NVIDIA GeForce RTX 3070 and RTX 3070 Ti Rumored Specifications Appear

[AleksandarK]

NVIDIA is slowly preparing to launch its next-generation Ampere graphics cards for consumers after we got the A100 GPU for data-centric applications. The Ampere lineup is getting more and more leaks and speculations every day, so we can assume that the launch is near. In the most recent round of rumors, we have some new information about the GPU SKU and memory of the upcoming GeForce RTX 3070 and RTX 3070 Ti. Thanks to Twitter user kopite7kimi, who had multiple confirmed speculations in the past, we have information that GeForce RTX 3070 and RTX 3070 Ti use a GA104 GPU SKU, paired with GDDR6 memory. The cath is that the Ti version of GPU will feature a new GDDR6X memory, which has a higher speed and can reportedly go up to 21 Gbps.

The regular RTX 3070 is supposed to have 2944 CUDA cores on GA104-400 GPU die, while its bigger brother RTX 3070 Ti is designed with 3072 CUDA cores on GA104-300 die. Paired with new technologies that Ampere architecture brings, with a new GDDR6X memory, the GPUs are set to be very good performers. It is estimated that both of the cards would reach a memory bandwidth of 512 GB/s. So far that is all we have. NVIDIA is reportedly in Design Validation Test (DVT) phase with these cards and is preparing for mass production in August. Following those events is the official launch which should happen before the end of this year, with some speculations indicating that it is in September.


View at TechPowerUp Main Site

 

[Assimilator]

The cath?

 

[cucker tarlson]

21 gbps on a 70ti

 

[droopyRO]

So RTX 3070 = ~RTX 2080 ? If the price is on par with a brand new RTX 2070 then it's gaming time starting this fall.

 

[TheoneandonlyMrK]

So, absolutely definitely a September paper launch with a dribble of cards until January/march.

 

[EarthDog]

Not sure I buy this... that is a ton of bandwidth for a 'mid-range' card.

The cath?

lol, 'the catch' I'd imagine.

 

[RedelZaVedno]

Only 2944/3072 CUDA cores on 7 nm die? This thing will have to be clocked way past 2 GHz in order to match/beat 2080TI (4352 CU). I'm hoping we're not in for another Turing like disappointment.

2.944/3072 shaders x 2 Ghz x 2 instructions per clock = 11.78/ 12.29 FP32 (float) performance... (RTX 2080TI FP32 = 13.45)

Either this thing can boost clocks higher than 2Ghz (+2.2Ghz), has +10% IPC gain, or we're f...ed.

GTX 670 was 6 % slower than GTX 590 ....... 40->28nm
GTX 770 was 51 % slower than GTX 690 ($1K GPU in 2012!) .......-> 28nm both
GTX 970 was 3% slower than GTX 780 TI ...... 28nm both
GTX 1070 was 12% faster than GTX 980TI 28nm .... 28nm ->14nm
RTX 2070 was 10% slower than 1080TI ...... 14->12 nm
RTX 3070 ? 2080TI ....... 12->10 nm or 12nm -> 7nm?


Are we in for another Turing moment? If my calculations stand, we might get 2080S level of performance for 500 bucks when Xbox X costs the same while offering similar TFlops horsepower.
Something's not right here. 3070(TI) being slower than 2080TI would just not be a worthy upgrade considering new consoles performance. Or Nvidia just don't care about PC gamers migrating to consoles?

 

[Deleted member 185088]

So RTX 3070 = ~RTX 2080 ? If the price is on par with a brand new RTX 2070 then it's gaming time starting this fall.

I was expecting something like the 1070 with performance similar to that of the 2080ti, given that they moved to a new node.

 

[dinmaster]

rumored specs... grain of salt, i also saw an article talk about 8nm from samsung instead of 7nm because amd took all the allocation at tsmc? until its announced the hype train continues..

 

[Valantar]

Definitely a lot of parts here that don't quite add up. Debuting a new type of memory on an xx70 Ti SKU? Expecting the same memory bandwidth on both SKUs despite the higher tier one having more advanced and faster clocking memory? Just a handful of shaders separating them?

Also, starting out with an xx70 Ti SKU in the lineup speaks to the price ladder extending higher yet again - adding rungs to the middle at launch isn't likely to push lower tier prices down, just push higher ones up. Hooray!

Also, mass production in August either means mass production of the die in August, which means 4-6 months before these hit the streets, or mass production of partner cards in August, which as noted above means some early cards might arrive 1-2 months after, no time to build up stocks, so availability will be bad for several more months.

 

[Tomorrow]

I still find it dubious there exists a memory type out there that no one has heard of before and that has speeds that match HBM2.
Even with GDDR5X there was something about it months before from Micron. You would thing there would be some proof of it's existance?

And would this be Nvidia exclusive with AMD using "regular" GDDR6?
Or does Nvidia think that AMD will use HBM2 and thus needs to have this special memory to compete?

I mean hey - if it turn out to be accurate - great. But i still have my reservations.

 

[stimpy88]

So no additional memory bandwidth for the 3070 over the 2070?

I bet the price will go up though...

 

[Punkenjoy]

Edited due to erronous infos got corrected !

 

[ppn]

8nm optical shrink of 2080/S.

 

[BoboOOZ]

Either this thing can boost clocks higher than 2Ghz (+2.2Ghz), has +10% IPC gain, or we're f...ed.

