News GPU Makers Slash Prices of GeForce RTX 4070 to $549

[Order 66]

same here. im ready to move on from my 1080ti and its gonna be an amd card after the next gen comes up.

the 1080ti is still a beast at 1080p though, why do you feel the need to upgrade?

 

[palladin9479]

makes sense given speed difference between storage and onboard vram. I wonder if it can compress it like windows can.

Depends on the API I think, textures take up the bulk and they are already compressed. I just know that people look at the GPU-Z's "memory in use" and freak up thinking that is how much that game "needs", when in reality it's similar to Windows Resource Monitor "memory free".

 

[AgentBirdnest]

Pleasantly surprising news for me, because this is the card I'm saving up for. : )
I really didn't think the price would drop. I can't even remember the last Nvidia price drop before this month...

 

[Colif]

Depends on the API I think, textures take up the bulk and they are already compressed. I just know that people look at the GPU-Z's "memory in use" and freak up thinking that is how much that game "needs", when in reality it's similar to Windows Resource Monitor "memory free".

it would just dump data on shutdown, since its only read only. Unlike windows that saves it to storage at shutdown. I am using 1.7gb, wonder what its actually got in there... you would think there are tools that can access the vram to see what is in there...like rammap

 

[JTWrenn]

"In response to AMD's Radeon RX 7800 XT?"

Why else would they do that!!

Quite a few reasons not the last of which would be a possible announcement of a 5000 series. Lower demand, all sorts of things.

 

[kiniku]

$50.00 less is not a slash. It's a lowering. Click bait headline.

 

[nitrium]

It should b e imo $350-$400ish based on the RTX 2060 (6GB) pricing from 4 years ago. The 4070 is the first RTX 4000 series that has the double the performance and double the VRAM of the 2060 (released in Jan 2019).... after well over 4 years you'd expect pricing reflective of the advancements from R&D, as opposed to paying almost double the price (at least here in NZ) for double the performance.

 

[oofdragon]

$500 7800XT > $500 4070.......... raster and VRAM way more important than fake frames and traced gimmick. The 4070 isn't worth more than 450

 

[footman]

Pretty much a redundant card now that the 7800XT has been released at $499. The free copy of Starfield, just added insult to the injury. These cards are aimed at 1440P, so the discussion around RT is irrelevant as neither card puts out decent framerates with RT enabled at this resolution. DLSS and FSR are also not worth using as both result in visual glitches. For now, the only metric I am interested in is pure rasterization and the 7800XT when overclocked gets mighty close to the 4070Ti in rasterization!
Nvidia needs to cut the price of the 4070 to $480 and the 4070Ti to $649 in order to be competitive with AMD.

 

[palladin9479]

it would just dump data on shutdown, since its only read only. Unlike windows that saves it to storage at shutdown. I am using 1.7gb, wonder what its actually got in there... you would think there are tools that can access the vram to see what is in there...like rammap

The game has to do it itself, not sure if an external entity can actually get a list of objects and references from outside the executing program. Maybe if the driver exposed that via an external API.

Programming DirectX with COM - Win32 apps

Programming DirectX with COM.

learn.microsoft.com

Managing a COM object's lifetime​

When an object is created, the system allocates the necessary memory resources. When an object is no longer needed, it should be destroyed. The system can use that memory for other purposes. With C++ objects, you can control the object's lifetime directly with the new and delete operators in cases where you're operating at that level, or just by using the stack and scope lifetime. COM doesn't enable you to directly create or destroy objects. The reason for this design is that the same object may be used by more than one part of your application or, in some cases, by more than one application. If one of those references were to destroy the object, then the other references would become invalid. Instead, COM uses a system of reference counting to control an object's lifetime.

An object's reference count is the number of times one of its interfaces has been requested. Each time that an interface is requested, the reference count is incremented. An application releases an interface when that interface is no longer needed, decrementing the reference count. As long as the reference count is greater than zero, the object remains in memory. When the reference count reaches zero, the object destroys itself. You don't need to know anything about the reference count of an object. As long as you obtain and release an object's interfaces properly, the object will have the appropriate lifetime.

