nvidia Rtx 4090 is a monster (Official Benchmarks)

[Shzzit]

Dude, a lot of the reviewers have wobbly ass fans.  You guys notice that?

 

For 1600 bucks I’d be pissed.

[Shzzit]

8 hours ago, shermantanker said:

The performance gains across the board are insane and fingers crossed I can snag a Strix from Newegg tomorrow.

Can I ask why am strix over an fe? Strix is 400 more dollars.  I wanted a strix but ya, not sure.

[BiG StroOnZ]

Here's a review roundup:

 

https://videocardz.com/140065/nvidia-geforce-rtx-4090-founders-edition-review-roundup 

[xg32]

looks like the first 4k/120hz card is here, and it likely won't lose much performance when undervolted and capped to about 350w

[Shzzit]

New msi afterburner patch you can adjust 4090 voltage, people are seeing 3200-330hz on air.  This is insane if the posts are real.

 

Hitting around 650 watts though hahaha

[bowrilla]

44 minutes ago, Shzzit said:

New msi afterburner patch you can adjust 4090 voltage, people are seeing 3200-330hz on air.  This is insane if the posts are real.

 

Hitting around 650 watts though hahaha

Fun for the benchmark run, pretty senseless for everyday usage.

[xAcid9]

I still can't believe the amount of L2 cache this thing have, 72MB compare a "mere 6mb" on 3090s. That's probably where most of the efficiency come from because of how power hungry the GDDR6X is. Infinity Cache look like childs play now imo. 

 

Also no game support the new shader reordering thingy atm iirc. 

[Spotty]

7 hours ago, Kisai said:

1. Sell the mainstream 450w "3 slot" parts with a x16 extension and mounting bracket to allow it to be used on chassis that have a dedicated "GPU" space. Below: Bequiet Silent Base 802

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.

[Stahlmann]

11 hours ago, CTR640 said:

You've got the AIB 3080 that caught my attention. How are the temps and performance? The 4090 is indeed a beast and a true 4K/120 GPU but it's not an option for me lol. Otherwise it means I have to change my case aswell as the PSU and all.

Stock it was pretty loud imo but i'm very particular about noise. And it was sitting behind a 360mm radiator that also had a pump/reservoir combo mounted, so airflow to the GPU was pretty bad. I mainly bought it because it uses a reference PCB and i wanted to include it in my custom water loop. Performance is as expected from a 3080.

[Stahlmann]

6 hours ago, xg32 said:

looks like the first 4k/120hz card is here, and it likely won't lose much performance when undervolted and capped to about 350w

From what Der8Bauer tested the best option is adjusting the power limit and leaving voltages alone. Undervolting using MSI Afterburner typically ended up in a larger performance drop. Still, limiting it to 70% power or so only saw a performance drop of around 5% in games, 10% in synthetic benchmarks. That means it'll run at around the same 320W TDP as my 3080 while having almost double the performance in most tasks. In terms of performance per watt this card is absolutely nuts.

[porina]

1 hour ago, xAcid9 said:

I still can't believe the amount of L2 cache this thing have, 72MB compare a "mere 6mb" on 3090s. That's probably where most of the efficiency come from because of how power hungry the GDDR6X is. Infinity Cache look like childs play now imo. 

Isn't Infinity Cache on RDNA2 up to 128MB? They way I see it, cache helps mitigate against not enough ram bandwidth so you get less loss in performance, especially with the relatively narrower bus this time around. It is balanced by being relatively expensive in silicon area. We'll have to wait for more in depth details on any execution changes as I don't think we can put it all down to the cache.

[leadeater]

5 minutes ago, porina said:

Isn't Infinity Cache on RDNA2 up to 128MB? They way I see it, cache helps mitigate against not enough ram bandwidth so you get less loss in performance, especially with the relatively narrower bus this time around. It is balanced by being relatively expensive in silicon area. We'll have to wait for more in depth details on any execution changes as I don't think we can put it all down to the cache.

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

[porina]

39 minutes 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.

They will design and optimise for the system, rather than one subsystem in isolation. I did a little digging and found the following:

 

 

This is based on an official AMD slide, with an estimated scale added for better indication. 40 series may not exactly follow this but the trends should be similar.

 

Source: (worth a look in the whole Twitter thread as there's more)

 

 

 

 

[xAcid9]

1 hour ago, porina said:

Isn't Infinity Cache on RDNA2 up to 128MB? They way I see it, cache helps mitigate against not enough ram bandwidth so you get less loss in performance, especially with the relatively narrower bus this time around. It is balanced by being relatively expensive in silicon area. We'll have to wait for more in depth details on any execution changes as I don't think we can put it all down to the cache.

