- Joined
- Aug 28, 2019
Now that I have a chance to look at this thing... it's a 20-core Arm CPU, with a powerful GPU, and 128 GB of LPDDR5X for running LLMs. Sound familiar?
GPUs & CPUs & Enthusiast hardware: Questions, Discussion and fanboy slap-fights - Nvidia & AMD & Intel - Separe but Equal. Intel rides in the back of the bus.
- Thread starter Smaug's Smokey Hole
- Start date
- Joined
- Dec 12, 2022
Just in raw core count, the 5070 doesn't beat the 4070 super according to Gamers Nexus. 7168 cores vs the 5070's 6144 cores. I'm not upgrading. Maybe the higher bandwidth gives it a edge. I also know, at least at 1080p 60 FPS, every game I throw at my 4070 Super combined with my i9 12900K runs natively without DLSS or any of the other shit. Seriously Nvidia, the fuck?
- Joined
- Feb 16, 2024
It’s useless to count cores and compare like for like before the cards release, just wait for benchmarks and reviews to inform your decisions.
- Joined
- Sep 10, 2021
The compute comparisons are done on fp8 for the 4xxx series and fp4 for the 5xxx series. What does this mean exactly? Is this just for AI training? I do compute for work (archviz) and I am not sure if this matters to me at all.
2000$ for the 5090 is absolutely fucking bonkers.
I am curious but skeptical about all their other tech. The latest slop games are already blurry and shimmery as it is.
2000$ for the 5090 is absolutely fucking bonkers.
I am curious but skeptical about all their other tech. The latest slop games are already blurry and shimmery as it is.
- Joined
- Aug 28, 2019
Wait for reviews. But also don't play video games in current year. The 5090 is a discount AI workstation card that is probably worth the $2k, but it can also quadruple your fake frames.
No idea about Archviz. What did previous benchmarks show?
- Joined
- Nov 15, 2021
You'll need a server CPU, and it's likely NVIDIA is soft-locking the cards from being used that way so they can sell high-margin server products.
- Joined
- Sep 10, 2021
This is the graph they show on their website, forgot to add it on my earlier post
I rely on Blender Opendata for my estimations for compute performance. The top performers are about what you would expect, but there are some outliers (4070ti being significantly better than 3090 for example, and the 4070 being a close equivalent to the 3090)
- Joined
- Aug 28, 2019
Looks like they drank some of Nvidia's Flavor Aid and spat out nonsense. The 4090 used 21 Gbps memory, 5090 uses 28 Gbps (+33%), 5080 uses 30 Gbps (+43%).
NVIDIA GeForce RTX 5090 Founders Edition is the only 5090 SFF-Ready model
Nvidia's reference designs may be far superior to anything coming out of AIBs.
- Joined
- Dec 18, 2022
Maybe it's time for me to throw out the old Broadwell server running my house, and build a whitebox threadripper.
- Joined
- Dec 26, 2014
I think we are hitting the moment that there is no any gains to get without gimmicks like frame generation and you have to choice between
You play AAA slop? 5080 you dont play slop? 5070
The 5090 was priced like that because nvidia probably believe that will deter scalpers but even they are showing in their own benchmarks that the gains compared with the 4090 are with gimmicks, without the gimmicks they are showing a 20% gains at best
AMD need to start reconsidering their strategy because at this rate there is no reason to get their cards unless they go for like 400 buck top
You play AAA slop? 5080 you dont play slop? 5070
The 5090 was priced like that because nvidia probably believe that will deter scalpers but even they are showing in their own benchmarks that the gains compared with the 4090 are with gimmicks, without the gimmicks they are showing a 20% gains at best
AMD need to start reconsidering their strategy because at this rate there is no reason to get their cards unless they go for like 400 buck top
Mister Fister
kiwifarms.net
- Joined
- Nov 17, 2024
Hot take: The APU/Handheld PCs are advancing rapidly enough to make the 60 series and below pointless.