Properly handling reference counting is a crucial part of COM programming. Failure to do so can easily create a memory leak or a crash. One of the most common mistakes that COM programmers make is failing to release an interface. When this happens, the reference count never reaches zero, and the object remains in memory indefinitely.

Programs are responsible for cleaning up stuff, if it's not explicitly deleted then it hangs around until it ages out or some sort of GC kicks off and evicts it from memory.

 

[wingfinger]

Recently, I was taking a look at a new card. It makes sense to me to plan for a potential 1440p monitor upgrade in the future, so lower cards are a problem.

The quandary for 4070 is:

Since the cards are so expensive, shouldn't you buy the card you want which is probably a 3 fan Asus or MSI, hopefully with an old-style power connector?

Or do you, cut your losses ($) by buying one on sale, including maybe a 2 fan version, with the hope that it will still be reliable for long time and won't end up having a problem with noise or temps?

 

[palladin9479]

If you absolutely have to buy a card, go with AMD this time around. Much better price to performance ratio.

 

[tommtajlor]

Still too much for a 3060 successor card. 350, very max 400 usd.

 

[Alvar "Miles" Udell]

$550 for a card that can barely play Starfield at 60fps at 1920x1080? No thanks!

 

[razor512]

VRAM amount is important, especially with the current gen games. For example even at 1080p games like Ratchet & Clank: Rift Apart are using nearly 12GB at 1080p very high settings and ray tracing disabled.

Though the game is a also one of the edge cases in terms of being able to utilize shared memory where it can pull around 3GB before you start getting frame time inconsistencies, thus an 8GB card like an RTX 3070 can run maintain decent performance while allocating a few GB of shared memory. Though it is likely how the game loads data.
When you go to a planet, with multiple rifts, the game loads them all with each rift literally randomly floating under the main map, hence virtually no extra loading even as you traverse multiple different locations in a chase scene. But if you have a card that has 12GB,+ of VRAM, those transitions are buttery smooth, while if there is shared memory, there will be some spikes in frame timing, though it is not much of an issue since it doesn't happen in combat. Furthermore the load on the SSD is not very high since it preloads a lot of the content needed(basically lots of empty sets where all it needs to do is load in the NPCs for each rift

Anyway even when a game is not going for photorealism, we are at a point where 1080p is already pushing the limits of 12GB.

While 16GB is not far off, we can at least be confident that games designed for multiple platforms will not be designed around expecting 16+GB of VRAM unless you push some very high resolutions, since they will likely design around the constraints of the current gen consoles. Thus we will likely see 16GB remain very usable at least until a new console generation comes out.

 

[Colif]

PS6 is all rumours now, no one knows. I see its meant to be based on RDNA4 but last I heard that might not have many high end GPU in it... but that is also based on rumours.

It would make more sense to me to stay at 16gb for longer to make some profit on the games already created. Console generations aren't normally only a few years. Games take longer to make now so less change means they might actually get released sooner and not need to change for new features just before launch.

The Graphic cards will improve and maybe even the ability of devs to make games that can run in native resolutions. No features required.

 

[btmedic04]

An 8% price cut is not a slash. Take off $125 and I'll consider that a price slash

 

[palladin9479]

VRAM amount is important, especially with the current gen games. For example even at 1080p games like Ratchet & Clank: Rift Apart are using nearly 12GB at 1080p very high settings and ray tracing disabled.

No it's not, it just doesn't immediate evict data from graphics memory when it is no longer needed. If there is no more room and new data is requested, the oldest asset with the lowest reference count will get evicted. The only time there is ever performance issues is if you are attempting to use more data in a scene then the graphics card has room for, meaning you would need a room / area with more then 12GB worth of data. Otherwise the engine will just evict the older unused data to make room for the new data.

That is how all the new graphics engines work. In the past a game would have to manually load assets whenever you loaded into an area, then evict / reload new during area transitions as part of a load screen. The developers would have an "expected GPU memory" target they would be working around and that value is what we say the game "needed". Modern engines did away with this and instead adopted the same model that is used for main memory with unused memory acting as a cache. Unfortunately programs like GPU-Z only show the maximum amount with data allocated, not how old that data is or if it's actually being used. Instead we'd have to test for stuttering effects, when the graphics engine has to unload stuff that is actually being used, load new stuff, then unload that new stuff to reload the stuff it just evicted.

