Speculation: GPU Performance Comparisons of 2020 *Spawn*

[Entropy]

He claimed to have access to something faster than 2080ti at work. It's unlikely any unreleased hw would be given out without nda's for people to test and spill the beans into internet. Likely there is nda between companies and another nda between company-employee/contractor. There is more than one post but you could read the existing discussion here(see posts before and after). Maybe he has access to something, maybe not.

https://forum.beyond3d.com/posts/2115508/

I was aware - I just pointed out that it doesn’t say anything meaningful without actual information. Of course there are faster products in the pipeline from all (2) pertinent players. Just saying that there is a faster prototype product in the wild - well duh. Which means that you can also pose as having one, since either AMD or nVidia is bound to release something vaguely like it and your credibility will be unharmed. In fact, you will have gained insider status, and can troll forums with cryptic statements forevermore.

Edit:Removed irrelevant stuff.

 

[gamervivek]

I'm not sure what maths they're doing but 2080Ti is 50% faster than 5700XT at 4k, so even if "Big" Navi is twice as fast as 5700XT then it'd still only be 33% faster than 2080Ti

https://www.techpowerup.com/review/asus-radeon-rx-5700-xt-tuf-evo/28.html

The Big Navi specs also look a little too optimistic for 2x5700XT performance( 15:15 ) with just a 427mm2 die and 384-bit bus.

I can see RDNA2 cards doing really well on clockspeed front, with PS5 gpu doing over 2.2GHz and Vega in Renoir doing 2.1GHz stock with a 2.4GHz overclocked 3DMark result floating around. So performance could "linearly" scale with number of units increase.



otoh, leakers are saying 3080 will be doing 2080Ti + 20%,

https://twitter.com/x/status/1285454528339259392

3080Ti/3090 should be another 20-30% on top of that for 45-60% over 2080Ti though RT performance would be more relevant anyway.

 

[Bondrewd]

with just a 427mm2

Way bigger.

I can see RDNA2 cards doing really well on clockspeed front

To put it rather bluntly, yeah.
Those are real speed deamons.

3080Ti/3090 should be another 20-30% on top of that

I'm sorry to disappoint you but nV isn't breaking 400W in client this year.
Maybe next one.

 

[Megadrive1988]

Way bigger.

To put it rather bluntly, yeah.
Those are real speed deamons.

I'm sorry to disappoint you but nV isn't breaking 400W in client this year.
Maybe next one.

GeForce Hopper in 2022.

 

[Deleted member 2197]

GeForce Hopper in 2022.

Thought the rumor for Hopper's release was sometime in 2021.

According to a report we already covered, AMD's aggressive approach with 7nm caught NVIDIA by surprise, and to rectify that problem, they have already pre-booked TSMC 5nm for Hopper GPUs. While the TSMC order is confirmed, we have also heard rumors that NIVDIA is courting Samsung for their 5nm process as well, so maybe we will see them split orders.
...
The relevant portion of the Digitimes report, via @chiakohua (and reproduced with permission):
For next-generation GPUs based on the Hopper architecture, Nvidia has already pre-booked TSMC's 5nm production capacity in 2021, and is also in discussion with Samsung for smaller volume orders...
...In order to prevent AMD from getting any bigger, Nvidia has decided to catch up, even leapfrog AMD, in adopting TSMC's 7nm and 5nm EUV nodes. (DigiTimes)

https://wccftech.com/nvidias-hopper-architecture-will-be-made-on-tsmcs-5nm-process-launching-2021/

 

[gamervivek]

A review by an MSI engineer has the iGPU in 4750G overclocking to 2.65GHz,

For iGPU, expects 2500mhz OC on average, he showed 2650mhz.

https://linustechtips.com/main/topic/1225149-amd-ryzen-7-pro-4750g-review/

Way bigger.

To put it rather bluntly, yeah.
Those are real speed deamons.

I'm sorry to disappoint you but nV isn't breaking 400W in client this year.
Maybe next one.

Over 500mm2 or over 700mm2?

Maybe 400W won't be broken by stock configuration, but going close to 300W would be enough for it to be bad on pef/W compared to 2080.

 

[Bondrewd]

review by an MSI engineer has the iGPU in 4750G overclocking to 2.65GHz

Yeah at this point idk what AMD GPU teams are snorting, but they've learned to design some speedy circuits recently.

Over 500mm2 or over 700mm2?

A bit over 500mm^2; >700mm^2 stuff is reserved for DC at AMD.

but going close to 300W would be enough for it to be bad on pef/W compared to 2080

It will be very hot, yes; but not bad in perf/W.
nV just had to do something with what they had and this was the answer.
At least Ampere will be cheaper since the xtors and the market demand it so.

 

[JoshMST]

I'm certainly hoping that lessons learned from the integrated Vega in the 4000 series will have been ported into RDNA-2. Hard to say considering release dates, but cross pollination is certainly something that is stressed between design groups.

