nvidia Rtx 4090 is a monster (Official Benchmarks)

[Shzzit]

8 minutes ago, shermantanker said:

I would take either card, but the performance in games does seem negligible from the reviews this morning. I do like the beefy cooling solution, and it does run silent from what I am seeing. I had the Strix in my cart with Newegg, but I didn't finish the transaction fast enough and will sit and wait to see what stock looks like overall over the next few weeks.

Ya i been refeshing best buy and new egg every 2 seconds for an hour now lol.  NOTHING>....  I never not ONCE seen FE cards in stock at Bestbuy, not sure whats going on there.

[Sir Beregond]

8 hours ago, leadeater said:

I don't think it matters if it's L2, L3 or L4 so long as it does it's job and is effective.

We'll have to see. I think rumors on the 7900XT are claiming 384-bit with Infinity Cache. The 256-bit bus of the top RDNA2 cards definitely didn't do it any favors in 4k and performance did not scale to 4k. I mean it did fine, but they definitely got held back at 4k where-as GA102 Ampere pulled ahead.

 

Which makes me all the more skeptical on the two 4080's, especially the 12GB.

[leadeater]

38 minutes ago, Sir Beregond said:

We'll have to see. I think rumors on the 7900XT are claiming 384-bit with Infinity Cache. The 256-bit bus of the top RDNA2 cards definitely didn't do it any favors in 4k and performance did not scale to 4k. I mean it did fine, but they definitely got held back at 4k where-as GA102 Ampere pulled ahead.

See the posts after that one, 4k lost out because the cache wasn't really big enough and the cache hit rate was much lower so needed extra VRAM bandwidth that just wasn't there.

 

Going 384bit solves that, adding more cache also solves that, doing both well that's even better. Correct my memory but wasn't RDNA3 actually going to have less cache?

[TOMPPIX]

hahahaha https://wccftech.com/nvidia-dlss-3-frame-generation-lock-reportedly-bypassed-rtx-2070-gets-double-the-fps/

 

[Sir Beregond]

34 minutes ago, leadeater said:

See the posts after that one, 4k lost out because the cache wasn't really big enough and the cache hit rate was much lower so needed extra VRAM bandwidth that just wasn't there.

 

Going 384bit solves that, adding more cache also solves that, doing both well that's even better. Correct my memory but wasn't RDNA3 actually going to have less cache?

I think you are correct. Less cache, more bandwidth. I think I heard 96MB Infinity Cache.

[Dracarris]

4 hours ago, Shzzit said:

ANY cards for sale yet?  I cant tell if they are sold out or not for sale yet.

At least some Swiss retailers seem to have stock on AIB cards. Not really at attractive prices but also not as insane as last time (1 CHF is basically 1 USD currently).

 

 

[Kisai]

17 hours ago, Spotty said:

I think I'm misunderstanding you here. Do you mean vertical GPU mounts? You'd just be putting the graphics card cooler up against the side panel. Most cases with vertical GPU mounts probably won't even fit these larger 3+ slot 4090 cards in the vertical mount.

The idea is that the chassis would have a "standard GPU location" for optimal cooling, and the GPU would ship with a card/cable/mount assuming it will be used in the mount rather than directly on the motherboard.

 

I'm not saying the BeQuiet chassis itself is the optimal solution here for a 4090, but it's not any worse than putting it on the motherboard and having it butt up against a another GPU in the slot beneath it because it's a full 3 slots wide. 

 

Like in my PC, the (ASUS TUF) 3090 butts up beside the PCIe capture card, and the capture card is only 1.5" tall, so it's not terrible, but not ideal. The (ASUS TUF) 3070Ti is exactly the same thickness and when I was swapping the cards I was like "wow these are easy to confuse" , but trying to remove the 3070 was a huge pain because the PCIe x16 clip was difficult to actually reach from underneath the heatsink, and the SSD's were in the way.

 

Anyway. I know "fully loaded" desktops aren't as common outside the enthusiast space, but the time has come to stop trying to put 3-slot GPU's on the motherboard x16 slots. 

 

One of the chassis I think that either LTT or GN had, put the GPU in a transverse mount where the GPU ports were facing the top of the chassis, that would also be an option if you have no mechanical hard drives. But we're kind of have a problem here that I don't think the GPU mounting bracket is going to be able to take the entire weight of the GPU if it was hung from it without there being something to mount the other side.

 

One of the odd advantages Dell and HP have, is they can design these brackets to fit their chonky one-size-fits-everyone chassis designs. But SI's do not have the luxury, and many chassis are smaller because they've removed the 5.25" and 3.5" drive bays.

 

 

[SimplyChunk]

On 10/11/2022 at 9:42 PM, tim0901 said:

Will one of the fastest gaming CPUs around be CPU bottlenecked?

