Wednesday, September 2nd 2020

NVIDIA Announces GeForce Ampere RTX 3000 Series Graphics Cards: Over 10000 CUDA Cores

Update 3 by

Sep 1st, 2020 12:04 Updated: Sep 2nd, 2020 10:05 Discuss (502 Comments)

NVIDIA just announced its new generation GeForce "Ampere" graphics card series. The company is taking a top-to-down approach with this generation, much like "Turing," by launching its two top-end products, the GeForce RTX 3090 24 GB, and the GeForce RTX 3080 10 GB graphics cards. Both cards are based on the 8 nm "GA102" silicon. Join us as we live blog the pre-recorded stream by NVIDIA, hosted by CEO Jen-Hsun Huang.

Update 16:04 UTC: Fortnite gets RTX support. NVIDIA demoed an upcoming update to Fortnite that adds DLSS 2.0, ambient occlusion, and ray-traced shadows and reflections. Coming soon.

Update 16:06 UTC: NVIDIA Reflex technology works to reduce e-sports game latency. Without elaborating, NVIDIA spoke of a feature that works to reduce input and display latencies "by up to 50%". The first supported games will be Valorant, Apex Legends, Call of Duty Warzone, Destiny 2 and Fortnite—in September.

Update 16:07 UTC: Announcing NVIDIA G-SYNC eSports Displays—a 360 Hz IPS dual-driver panel that launches through various monitor partners in this fall. The display has a built-in NVIDIA Reflex precision latency analyzer.

Update 16:07 UTC: NVIDIA Broadcast is a brand new app available in September that is a turnkey solution to enhance video and audio streaming taking advantage of the AI capabilities of GeForce RTX. It makes it easy to filter and improve your video, add AI-based backgrounds (static or animated), and builds on RTX Voice to filter out background noise from audio.

Update 16:10 UTC: Ansel evolves into Omniverse Machinima, an asset exchange that helps independent content creators to use game assets to create movies. Think fan-fiction Star Trek episodes using Star Trek Online assets. Beta in October.

Update 16:15 UTC: Updates to the AI tensor cores and RT cores. In addition to more numbers of RT- and tensor cores, the 2nd generation RT cores and 3rd generation tensor cores offer higher IPC. Making ray-tracing have as little performance impact as possible appears to be an engineering goal with Ampere.

Update 16:18 UTC: Ampere 2nd Gen RTX technology. Traditional shaders are up by 270%, raytracing units are 1.7x faster and the tensor cores bring a 2.7x speedup.

Update 16:19 UTC: Here it is! Samsung 8 nm and Micron GDDR6X memory. The announcement of Samsung and 8 nm came out of nowhere, as we were widely expecting TSMC 7 nm. Apparently NVIDIA will use Samsung for its Ampere client-graphics silicon, and TSMC for lower volume A100 professional-level scalar processors.

Update 16:20 UTC: Ampere has almost twice the performance per Watt compared to Turing!

Update 16:21 UTC: Marbles 2nd Gen demo is jaw-dropping! NVIDIA demonstrated it at 1440p 30 Hz, or 4x the workload of first-gen Marbles (720p 30 Hz).

Update 16:23 UTC: Cyberpunk 2077 is playing big on the next generation. NVIDIA is banking extensively on the game to highlight the advantages of Ampere. The 200 GB game could absorb gamers for weeks or months on end.

Update 16:24 UTC: New RTX IO technology accelerates the storage sub-system for gaming. This works in tandem with the new Microsoft DirectStorage technology, which is the Windows API version of the Xbox Velocity Architecture, that's able to directly pull resources from disk into the GPU. It requires for game engines to support the technology. The tech promises a 100x throughput increase, and significant reductions in CPU utilization. It's timely as PCIe gen 4 SSDs are on the anvil.

Update 16:26 UTC: Here it is, the GeForce RTX 3080, 10 GB GDDR6X, running at 19 Gbps, 238 tensor TFLOPs, 58 RT TFLOPs, 18 power phases.

Update 16:29 UTC: Airflow design. 90 W more cooling performance than Turing FE cooler.

Update 16:30 UTC: Performance leap, $700. 2x as fast as RTX 2080, available September 17. Up to 2x faster than the original RTX 2070.

