As usual Sander an excellent article

I find it a bit worrying that some new video cards are looking at using system memory. Ok I do get it that Turbo Cache is offering a boost in performance on a cheaper hardware platform, but at the expense of system RAM.

This worries me a bit because frankly I run out of system RAM enough as it is (I edit large image and video files) and don't want my graphics card poaching it.

Does the advent of Turbo Cache mean that high RAM video cards are likely to become a thing of the past or is it just a way of producing a cheap high RAM substitute to run alongside the more conventional offerings?

Tim


AMD XP2800+
GeForce FX5600
Onboard Audio (cloth ears)
1GB PC2700 RAM
Gigabyte GA-7VT600 MOBO
Silicon Image SiI 0680 Ultra-133 Medley ATA Raid Controller
Pinnacle DC10 Video capture board
7200 rpm Samsung SP4002H 40GB HDD (Primary)
7200 rpm Western Digital WD 2000JB 200GB (2ndary)
Toshiba SDM-1222 DVD ROM
Pioneer DVR-106D DVD-RW
Liteon LTR-48246K 48x24x48 CDRW

The latter Tim, just for the reasons you outlined this is an entry-level solution for now, also because there's not enough bandwidth to fuel the bandwidth hungry graphic processors in the mid-range and high-end segment. And thanks for the compliment, I try to bring you the gist of the matter, without too much complicated technical fluff around it. If there's something interesting that needs to be elaborated on rest assured I will do that though.

If you failed to notice, the first two cards released with the caching architecture will feature 128MB and 256MB.
These cards will only use system memory as an extended pool; one that it'll rarely tap into.

Nvidia's site says it's strictly being limited to the 6200 budget end, so we'll never see an enthusiast card making significant use of this feature, if in fact they decide to include it at all. It's not a bad feature to have, but I doubt it'll be made use of with graphics cards that have more than 64MB of memory.

One more thing we haven't factored in is what this does for laptops. The performance is currently lacking, so this could definitely tackle the lack of bandwidth without sacrificing motherboard realestate.


The fact of the matter is, graphics cards don't use a fraction of the memory they're outfitted with. The human eye can't distinguish the difference between 128MB and 256MB, and games don't care much either.
Graphical memory is only a factor in benchmarks... and even then they aren't being cached with usable data.


I don't know if you knew this, but your graphics card doesn't affect video editing, encoding, and professional 3D rendering whatsoever. Most professional applications don't even recognize the 6800's onboard video processor yet.

As far as 3D rendering is concerned, if you're displaying it on the screen, the graphics card is made use of. However, if you're doing a full scene render, that's entirely processor and memory bandwidth reliant.

Sean, it depends on the game you play. If you stick to old stuff you're probably right but if you try to turn doom 3 up all the way it gives you a warning about needing atleast 520MB just for textures. Also Id made the engine scale itself down going from a 256MB to a 128MB and below card.

The thing that I don't like about this feature is that It may put more strain on the rest of the system, I would hate to think that I was loosing framerates because the video card was sturating my memory controllers available bandwidth, and the CPU was having to wait to get what it needed.

Ok I had to respond to this, one if you don't know what your talking about stop spreading fud. Most videocards use onboard memory for rendering, and when that runs out which is does very fast even at 1024x768 they have to swap to main memory, the same way your ramm is swaped to you hard drive's swap file. That's the simplied version, but anyone who says your videocards NOT using all it's memory should made to walk in circle for an hour, for misleading you. The only games that don't use it are 2d prerendered games like freecell, but if all your playing is freecell why buy a videocard at all? Second far more annoying all high end rendering software reconizes Quadro series, which is nvidia workstation cards, and dump as much of the work load on the card as possible. Also if your cpu did real time rendering you would be watching a slideshow, CPU are not designed for that, they do batched rendering, which means their doing math and simply do it a task at a time so that the math for one animation may be done before another that will display at the same time. When the batch render is done your videocard still has to decode the file and render it in real time to the screen.

"Sean, it depends on the game you play. If you stick to old stuff you're probably right but if you try to turn doom 3 up all the way it gives you a warning about needing atleast 520MB just for textures. Also Id made the engine scale itself down going from a 256MB to a 128MB and below card."