According to this:

It'll push the limit of a 5800X3D at 1440p high/ultra

[Kisai]

6 hours ago, Dracarris said:

At least some Swiss retailers seem to have stock on AIB cards. Not really at attractive prices but also not as insane as last time (1 CHF is basically 1 USD currently).

 

 

Memory Express says they have at least one AIB card in stock.

[Shzzit]

 

 

[ewitte]

I've been testing the MSI Gaming Trio I picked up from Microcenter.  This thing is insane.

 

80% power target (450W base) and very quick custom voltage curve.   That has a power target of 360W.  At about 1200 RPM it is in the 60s in a room that is 29C (85F).

 

Port Royal is about 24600 (25300 out of box)

Tried DLSS3 in spiderman.  160-180+ fps at 4k vs about 80 on my 3080 (with DLSS). - Watts is lower than my 3080 (under 300).

 

Not pushing this very much I have no AC.

 

[Dracarris]

1 hour ago, ewitte said:

At about 1200 RPM it is in the 60s in a room that is 29C

holy smack that's hot, how can you stay in such a room? I'd be sweating like crazy.

[ewitte]

59 minutes ago, Dracarris said:

holy smack that's hot, how can you stay in such a room? I'd be sweating like crazy.

It is not fun I have a few more days before there is a fairly big cool front.  About the time I get the parts I need to fix the AC it will drop from 93/70 outside to about 72/48...

 

BTW I do have an evaporative cooler I'm throwing ice in but it is pretty humid here it is probably helping about 5 degrees.

[Shzzit]

Damn bots got all the 4090s

[xAcid9]

On 10/12/2022 at 9:41 PM, leadeater said:

Capacity is a huge factor in hit rate, small caches have worse hit rate without significant optimization on the software side that specifically targets the architecture. Capacity is the brute force way of increasing hit rate basically universally, not "optimization" required.

That's why I mentioned what kind of software optimization Nvidia do in the software side to optimize the huge L2. This is Nvidia we're talking about, they've done software wizardry in the past to optimized their GPU. Maybe they need large L2 for a new type of caching or a new version of tiled caching that they introduced since Maxwell. 

On 10/12/2022 at 4:29 PM, xAcid9 said:

I remember one of Maxwell secret sauce was tiled based rendering + much bigger L2(2MB) compare to Kepler(256kb).

I wonder what Nvidia do in software side to utilize that huge L2. Where is David Kanter? 

 

[leadeater]

5 minutes ago, xAcid9 said:

That's why I mentioned what kind of software optimization Nvidia do in the software side to optimize the huge L2

Software side means the games, or rather the game engines.

 

6 minutes ago, xAcid9 said:

This is Nvidia we're talking about, they've done software wizardry in the past to optimized their GPU. Maybe they need large L2 for a new type of caching or a new version of tiled caching that they introduced since Maxwell. 

Well it doesn't actually matter though, it could be L2, L3 or L4. These caches are simply a store of data so you don't have to go back to VRAM so no matter what they want to do any cache layer will support whatever they want to do fundamentally.

 

The problem however is fundamental, larger caches are higher latency. Higher cache tiers are, typically, higher latency than lower tiers etc etc. 

 

On 10/12/2022 at 9:29 PM, xAcid9 said:

It's L3 vs L2 though? Much faster bandwidth(3x-4x faster) and lower latency.

I was responding to this, the above simply doesn't really matter and it's likely not even the case anyway. As per evidence already in this topic anyway (provided by yourself from memory).

 

There are also cache layer labelling difference between Nvidia and AMD also, so talking about L0, L1 and L2 etc aren't necessarily the same thing between the two so be careful of that.

 

Quote

With the newer test, RDNA 2 and Ampere have similar latency to their fastest cache, but Ampere’s L1 is larger than RDNA 2’s L0. Nvidia can also change their L1 and shared memory allocation to provide an even larger L1 size (up to 128 KB according to the GA102 whitepaper). But for OpenCL, it looks like Nvidia chose to allocate 64 KB as L1.

 

Additionally they are most likely the same SRAM type and construction anyway and have very similar characteristics. The penalty AMD gets over Nvidia is basically just around having an extra cache layer at all and the implications of that. The caches are also segmented so scan/walks will be nice and fast even though the total cache is large, I would also assume Nvidia is also doing this and it's just not literally shown in the diagrams.

 

TL;DR L2 vs L3 here doesn't really matter and it's likely not even that much different on the latency and bandwidth side anyway.

 

RDNA1

 

RDNA2

 

Ampere

 

Ada

[ewitte]

One I ran against my 3080 (had a different CPU). Raytracing feature test.  At 82% power target, +150/+200