- Joined
- Sep 7, 2019
it certainly is part of the shilling material, I'm seeing an awful lot of BRO 4090 PERFORMANCE FOR $500 conveniently not mentioning the 4x framegen. but hey graph is higher that's all that matters.
fake input for fake frames. at some point people will pay $3000 to watch a game play itself instead of watching a stream...
Last edited:
- Joined
- Dec 12, 2022
That's not that much of a hot take. I think the wider market would agree since it seems like a lot of companies are trying get into handhelds.
- Joined
- Jul 18, 2019
It was already known that Blackwell wasn't going to be as big of a step as Lovelace since both of them are fabbed on roughly the same process node. The performance uplifts are coming from refinement of the underlying architecture and faster VRAM. Nvidia and AMD have been reluctant to use TSMC's 2nm process because of costs but apparently Nvidia is considering going to Samsung so we'll likely see some big changes for the 6000-series.
The 5090 is priced that way because the 4090 already effectively has a price of $2000 and there's no more left on the market since they stopped production.
Correct. The 5070 looks like it might trade blows with the 4080 which puts AMD in a weird situation. At the very least, they cannot charge more than $549 for flagship.
- Joined
- Dec 18, 2022
There's no point to having a 90 series for the average consumer or even the enthusiast at this point.
I think the 90 series' new market share is people and small businesses who have a standard ATX/ITX desktop/workstation/tower server, and want to do mid to high-level ML inference/rendering at a lower cost of entry than an enterprise card. It's the best tool for the job for an independent contractor or small business that needs GPU compute. For a small business, 2k for an important tool is negligible if it can bring in more revenue.
It doesn't make sense as a server card because their coolers make them take up 3 slots even with the FE. Compared to their enterprise cards with 2 Slots with passive cooling on Rack servers.
I think the 90 series' new market share is people and small businesses who have a standard ATX/ITX desktop/workstation/tower server, and want to do mid to high-level ML inference/rendering at a lower cost of entry than an enterprise card. It's the best tool for the job for an independent contractor or small business that needs GPU compute. For a small business, 2k for an important tool is negligible if it can bring in more revenue.
It doesn't make sense as a server card because their coolers make them take up 3 slots even with the FE. Compared to their enterprise cards with 2 Slots with passive cooling on Rack servers.
- Joined
- Dec 12, 2022
The 7950X3D, yes. Having the VCache on only one CCD let some games run on the non-VCache set of cores. I think of lot of the problems were fixed. But since the 9950X3D does the same thing, who knows if those problems will pop up again.
- Joined
- Dec 26, 2014
I still find stupid that the not ridiculous big version is Founder edition exclusive
- Joined
- Jul 10, 2023
I wonder how the temps will differ depending on how this new FE model is mounted.
- Joined
- Nov 15, 2021
Just generating even more frames is extremely underwhelming. AMD recently unfucked FMF in Adrenaline so I could experiment with it in Diablo 4 and Modern Warfare a bit, and my conclusion is it's fine enough if your base frame rate is 60 fps, but Lossless Scaling's FSG1, which doesn't use any fancy motion vectors at all, just purely what it can get from the pixels on screen, looks just as good.
I think the reason is that if you're running fast enough for frame generation to actually work, frames are flipping by you too fast for your eye to notice much difference between one algorithm and another. This is very different than upscaling, where naive interpolation (like FSR1 or bilinear) looks like shit compared to inferencing. So this is a case where I think NVIDIA's overshot the value that inferencing actually provides.
If you play AAA slop, its main target is going to be the PS5. That's a 6650 XT, more or less. Anything better will be fine. People vastly overestimate how much processing power they need to be fine.
I think the reason is that if you're running fast enough for frame generation to actually work, frames are flipping by you too fast for your eye to notice much difference between one algorithm and another. This is very different than upscaling, where naive interpolation (like FSR1 or bilinear) looks like shit compared to inferencing. So this is a case where I think NVIDIA's overshot the value that inferencing actually provides.
If you play AAA slop, its main target is going to be the PS5. That's a 6650 XT, more or less. Anything better will be fine. People vastly overestimate how much processing power they need to be fine.