 

[phxrider]

I bought a couple of 6800xt's during the crypto craze because I was mining, they were available, and the prices were justified for how much profit I was making... These were added to the Nvidia cards I already had. I also game with them, and I have to say I am really impressed with the AMD cards and am pretty solidly on team red at this point. Other than if you have to have a 4090 at nearly twice the cost, AMD is the way to go if you value price versus performance. The media still seems to be biased toward team Green, but as a real world user I can say AMD is a better value for the money, and contrary to reputation the drivers are fine and maybe more stable than Nvidia, and the adrenaline software package is nice especially if you have an AMD CPU, then it all integrates together (but it works fine on Intel, I have that combo too). And the 7900 series seems like a genuine upgrade over the previous generation, unlike the lower series.

 

[grogi]

Had both AMD/Ati and Nvidia cards in the past but went with Nvidia this time based on superior RT. Anyways RAM size is important but sometimes more doesn't help. Look at the performance of 460 TI (16 GB) vs 460 TI (8 GB) one can hardly tell them from each other.

 

[scottscholzpdx]

Still a tremendously bad value compared to the 6800XT/7800XT.

 

[razor512]

No it's not, it just doesn't immediate evict data from graphics memory when it is no longer needed. If there is no more room and new data is requested, the oldest asset with the lowest reference count will get evicted. The only time there is ever performance issues is if you are attempting to use more data in a scene then the graphics card has room for, meaning you would need a room / area with more then 12GB worth of data. Otherwise the engine will just evict the older unused data to make room for the new data.

That is how all the new graphics engines work. In the past a game would have to manually load assets whenever you loaded into an area, then evict / reload new during area transitions as part of a load screen. The developers would have an "expected GPU memory" target they would be working around and that value is what we say the game "needed". Modern engines did away with this and instead adopted the same model that is used for main memory with unused memory acting as a cache. Unfortunately programs like GPU-Z only show the maximum amount with data allocated, not how old that data is or if it's actually being used. Instead we'd have to test for stuttering effects, when the graphics engine has to unload stuff that is actually being used, load new stuff, then unload that new stuff to reload the stuff it just evicted.

While it is hard to tell exactly how much VRAM is actively being used vs allocated, you can get clear indicators of when a spillover happens and system memory is actively being used to supplement the VRAM. For example, if a game gradually goes from allocating 90% of the VRAM, to virtually all of it, in addition to 1-2GB of shared memory, but the PCIe bus usage doesn't increase, then you can reasonably infer that much of that spillover is caching, which is used in some cases, for example, the PC version of Ratchet & Clank: Rift Apart where it uses a lot of extra VRAM, but a portion of it is not actively used all of the time since thee various "rifts" are different rooms and different versions of the same area simply placed under the map.

On the other hand, there are games like the pre performance patch version of hogwarts legacy where it will on an 8GB card, 2-3GB can spill over to shared memory, and you will see the PCIe bus usage jump from around 4-10%, to 50+% (saturating fully in one direction, = 50% usage, and saturating both directions will get it to 100%). Other signs of the spillover are reduced GPU power consumption to go along with the significant drop in frame rate. After the patch, the game would simply stop loading in all of the textures in order to avoid the spillover into system memory, and that effectively eliminated nearly all frame rate drops, and allowed the game to remain smooth on an RTX 3070, even when maxed out.

No amount of VRAM bandwidth can make up for not having enough VRAM. If there isn't enough memory to hold the full working set, then the working set either needs to spillover and take a large performance hit as you effectively end up with part of the working set running at a tiny fraction of the speed of the VRAM, or you compromise it by simply not loading all of it.

This is why in the past, you saw cards like the Radeon R9 Fury X lose out to a number of lower spec cards with more VRAM even though many of the cards it lost to had less than half the VRAM read/ write speed of the HBM that was on the R9 Fury X. Simply put, the performance hit from having part of the data the GPU needs to work with trickling in at 20-25GB/s lead to many issues that cards with GPUs that were a little slower but with more RAM didn't have. The R9 Fury X aged very poorly compared to its peers.