Speculation: GPU Performance Comparisons of 2020 *Spawn*

And I ran 7970 OCed so I'd rather not to.

No broken display PHYs here henceforth no driver issues.
Nice.

 

Yep, none.

 

Well, yeah, something is really apocalyptic about DCN2 or its PHYs.
One cable swap into spazzing is very odd when the GPU and the driver version is the same.

 

I’d expect more the usual issues on launch and not a Navi10 kind of mess. It seems that bugs in the hardware really were responsible for a part of that. I also expect the drivers to be in a better state in general. Personally wouldn’t buy anything on launch, anyway.

 

Don't forget Navi 10 is 40 CU's, the PS5 is only 36. That already makes up for most of the clock speed difference.

And they both have 448GB/s of bandwidth, but the PS5 has to share it's bandwidth with the CPU.

 

It literally depends on the proportion of transistors that are dedicated to ray tracing.

I haven't read the AMD-specific ray tracing patent documents, so I have no idea if we can establish this proportion.

 

Bro, I thought we're talking about Zen 3 vs Willow Cove on desktop.

 

Zen3 on DT gets an even nicer mem setup, which is moar %% ST, and it's not WLC in RKL.
Not at all

 

That remains to be seen. If Rocket Lake is based on a Skylake-derivative uarch then yes Zen3 will wipe the floor with it.

 

RKL is a genuine frankencore.
It's stuck somewhere between SKL and SNC.

 

I'm as skeptical of @Bondrewd 's statements as the next guy, but this comparison couldn't be farther from fair and honest.

A100 is nvidia's flagship chip (i.e. the one probably selling with the highest margins), and you're comparing it to AMD's CPUs with the lowest margins.
AMD's flagship CPU at the moment is the Epyc 7742 which is selling for $5000-7000 each. It contains one 125mm^2 IO die made on the super cheap GF 14nm, and 592mm^2 worth of 7nm Zen2 chiplets, split into 74mm^2 chiplets, meaning they maximize yields.
The A100 is probably getting a higher revenue-per-chip, but its monstrous monolithic size means yields can't be spectacular and nvidia can't sell it without putting it into an interposer together with six HBM2 chips and then in a PCB with voltage regulation. Oh and then they need to put it into a motherboard with a couple of Epyc 7742 CPUs, which they buy from.. AMD.


And then there's the fact that the TAM for Epyc is probably on a completely different order of magnitude compared to an A100.
Intel's quarterly datacenter revenue has been around $7B, which is significantly greater than nvidia's latest projected $1.7B (of which 1/3rd come from Mellanox network hardware sales, meaning GPU sales account to less than $1.2B).


If AMD has the superior product and has successfully gained traction in the server space, it's a no brainer that this should be their focus. Navi is indeed just a child's toy when compared to the revenue potential of Zen 2/3/4.

 

400-something.
Rome sIOD is the size of SKL-SP HCC die, lol.

Rome is fairly scarce which is why GOOG has deployed A100...
...with Cascade Lake.

They finally did, Q2 was double digit unit share (not accounting for Octeon-tier edge networking where Intel now plays and AMD is yet to) and Rome is selling out, more or less.

Eh, the mainstream ones (22/23) will make some nice money and will certainly anchor AMD in expensive-ish gaming laptops.

 

Exactly my point, that's why I exaggerated it soo hard. One-liners do not help the discussion.

 

If the 290X is a furnace then what are the RTX30 series to you?

This new nvidia gen and all the people who are now super excited for it just proves that absolute power consumption on desktop graphics cards was never a real concern, but one that was manufactured by marketing divisions and further picked up by fanboys.


Now that nvidia cards pull over 300W, let's see how many times we'll have with people - and reviewers - doing those napkin maths of how much more they need to pay in annual power bills by using completely bonkers numbers like assuming everyone will play 8 hours a day every single day of the year.
All of a sudden I also don't see many people complaining how there aren't any mITX cards on the RTX30 range to put in the tiniest cases they saw in the Internet.

What matters is whether or not the card fits the PSU (most decent 650W ones will), and whether or not the cooling is adequate for its chip and silent.

Nvidia seemingly went above and beyond with their new coolers which is great, but it also proves that the FUD they and their fanboys generated over cards that consume 300W+ is completely manufactured.

 

290x was inadequately cooled which resulted in not very nice temps and sound signature.

Narrative always shifts and goes wherever the market leader moves.
It's okay, only a month left.

 

"Rays per watt" - there, I said it.

 

I'm puzzled.
Why would Nvidia distribute PCAT to select NDA partners just now? They are probably concerned about having a single "highest power number" with their new cards without corresponding increase in perf. But at the same time, total system consumption always masks graphics cards power as well.

If Nvidia was particularly concerned about AMD beating them in perf/watt, why would they make available PCAT right now?

 

what if pigs flied?

 

True, but everything in this thread has been IF so far.