Update 17:05 UTC: GDDR6X was purpose-developed by NVIDIA and Micron Technology, which could be an exclusive vendor of these chips to NVIDIA. These chips use the new PAM4 encoding scheme to significantly increase data-rates over GDDR6. On the RTX 3090, the chips tick at 19.5 Gbps (data rates), with memory bandwidths approaching 940 GB/s.

Update 16:31 UTC: RTX 3070, $500, faster than RTX 2080 Ti, 60% faster than RTX 2070, available in October. 20 shader TFLOPs, 40 RT TFLOPs, 163 tensor cores, 8 GB GDDR6

Update 16:33 UTC: Call of Duty: Black Ops Cold War is RTX-on.

Update 16:35 UTC: RTX 3090 is the new TITAN. Twice as fast as RTX 2080 Ti, 24 GB GDDR6X. The Giant Ampere. A BFGPU, $1500 available from September 24. It is designed to power 60 fps at 8K resolution, up to 50% faster than Titan RTX.

Update 16:43 UTC: Wow, I want one. On paper, the RTX 3090 is the kind of card I want to upgrade my monitor for. Not sure if a GPU ever had that impact.

Update 16:59 UTC: Insane CUDA core counts, 2-3x increase generation-over-generation. You won't believe these.

Update 17:01 UTC: GeForce RTX 3090 in the details. Over Ten Thousand CUDA cores!

Update 17:02 UTC: GeForce RTX 3080 details. More insane specs.

Update 17:03 UTC: The GeForce RTX 3070 has more CUDA cores than a TITAN RTX. And it's $500. Really wish these cards came out in March. 2020 would've been a lot better.

Here's a list of the top 10 Ampere features.

Update 19:22 UTC: For a limited time, gamers who purchase a new GeForce RTX 30 Series GPU or system will receive a PC digital download of Watch Dogs: Legion and a one-year subscription to the NVIDIA GeForce NOW cloud gaming service.

Update 19:47 UTC: All Turing cards support HDMI 2.1. The increased bandwidth provided by HDMI 2.1 allows, for the first time, a single cable connection to 8K HDR TVs for ultra-high-resolution gaming. Also supported is AV1 video decode.

Update 20:06 UTC: Added the complete NVIDIA presentation slide deck at the end of this post.

Update Sep 2nd: We received following info from NVIDIA regarding international pricing:

UK: RTX 3070: GBP 469, RTX 3080: GBP 649, RTX 3090: GBP 1399
Europe: RTX 3070: EUR 499, RTX 3080: EUR 699, RTX 3090: EUR 1499 (this might vary a bit depending on local VAT)
Australia: RTX 3070: AUD 809, RTX 3080: AUD 1139, RTX 3090: AUD 2429

Add your own comment

502 Comments on NVIDIA Announces GeForce Ampere RTX 3000 Series Graphics Cards: Over 10000 CUDA Cores

#301

PowerPC

Vya DomusI tried, yet you made sure to prove to me you have absolutely no idea what you are talking about on every occasion. Do you realize how idiotic it would be if I were to write say a + b in a GPU shader and the CPU would have to instruct the GPU step by step on how to do that ? What would there even be the point of having a GPU ?

Shaders run exclusively on the GPU, line by line, instruction by instruction with no intervention from the CPU side, that's what GPUs do. You know precisely nothing about this subject yet you insist to correct me on everything I say. May you delve in your ignorance for as long as you want, I'm out.

en.wikipedia.org/wiki/Dunning–Kruger_effect#:~:text=In%20the%20field%20of%20psychology,recognize%20their%20lack%20of%20ability

You are just insanely inaccurate about everything you say.

I never said CPU does the computation of the GPU line by line. That is something you just invented somehow. You just can't stop lying about what I said, you really don't give a f*** it seems.

The CPU can give instructions to the GPU and the GPU will actually compute those instructions. The CPU doesn't need to do it, but it does need to orchestrate what the GPU does. That is what that program is doing that someone wrote in C. The program itself doesn't do anything either, it's just the instructions for the CPU and the CPU passes those instructions on to the GPU to compute whatever that program wants it to compute.