That 520MB is an accumulation of a small fraction of video memory and the majority of it is system memory.


"The thing that I don't like about this feature is that It may put more strain on the rest of the system, I would hate to think that I was loosing framerates because the video card was sturating my memory controllers available bandwidth, and the CPU was having to wait to get what it needed."

It's putting strain on a component that exists but isn't being used effectively. The AGP bus was never used to it's full potential, at least this caching architecture allows PCIe to be put to good use. Although, I can think of better uses.



Bob, name one game that has enough individual textures with sizes ranging from 5MB to 10MB being rendered all at once.

The maximum amount of simultaneous textures in Doom III was about 50 in certain areas, and most of them are compressed to mere kilobytes. Not to mention, with bump mapping(which Doom III uses a lot if you haven't noticed) the only 3D models are objects, not enviornments. So, you have a lot of textures that "look" detailed, but have simply been bump and light mapped, on top of already being compressed.


I don't think you understand how the professional 3D animation industry works.

Production studios such as Dreamworks, Pixar, and ILM start on workstations, which yes, have Quadros or better yet, 3D Labs Wildcat 4 and Wildcat Realizm (Dual GPUs and up to 640MB onboard memory).


After they finish raw cuts, which entail extensive character and enviornment modelling, keyframe animation(or motion capturing, depending on the project), and overall synchronization(which includes synchronizing characters to voice actors, and enviornments to ambient sounds), they then throw the entire project on a batch renderfarm.

RENDERFARMS DO NOT USE GRAPHICS CARDS! Renderfarms rely on large amounts of system memory, coupled with fiber channel disk storage and RISC based processors.

Pixar's renderfarm is based on Sun UltraSPARC, while Dreamworks facillitated HP PA-8800 based farms, both of which are RISC processors. I don't know about you, but I don't know any chipset that supports either processor and an AGP slot. Sun does have it's own professional graphics card that supports UltraSPARC boards, but they cost an exorbitant amount of money... and when you multiply that by 200+ systems, you're significantly over budget.

Not to mention, there isn't much advantage in using FrameLock or GenLock architectures, because clusters do such processing more effectively using just processors.


Furthermore, professional batch rendering software, such as Pixar's Renderman and Avid MentalRay don't even consider the graphics card- they work strictly from the processor. Nvidia Gelato is one of the only batch renderers that considers the graphics card(namely Quadro FX) before the processor.

Phew!

Thanks guys, I'm working on progressing to high end cards soon, wife permitting, (current on is really pants for some reason!) and didn't want to see my system RAM eaten away wholesale.

Am I getting my Maths right here? I get that it takes 18MB of video memory to render each frame of graphics at 1024x768, in 24 bit colour (3 colours 256 shades of each) not counting space taken up by the actual code doing the rendering. OUCH!

Cheers

Tim

AMD XP2800+
256MB GeForce FX5600
Onboard Audio (cloth ears)
1GB PC2700 RAM
Gigabyte GA-7VT600 MOBO
Silicon Image SiI 0680 Ultra-133 Medley ATA Raid Controller
Pinnacle DC10 Video capture board
7200 rpm Samsung SP4002H 40GB HDD (Primary)
7200 rpm Western Digital WD 2000JB 200GB (2ndary)
Toshiba SDM-1222 DVD ROM
Pioneer DVR-106D DVD-RW
Liteon LTR-48246K 48x24x48 CDRW

The GPU doesn't keep a cache of any of the engine, however, it does recieve and cache instructions, just like a processor. None of the instructions that the GPU interprets are cached on video memory, just as none of the CPU's instructions are cached in ram. Both have their own cache for this.

The only viable means for video memory is that it's suppsed to act as a fast cache for textures and models that need to be processed, but software is written in a way where it doesn't use this to its advantage.

Furthermore, Doom III's textures were insanely low resolution, so even without compression, they consumed very little memory and could even be called on the fly.

One question I have:

What is the difference between this TurboCache thing and AGP aperture RAM (except, that AGP aperture RAM uses the AGP bus)?