So many questions though on RDNA-2 and how it can be another big jump from the original. Also curious how NV will be able to address challenges in 7nm design, since it is such a big jump from current 16/12.

All I really want is competition in this space again and some parity in feature sets that gives consumers some real choices between the two.

 

[Bondrewd]

I'm certainly hoping that lessons learned from the integrated Vega in the 4000 series will have been ported into RDNA-2.

Those were the training grounds.

Also curious how NV will be able to address challenges in 7nm design

They already did with A100.

All I really want is competition in this space again and some parity in feature sets that gives consumers some real choices between the two

Choice doesn't matter, 80% of the market will get nVidia anyway.
AMD only ever exists to keep nV prices down.

 

[Jay]

Choice doesn't matter, 80% of the market will get nVidia anyway.
AMD only ever exists to keep nV prices down.

Can actually change, same way it has on the cpu side.
Will take a lot of good execution and stars aligning though. Something that haven't seen much of on the gpu side, but you never know.

 

[Bondrewd]

Can actually change, same way it has on the cpu side

Not even bargain bin Polaris moved the mss % needle. People don't want to buy Radeon, they want them to exist and be "competitive", whatever that means in their minds.

Will take a lot of good execution and stars aligning though

Let's see, they'll have 3 major uArch updates and 2 full node shrinks in 3 years, 2019-2021. Not accounting fancy DC stuff they're cooking too.
Is that enough?
I hope it's enough.

 

[Samwell]

Let's see, they'll have 3 major uArch updates and 2 full node shrinks in 3 years, 2019-2021. Not accounting fancy DC stuff they're cooking too.
Is that enough?
I hope it's enough.

Sounds interesting, what company you're talking about, which is doing this? Cause it won't be AMD, at least beside someones dreams. Only if you take steps like GCN1.0 to GCN1.1 as major Arch updates and extend the timeframe to 2022.

 

[Bondrewd]

Sounds interesting, what company you're talking about, which is doing this

Yeah.

Cause it won't be AMD

Yeah.
I ain't even talking DC roadmaps yet.
Almost makes me wish their software was ready by now and not late'21, but alas.

at least beside someones dreams

Roadmaps and no less than one taped out N5 die aren't dreams, unfortunately.

Only if you take steps like GCN1.0 to GCN1.1 as major Arch updates and extend the timeframe to 2022.

Oh noes, no.
One's gotta smash and fuck when they're behind.

 

[yuri]

Yeah at this point idk what AMD GPU teams are snorting, but they've learned to design some speedy circuits recently.

There were talks about that snorting stuff is coming from the CPU know-how realms. It's *a bit* late given 14 years since the AMD+ATi merge, but better late then never, I guess.

Not even bargain bin Polaris moved the mss % needle. People don't want to buy Radeon, they want them to exist and be "competitive", whatever that means in their minds.

The HD 4000-series did that trick having low cost, high performance and quite good drivers. People jumped on AMD bandwagon rather quickly back then. Nowadays, Polaris is a over 4 years old rebranded technology which wasn't stellar even at launch.

 

[Bondrewd]

There were talks about that snorting stuff is coming from the CPU know-how realms.

Yeah they borrowed quite a lot of CPU teams design methodologies.
And devised some new internally.
Seems very much working.

It's *a bit* late given 14 years since the AMD+ATi merge

Yeah this is very pre-Lisa AMD.
But they're a wiser bunch now.

The HD 4000-series did that trick having low cost, high performance and quite good drivers

Yeah problem is this was before GTX970 permanently fucked this market up.
Sorry I just don't have high hopes for AMD having a lot of traction in DIY GPU these days.
OEMs? That might be a fair shot.

 

[DegustatoR]

Sorry I just don't have high hopes for AMD having a lot of traction in DIY GPU these days.

There's nothing I can imagine in stopping this from happening besides the simple fact of their GPUs not actually being that much better than NV's despite what some of people are saying about them.

 

[PSman1700]

I read today that a 3080 is going to be 50% faster then a 2080, 20% faster then a 2080Ti. Sounds too good to be true?

 

[P_EQUALS_NP]

I read today that a 3080 is going to be 50% faster then a 2080, 20% faster then a 2080Ti. Sounds too good to be true?

it is a full node shrink so 50% more sees realistic, but i also expect a lot of the performance to be obtained as a frequency boost rather then 50% more cores.

 

[Bondrewd]

There's nothing I can imagine in stopping this from happening

Their brand is very dead and it'll take probably one more launch for people to start maybe considering Radeon as something besides nVidia cost cutting measure.

Sounds too good to be true?

No, but they've cranked the power up which is maybe unwise.
AMD? Turing backlash? who knows.

 

[madhatter]

I read today that a 3080 is going to be 50% faster then a 2080, 20% faster then a 2080Ti. Sounds too good to be true?

I would hope that it’s true, given the rumor that it uses GA102 and the rumored power draw.