harmonics just maybe?
So if Nvidia can not actually do async compute do you think they will still be able to perform with DX12? If they cannot support it at a hardware level then wouldn't their cards eventually become obsolete?
hey science studio what lights are you using in your case? i have the same case but black and the lights make it look great
This was one of the things I always wondered about, thanks for explaining it!
Why are there dislikes on this video WHY?
A great video for an intriguing topic. Good work!
Wow a video thats not uploaded at like 1 am
It's always uploaded at 1AM somewhere on Earth.
+Science Studio Rekt
EST
+Science Studio it's LITEALLY 1 am here, ironic isn't it ?
Got'em
Great video dude!
Wow, good video and good job explaining!
The only thing I have to criticize is this weird comparison with engines pistons :D... This really didn't made it any easier to me to understand but rather more abstract and difficult... Anyways, other people maybe see it different...
Well informed. I have a msi 8400 gs which only has 8 cuda cores and I have it as a dedicated physx card. seems pointless but without it the system crawls. paired with a GeForce 630 gts.
thanks! needed someone to clear this up 😊