Past rumors were indeed for more than 10% IPC gain.

But wait&see and big grains of salt seem to apply to most of the latest rumors...

 

[EarthDog]

You're new to this it seems.

FTFY.

But wait&see and big grains of salt seem to apply to all rumors...

 

[Vya Domus]

Only 2944/3072 CUDA cores on 7 nm die? This thing will have to be clocked way past 2 GHz in order to match/beat 2080TI (4352 CU).

Who said it's supposed to beat a 2080ti ? The days of true generation improvements are over, for technical and business reasons.

HBM2 and HBM2E have the possibility to deliver more bandwidth, but you would have to have a lot of memory on board.

No, it's the other way around. The reason HBM is faster is because the bus width is very wide, you can connect a single module with a 1024 bit wide bus, GDDR6 chips are each connected with a measly 32 bit interface. You need far, far less chips with HBM than with GDDR6 or any other conventional memory type. Just to make a point, if you were to use the fastest HBM modules avaiable you'd only need two of them to get 920 GB/s, almost a TB/s.

HBM give less bandwidth per GB than GDDR6/6x

It doesn't even makes sense to think of it in terms of "bandwidth per GB". It's bandwidth per module and as explained above HBM is basically an entire order of magnitude superior in this regard.

 

[BoboOOZ]

You're new to this it seems.

FTFY.

Well, rumors/leaks from a few months ago were at least coherent with themselves.
Here I don't even understand the numbers. 21Gbs is what kind of speed, sounds more like bandwidth? 21GHz?
If it's 21GHz, why the final bandwidth is so low, only 512GB/s?

 

[Vya Domus]

If it's 21GHz, why the final bandwidth is so low, only 512GB/s?

Supposedly "it's up to". I don't think we'll see 21 Gbps any time soon.

 

[EarthDog]

It's rumors man. Some make sense, most don't. Your advice is sound.

 

[Valantar]

The main thing that slow down HBM is the cost. HBM2 and HBM2E have the possibility to deliver more bandwidth, but you would have to have a lot of memory on board.

HBM2E can deliver 460 GB/s per 16 GB stack. Something similar to GDDR6 can deliver with 8 GB on a 256 bit bus on the Radeon RX 5700 XT

For a memory that already cost more per GB than GDDR6 and require a costly interposer instead of using printed circuit board, that make this memory way costlier than this.

That is the main reason. you can still get more bandwidth if you go with a 4 chips design like Vega64 cpu, but it would mean a 64 GB GPU.

HBM give less bandwidth per GB than GDDR6/6x

There's a significant misunderstanding here: HBM(2(E)) stacks are flexible in their sizing. A HBM2E stack can (IIRC) be 2-16GB (1-high to 8-high), without affecting the bandwidth of the stack (as that is determined by the controller die on the bottom of the stack, not the number of stacked chips). In other words you could get a ridiculous bandwidth 8GB HBM2e setup with four 2GB stacks, but you'd then need to be willing to pay the cost (more stacks is more expensive than taller stacks; more stacks requires more controllers and physical connections on the GPU/CPU die; more stacks requires a larger interposer). Something like the 8GB 2-stack Vega setup is a reasonable compromise, though any new product with something similar would likely use much faster memory as well.

 

[RedelZaVedno]

Who said it's supposed to beat a 2080ti ? The days of true generation improvements are over, for technical and business reasons.

GTX 670 was 6 % slower than GTX 590 ....... 40->28nm
GTX 770 was 51 % slower than GTX 690 ($1K GPU in 2012!) .......-> 28nm both
GTX 970 was 3% slower than GTX 780 TI ...... 28nm both
GTX 1070 was 12% faster than GTX 980TI 28nm .... 28nm ->14nm
RTX 2070 was 10% slower than 1080TI ...... 14->12 nm
RTX 3070 ? 2080TI ....... 12->10 nm or 12nm -> 7nm?

As you can see every time NGreedia went to double node density, we got around previous generation's xx80 TI level of performance. Why would anyone even consider buying 3070 if it offers 2080(S) level of performance for 500 bucks? Nvidia's 3070 GPU has to beat XBox X by a decent margin or many PC gamers will have hard time justifying the purchase and might skip Ampere or even decide to buy new console instead. For example, I own 3.5 yo 1080TI and I see no reason in upgrading to Ampere if it doesn't offer 2080TI level of performance for 500 bucks so Nvidia wouldn't see my money for another 2-3 years if performance isn't there. I don't know if that's in their best business interest.

 

[moproblems99]

NGreedia

Oh shit, here we go again. Want lower prices? Stop buying things. And convince everyone else to do the same.

 

[ppn]

Clearly our upgrade path is 3080Ti, so you get 46% more Cuda with 30% IPC async compute variable shader rate capable and 50% more power hungry, also 1TB s bandwidth In the case of the mythical GDDR6X. 1080Ti/2070S +80%.

 

[xkm1948]

Newest rumor from Chiphell claims the upcoming 3080Ti (or the non Titan flagship equivalent) will boost 24GB of VRAM.

旧帖归档 - Chiphell - 分享与交流用户体验

旧帖归档Chiphell - 分享与交流用户体验 ,Chiphell - 分享与交流用户体验

translate.google.com

$1499 for lowest end 3080Ti would I dare say?