I thoiught, that something similar to AGP aperture RAM already exist for PCI-E. is this correct?


"If you failed to notice, the first two cards released with the caching architecture will feature 128MB and 256MB.
These cards will only use system memory as an extended pool; one that it'll rarely tap into."

Doom 3 has a mode, which can only be activated on GPUs w/ 512MB. this mode uses uncompressed textures. Some said, who used this mode w/ 512Mb cards or SLI said, that they recognized a difference to the best mode, which can be used on graphics cards w/ <512MB. The only difference between these modes is that the 512MB mode uses uncompressed textures.

Also I read in my Doom 3 walkthrough, that you should only activate some features on graphics cards w/ at least 128MB and some other you should only activate w/ at least 256MB.


"The only games that don't use it are 2d prerendered games like freecell, but if all your playing is freecell why buy a videocard at all?"

you need a video card even for 2d apps (if you dont have a graphics chip onboard of your mobo), because otherwise you dont get a output for your monitor. you only dont need cards 3d cards, but you will hardly get cards w/o 3d.


"Thanks guys, I'm working on progressing to high end cards soon, wife permitting, (current on is really pants for some reason!) and didn't want to see my system RAM eaten away wholesale."

High end cards will not use TurboCache. At least not in the foreseeable future.

AGP aperature and texture cache settings through your video card are very rarely used. This is because the maximum bandwidth AGP has available to it is 2.1GB/s. Granted, AGP cards never even get to tap into half of that bandwidth. If they did, however, you would see more performance constraints than without it, because there is no allocation pipeline for either aperature or software texture caches through ram... which means that it would have to partially run on the same pipeline that carries system memory to and from the processor.

There are tests that show a larger aperature is better, but with 128MB and 256MB cards, this feature is ultimately useless and only still available as a safety net.

TurboCache, however, is designed to have its own partial memory controller, which alleviates the northbridge or AMD IMC(integrated memory controller) from taking on the task of allocating memory for the AGP or PCIe bus.


That "special" mode in Doom III isn't so hidden and isn't so special. I was able to run the game without texture or lightmap compression at 1280x1024x32@75hz on a Geforce FX 5900XT with 128MB ram. The game ran fluidly at 30-80fps, the lowest being with several monsters on screen and the highest being in partial darkness.

What you have to realize, is that Doom III was a flop as far as textures are concerned. They promised high resolution textures, yet failed to deliver a SINGLE texture above 512x512 resolution. This is the same maximum reached with both Halflife and Quake 3.

There is only a slight difference in clarity between having texture compression on and turning it off with Doom III. This is because the textures are so low resolution that they aren't even able to be compressed more than 15%. So, if you thought you were missing some level of detail by having texture compression on, you weren't.


High end cards might have this TurboCache feature as an additional performance measure, not a budget measure.

If you go to the Nvidia website, they bench a 6200 with 128MB of ram featuring TurboCache against an ATI Radeon X300 PE with 128MB of ram. (http://www.nvidia.com/page/turbocache.html)

It used to be a year ago that a 128MB video card was considered top of the line. Now that 256MB has become the pinnacle for enthusiast cards, people are waiting for 512MB.

But because of this TurboCache feature, we might not ever see it.

my response if this post was in 2008 =

LMFAO. ROFLMAO AND LOLLERCOPTEr. You my friend are a noob. This guy has no idea. Bless the noob. Crysis requires alot of memory in a card and even my 2 x 8800 ULTRA with 768 MB each do not run Crysis maxed out at full resolution. Please can you research into what you are saying before posting here.

Memory largely effects gaming fps and is directly linked to gaming performance all around alongside a diecent processor and a large quantity of ram.

For more information just google or wiki it.

TurboCache was developed as a means to provide a better cost/performance ratio by reducing the amount of memory modules on the video card. According to NVIDIA, a GeForce 6200 with TurboCache will perform about four times better than the Intel GMA 900. As with integrated graphics, the operating system may report a lower amount of main memory than is physically present when main memory is used. (I'm not personally sure of how this effects benchmarks but i'll have a look into it as i just got a crap laptop with a crap 7300 GO turbo cache and i'm interested to see.

All the best.