If you knew something about C you would know it gets compiled down to Assembly language and then down to Machine code for x86 CPUs. Assembly usually just knows one instruction set, so it can't just run on anything you would like.

You are only projecting your own lack of knowledge on to others. You don't seem to be that bright either, literally everything you said was wrong so far and even others have corrected you, it's actually hard for me to remember encountering someone as willfully ignorant or even just as purposefully deceptive as you on here before.

#302

Vya Domus

Maybe you have trouble remembering what you wrote so I will post the exact comments and order in which you replied.

Vya DomusYes, it runs purely on the GPU, instruction by instruction for each instance of the shader.

To which you responded very confidently that this is "garbage" :

PowerPCThat is garbage. CPU always works together with the GPU. CPU instructs the GPU to do things all the time.

Now you write :

PowerPCI never said CPU does the computation of the GPU line by line. That is something you just invented somehow. You just can't stop lying about what I said, you really don't give a f*** it seems.

No buddy, you thought that idea was "garbage", you wrote that yourself not knowing even what I was talking about.

Stop, you are the laughing stock of everyone reading this.

#303

PowerPC

Vya DomusMaybe you have trouble remembering what you wrote so I will post the exact comments and order in which you replied.

To which you responded very confidently that this is "garbage" :

Now you write :

No buddy, you thought that idea was "garbage", you wrote that yourself.

Stop, you are the laughing stock of everyone reading this.

How is that even supposed to be against what I said???

YOU said it runs line by line on the GPU. Then I replied that's garbage / totally wrong. I said it runs on both every time and you don't seem to get that simple principle. You have posted total non-sense over and over again.

The CPU is still always the host processor, it orchestrates what the GPU has to do. The program just doesn't purely run on the GPU, but also on the CPU. You have no sense of the abstraction happing with that C program. I'm still accurate on that, despite your silly attempts to frame this otherwise. You make extremely illogical points, so that won't happen either way. But somehow you are still trying.

#304

Vya Domus

PowerPCI said it runs on both every time and you don't seem to get that simple principle.

It doesn't run on both, you have no clue in the slightest how these things work. I said above how unimaginably idiotic it would be for that to be the case. A shader is compiled and runs on the GPU alone, not on both. Can you still not understand how dumb that would be ?

PowerPCThe program just doesn't purely run on the GPU

The shader runs purely on GPU, period. I don't know what's this "program" you talk about, I wrote exclusively about shaders.

PowerPCYOU said it runs line by line on the GPU.

Which is 100% correct. Everything in the shader, the code that I posted earlier, runs on the GPU. All of it, never on both.

PowerPCYou have no sense of the abstraction happing with that C program.

:roll: :roll: :roll: :roll:

IT'S NOT A C PROGRAM. It's a GLSL shader using a C syntax style language, I wrote that as clear as possible, of course you don't have any idea of what that means. That's why you don't understand any of this either. This is what I am trying to show you.

#305

PowerPC

Vya DomusIt doesn't run on both, you have no clue in the slightest how these things work. I said above how unimaginably idiotic it would be for that to be the case. A shader is compiled and runs on the GPU alone, not on both. Can you still not understand how dumb that would be ?

The shader runs purely on GPU, period. I don't know what's this "program" you talk about, I wrote exclusively about shaders.

The computation for the shader runs on the GPU, that's it. The C programs run on both, that's it. The discussion was about the C program, and you started it, so you're just out of your mind now suddenly changing the facts up to appear right. That's some weird, childish reasoning.

Let's sum up:
You were wrong on the TFLOPS argument.
You were wrong on the general-purpose argument.
You are still wrong on the C program argument, which btw. you stared and which has nothing at all to do with this whole discussion.

Or what did this have to do with anything now? You wrote this:

Vya DomusShading languages don't run on specialized hardware, they can't, they need generic all-purpose processors.

Of course they run on specialized hardware. It's called a Shading Unit inside a GPU. That's definitely as specialized as it gets, my friend. And there is no such thing as a "shading language", if it's written in C. C is an all-purpose language, no a shading one.... Whatever that is even supposed to mean.

Now stop this completely silly discussion. But you just don't seem to know when to stop, do you?

#306

Vya Domus

PowerPCThe discussion was about the C program

IT'S NOT A C PROGRAM

You can't figure that out even now ? Read back that comment.