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

 

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? 

[porina]

7 minutes ago, xAcid9 said:

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

I was struggling to find a good write up of Infinity Cache to refresh myself on how it fits in. Based on CPUs, I find overall capacity a good general indication, more so than focusing on bandwidth and/or latency numbers of each tier.

[xAcid9]

1 hour ago, porina said:

I was struggling to find a good write up of Infinity Cache to refresh myself on how it fits in. Based on CPUs, I find overall capacity a good general indication, more so than focusing on bandwidth and/or latency numbers of each tier.

It's low level caching, latency and hit rate is just as important if not more than capacity imo.  

 

Apparently on Ampere the L2 latency is higher compare to RDNA2 L2 and L3.

I hope the guys at ChipHell release memory latency graph for 4090 soon.

[porina]

41 minutes ago, xAcid9 said:

It's low level caching, latency and hit rate is just as important if not more than capacity imo.  

As it is a tiered system you would expect them to balance it to their needs, hence I wouldn't look too hard at each layer in isolation. Hit rate is proportional to size though.

 

41 minutes ago, xAcid9 said:

Apparently on Ampere the L2 latency is higher compare to RDNA2 L2 and L3.

I hope the guys at ChipHell release memory latency graph for 4090 soon.

Would be interesting to see that data.

 

I recently became aware of the following tool to measure bandwidth/latency for CPUs. Anyone know of similar for GPUs? 

 

 

[ewitte]

4 hours ago, Stahlmann said:

From what Der8Bauer tested the best option is adjusting the power limit and leaving voltages alone. Undervolting using MSI Afterburner typically ended up in a larger performance drop. Still, limiting it to 70% power or so only saw a performance drop of around 5% in games, 10% in synthetic benchmarks. That means it'll run at around the same 320W TDP as my 3080 while having almost double the performance in most tasks. In terms of performance per watt this card is absolutely nuts.

That is entirely card dependent some are already at near the best voltages they can be.  It is like a lotto.  

[Shzzit]

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

[jasonvp]

13 minutes ago, Shzzit said:

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

I've heard stories of folks getting AIBs.  Unfortunately, NVidia decided (again) to lean on BestBuy in the US to sell their FE cards.  BestBuy's site alternates between "Sold out" and "Sold out, and we don't ship this product".  The depletion happened in less than 5 seconds, but it's still not clear whether BestBuy will actually ship, or require you to pick up.  Either way, I'll bet we see a lot of these on Ebay later today.

Embrace the scalp.  It's going to happen again.

[xAcid9]

2 hours ago, porina said:

As it is a tiered system you would expect them to balance it to their needs, hence I wouldn't look too hard at each layer in isolation. Hit rate is proportional to size though.

Would be interesting to see that data.

I recently became aware of the following tool to measure bandwidth/latency for CPUs. Anyone know of similar for GPUs? 

My bad, the data was from Chip and Cheese website and not Chiphell. https://chipsandcheese.com/2021/05/13/gpu-memory-latencys-impact-and-updated-test/

 

 

Also from what I read, the tool that they used to bench gpu memory latency test was taken from this patreon. https://www.patreon.com/nemezor
Not sure how good/reliable the tool though. 

[leadeater]

3 hours ago, xAcid9 said:

It's low level caching, latency and hit rate is just as important if not more than capacity imo.

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.

 

Just look at the 4k resolution line on the AMD slide with the hit rates added (estimated), it's basically a linear line with capacity increase. That should tell you all you need to know for how important having the necessary capacity is.

 

That said a new type of cache does itself need to be optimized for but you can probably do most of that in the silicon design and drivers.

 

3 hours ago, xAcid9 said:

Apparently on Ampere the L2 latency is higher compare to RDNA2 L2 and L3.

Larger the capacity of the cache the higher the latency is has, like for like.

[ewitte]

13 minutes ago, 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.

Take it 24GB of cache would mean a 100% hit rate...

[shermantanker]

15 hours ago, Shzzit said:

Can I ask why am strix over an fe? Strix is 400 more dollars.  I wanted a strix but ya, not sure.

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.

[leadeater]

36 minutes ago, ewitte said:

Take it 24GB of cache would mean a 100% hit rate...

haha well no sadly, not unless cache entries are never retired and always updated but then it's not cache and just the vram