PowerPCYou were wrong on the TFLOPS argument.
You were wrong on the general-purpose argument.
You are still wrong on the C program argument, which btw. you stared and which has nothing at all to do with this whole discussion.

I was correct about all of them.

More TFLOPS indicate more performance.
GPUs are general purpose, that's why they are used for all sorts of things besides graphics. That's also why they're capable of being GPGPU (General Purpose Graphics Processing Units). It's in the name.

Thoroughly explained why, can you do the same ?

PowerPCAnd there is no such thing as a "shading language", if it's written in C. C is an all-purpose language, no a shading one.... Whatever that is even supposed to mean.

en.wikipedia.org/wiki/OpenGL_Shading_Language

Just for how long will you make me do this ? Have I not proven you wrong enough times ? I feel like a bully.

You're gonna use that cognitive dissonance thing and tell me that you never said that, right ?

#307

B-Real

xkm1948$699 3080???????????????????

OH MY GOD

Wut? This may be the same pricing as the RTX 2000 series. No Founders Edition mentioned in the charts, so the cheapest AIB ones may start from this price range too - just like with the RTX 2080.

xkm1948Nope. To me Nvidia is only competing with itself at this point. In other words, how to get the Pascal crowd to move on and upgrade? Jensen even mentioned that in his talk!

Competition doesn't mean only the high end. It's there at mid and entry range too.

#308

PowerPC

Vya DomusIT'S NOT A C PROGRAM

You can't figure that out even now ? Read back that comment.

You can keep changing and editing posts but that's ok.

You haven't answered the question of how that language (C or not, it doesn't matter) purely runs on the GPU. Everything has to go through the CPU at some point, GPU and CPU work together constantly and for every process, they do to reach the result of a game on the screen. That's why you still need a CPU to play games. We wouldn't even need a CPU, if you could just program on the GPU. It's definitely not how it works.

And either way, what is this point even trying to accomplish? What are you trying to reach with this exact point, besides all the other points that you were wrong on? You still haven't responded to this, so I'll quote you again:

Vya DomusShading languages don't run on specialized hardware, they can't, they need generic all-purpose processors.

Are you actually saying Shading Units aren't specialized hardware when it's literally in the name??? :kookoo::kookoo::kookoo:

#309

Parn

The power usage is crazy. Both the 90 and 80 are above 300W and even the 70 is approaching the territory that used to be reserved for the 80 Ti cards. But then if 3070 is actually faster than 2080 Ti by a good margin, that means power efficiency has been improved with Ampere. BTW where is that 12-pin pcie aux power connector?

#310

biffzinker

efikkanSpecs are looking pretty good (on paper), perheps except TDP numbers.

I thought the 220 watts for the 3070 wasn’t a bad TDP for the performance their advertising.

#311

Vya Domus

PowerPCYou haven't answered the question of how that language (C or not, it doesn't matter) purely runs on the GPU. Everything has to go through the CPU.

You write "int a = 1;" in a shader, it's compiled then sent onto the GPU where it's executed by it. Is that too much to comprehend ? Do you understand what "run" means ?

PowerPCAre you actually saying Shading Units aren't specialized hardware when it's literally in the name??? :kookoo::kookoo::kookoo:

Both AMD and Nvidia have stopped calling them like that for a very long time. Anyway, you still haven't addressed how shading languages don't exist despite this little thing :

en.wikipedia.org/wiki/OpenGL_Shading_Language

(GLSL) is a high-level shading language with a syntax based on the C programming language.

PowerPCAnd either way, what is this point trying to accomplish?

That you know nothing, you are a compulsive liar, and really, really stubborn. It wasn't my initial goal, but alas.

#312

PowerPC

Vya DomusYou write "int a = 1;" in a shader, it's compiled then sent onto the GPU where it's executed by it. Is that too much to comprehend ? Do you understand what "run" means ?

So you still really think everything that code does only and purely runs on the GPU. That is completely wrong and I won't keep repeating this point because you don't seem to understand how computers or Von Neumann Architecture works on a basic level.

Vya DomusBoth AMD and Nvidia have stopped calling them like that for a very long time. Anyway, you still haven't addressed how shading languages don't exist despite this little thing :

No, it's still called that because Nvidia or AMD don't determine what basic hardware components are called. Intel definitely also uses Shading Units in their iGPUs.
en.wikipedia.org/wiki/Unified_shader_model

Unified shader architecture (or unified shading architecture) is a hardware design by which all shader processing units of a piece of graphics hardware are capable of handling any type of shading tasks. Most often Unified Shading Architecture hardware is composed of an array of computing units and some form of dynamic scheduling/load balancing system that ensures that all of the computational units are kept working as often as possible.

Vya DomusThat you know nothing, you are compulsive liar, and really, really stubborn. It wasn't my initial goal, but alas.

Again, totally projecting. What was your initial goal other than presenting how much you know about shaders but not the overall picture of how a computer works? I don't need to know how everything works because I know how abstraction in computer architecture works, that's the point of having it. So, you still haven't told me about your initial goal. It seems to be that you just wanted to talk s*** at other members of the forum and nothing else. You are truthfully a sad and silly human being. I kind of pitty you that you had to go through this length to defend a useless point that doesn't even further anything about the discussion. You seem to just vent because something went wrong in your life?

#313

Vya Domus

PowerPCSo you still really think everything that code does only and purely runs on the GPU. That is completely wrong

It's a fucking GPU shader ! Of course it runs only on the GPU, THAT'S WHAT IT'S FOR ! My God, you can't be this dense, really I hope you're just a bad troll. :roll:

PowerPCyou don't seem to understand how computers or Von Neumann Architecture works on a basic level.

Lmao you're just writing some random ass computer science thingies that you know about.

PowerPCAgain, totally projecting. What was your initial goal other than presenting how much you know about shaders but not the overall picture of how a computer works? I don't need to know how everything works because I know how abstraction in computer architecture works, that's the point of having it. So, you still haven't told me about your initial goal. It seems to be just wanted to talk s*** at other members of the forum and nothing else. You are truthfully a sad and silly human being. I kind of pitty you that you had to go through this length to defend a useless point that doesn't even further anything about the discussion. Did you just need to vent from something that went wrong in your life?

Damn are you getting emotional, want a tissue or something ? You're not gonna cry on me are you ?

#314

Flanker

Over9000.gif

#315

PowerPC

Vya DomusIt's a fucking GPU shader ! Of course it runs only on the GPU, THAT'S WHAT IT'S FOR ! My God, you can't be this dense, really I hope you're just a bad troll. :roll:

You have no idea what a small line of higher-level code is compiled to in the computer and what components carry out that code. And get this, it doesn't even matter, it's called abstraction! But you have no clue what that is, do you?

Now, you still haven't answered why you even wrote that comment about languages not running on specialized hardware, which they actually can do with the help of the CPU, since it literally had nothing to do with anything being discussed. It's not relevant to the specialized hardware point that you were wrong about. You can still have specialized hardware, like Shader Units or TMUs or ROPs... It's also not relevant to the TFLOPS argument that you were also wrong about. So what was it even relevant for, you silly goose? Tell me, I asked at least 3 times already and you can't answer, but I can continue asking. :laugh: :laugh: :laugh:

Vya DomusLmao you're just posting some random ass computer science thingies that you know about.

Sure I am. I just have a CS degree for nothing. But you are the expert... Probably you tried programming a game once and didn't even work out and now you like to play yourself up as someone of importance here. It's not working for you, sorry to be the one who tells you the truth.

Vya DomusDamn are you getting emotional, want tissue or something ? You're not gonna cry on me are you ?

I'm not the one who needs to cry, but people who are around you. And they probably also need to run from what kind of pathetic lier and wannabe you are. I won't even say troll because you somehow manage to be way beneath that.

#316

Unregistered

Finally a real leap with focus on Cuda cores, even the 3070 has more than the 2080ti, a bit disappointed by the memory amounts though, 12 GB for the 3070 and 16gb for the 3080 would've been better.
Wonder what will AMD do now, they have to launch a big GPU, we need the power to drive to run games at 4k.

#317

mrthanhnguyen

can anyone calculate how many more fps for 1440p going from 2080ti to 3090?

#318

ViperXTR

wait,
Oh so from what ive read, each core can do INT+FP, FP+FP vs previous generation of INT+ FP only? Its still 2 core inside but only one when doing specific operations?

#319

AteXIleR

I'm thankful and glad for the probable generational leap, and for the price of the two lower more mainstream models.

#320

Unregistered

ViperXTRwait,
Oh so from what ive read, each core can do INT+FP, FP+FP vs previous generation of INT+ FP only? Its still 2 core inside but only one when doing specific operations?

Interesting, can someone explain this for the laymen among us.

#321

Lionheart

I know Nvidia recommends a 750W PSU for the 3080 but I'm hoping my 650W Gold rated PSU will suffice, reviews/time will tell.

#322

nguyen

Xex360Interesting, can someone explain this for the laymen among us.

Read through this concurrent FP+INT article
www.nvidia.com/en-us/geforce/news/geforce-gtx-1660-ti-advanced-shaders-streaming-multiprocessor/

Basically 2080 Ti has 13.5TFLOPs of FP32 and 13.5TFLOPs of INT32, if a game fully leverage both FP32 and INT32 instructions then 2080 Ti would effectively have 27TFLOPs of combined FP+INT. So depends on how much INT32 instructions are used, 2080 Ti's effective TFLOPs range from 13.5 to 27. From the SoTR example, 38 out 100 instructions are INT, that means effectively 2080 Ti has 13.5 + 13.5(x 38/62) = 21.77 TFLOPs

Meanwhile the 20TFLOPs of the 3070 are fixed (FP32 or FP32+INT32) and does not depend on the game engine's usage of INT instructions.

@Vya Domus and @PowerPC You guys forgot that 2080 Ti can do concurrent FP32+INT32 so effectively 2080 Ti has 27TFOPS in rare instances.

#323

yotano211

LionheartI know Nvidia recommends a 750W PSU for the 3080 but I'm hoping my 650W Gold rated PSU will suffice, reviews/time will tell.

I wouldnt push it.
If I ever build a desktop again, I would never mess with or cheap out on the power supply, I would buy everything else used but the power supply, I would buy new. So I wouldnt push it.

#324

wolf

Performance Enthusiast

Well well well, looks like a lot of rose coloured crystal balls were off the mark.

A 3080 is definitely the card I was waiting for!

#325

Valantar

LionheartI know Nvidia recommends a 750W PSU for the 3080 but I'm hoping my 650W Gold rated PSU will suffice, reviews/time will tell.

Depends on the rest of your hardware and how noisy your PSU is, but 650W with a 320W GPU should be perfectly fine unless you're running very high end hardware or overclocked. If Ampere is power limited to TDP like Turing, you have 320W for the GPU, plus 20-30 for the motherboard and RAM, ~5 per SSD, 15 per 3.5" HDD, ~5W per couple of fans, 10 per AIO pump, and however much power your CPU needs. For a 65W AMD that is 88W, for a 95W AMD it's 144W, for a 10th gen Intel it's 150-250W depending on the SKU. That's at stock clocks within the boost window (which might be infinite depending on your motherboard). I would add 20% margin on top of that for safety, and at least another 20% if you're overclocking - likely more. Of course it's highly unlikely for all components in the system to draw maximum power at once, and CPUs pretty much never run at 100% while gaming, so there's some extra margin in there based on that too. 650W would as such be rather slim for a system with something like a 10700K or 10900K (my formula ends up at 675W minimum assuming a couple of SSDs and a few fans), but should work fine with a less power hungry CPU or if you undervolt and/or tune the power limits of one of those power hogs.

Add your own comment

NVIDIA Announces GeForce Ampere RTX 3000 Series Graphics Cards: Over 10000 CUDA Cores

502 Comments on NVIDIA Announces GeForce Ampere RTX 3000 Series Graphics Cards: Over 10000 CUDA Cores

Latest GPU Drivers

New Forum Posts

Popular Reviews

Controversial News Posts

NVIDIA Announces GeForce Ampere RTX 3000 Series Graphics Cards: Over 10000 CUDA Cores

Related News

502 Comments on NVIDIA Announces GeForce Ampere RTX 3000 Series Graphics Cards: Over 10000 CUDA Cores

Latest GPU Drivers

New Forum Posts

Popular Reviews

Controversial News Posts