Oppinions: 3850 256MB or 8800GT - Page 2 - Overclock.net

maroon1 · 12-06-07

According to AnandTech, 8800GT performs around 15% faster than HD3870 on average

look at these benches.
http://www.anandtech.com/video/showdoc.aspx?i=3151&p=6

Vegnagun · 12-06-07

The 3850 is similar to the 2900gt.. Slightly faster
http://www.gpureview.com/show_cards....=546&card2=537

I'd either go with the 3850 or 2900gt.. Cost effectiveness is important

dragosmp · 12-06-07

Quote:

Originally Posted by maroon1

According to AnandTech, 8800GT performs around 15% faster than HD3870 on average

look at these benches.
http://www.anandtech.com/video/showdoc.aspx?i=3151&p=6

They are both too expensive for me, that's why I went for the 3850. The 8800GT 256MB is likely to cost around the 3870 level, so it's too much for me. 10x for the imput though.

Quote:

Originally Posted by Vegnagun666

The 3850 is similar to the 2900gt.. Slightly faster
http://www.gpureview.com/show_cards....=546&card2=537

I'd either go with the 3850 or 2900gt.. Cost effectiveness is important

Cool

didn't knew the site. I was hovering between the 2900GT (131€) and the 3850 (151€) but I went for the latter, it seems to have a slight edge. With the low power consumption it's a lot better than the 2900GT in my book.

Anth0789 · 12-06-07

Id say go for the 8800GT

Nvidia wont let you down

Lt.JD · 12-06-07

If price is a concern the 3850 is good but the 8800GT pwns all

Lelin · 12-06-07

Have you considered the 8800gt 256MB? It wont lack memory for low resolutions and beats the HD3850.

WhiteCrane · 12-07-07

You all seem to put a lot of emphasis on the card's RAM for its ability to handle high resolutions.

I'd think thjat Hi Res has a lot more to do with its GPU, and your CPU and system RAM, than your V-RAM. Anyone care to get technical to explain that one? I'm very curious. I think of V-RAM as texture memory.

dragosmp · 12-07-07

Quote:

Originally Posted by WhiteCrane

You all seem to put a lot of emphasis on the card's RAM for its ability to handle high resolutions.

I'd think thjat Hi Res has a lot more to do with its GPU, and your CPU and system RAM, than your V-RAM. Anyone care to get technical to explain that one? I'm very curious. I think of V-RAM as texture memory.

I can't be too techy, but a thumb explanation could sound like this.

Let's say we play a game like Crysis, it's advanced enough to make my point. A certain image/frame has a certain amount of stuff on it: leaves, sky, trees, why not buildings and enemies. In this particular game the quantity of information contained in every frame is enormous when compared to other games. All this info is well compressed and stored in the texture buffers, Vram. The size of the frame that has to be stored is proportional to the resolution, assuming the same detail level. Any vid card renders about 5 frames ahead, so it needs the place to store them.

The VID cards aren't quite like CPUs, they don't need an equal amount of Vram for the instructions, because there aren't really that many things that can be applied to a certain frame. Among those we can find AA and AF. The AF roughly speaking makes the edges in old games look smother, in new games it has next to no real impact. AA on the other hand makes textures more sharp by applying filters. Each run by a certain filter (AA) works on the frame, so it has to store the original frame, the intermediate results and then the final frame before rendering. This makes the VID ram usage to explode...

Now VISTA has a preety feature called VID ram virtualisation, a DX10 demand btw. It makes the VID ram and a part of the system ram (equal to the VID ram) virtual VID ram. The VID card's shaders have direct access to the whole thing (say 256VID ram +256sys ram = 512MB). The VID drivers though tend to force the VID card to use firstly the VID ram, logic since that's extremely low latency and 2GHz for some cards. This is a feature used only by very low end VID cards, where the added memory and PCIe 2.0 interface works wonders, as a 1.8GHz DDR3 ram is as fast if not faster than the onboard ram, 128bit (vs 64bit onboard). The only issue is the added PCIe 2.0 latency in the loop, but hey. This maybe the only real way in which system ram can influence significatively card's performance.

The CPU plays a great role, it runs the AI, physics, the load/store instructions, so it's very important. High speed system memory helps the CPUs with high mem latency (C2D for ex), so by speeding the CPU calculus the RAM becomes somewhat more important. A large amount of memory also helps when playing huge maps with lots of AI (C&C3 for ex) that need to be stored somewhere. The sys ram is also an antechamber for the textures that are loaded from the HDD and shipped by the CPU when they are needed.

The GPU is the key for high res, just as the Vmem. Those frames, that are stored in the Vmem have to be computed, and here's where the GPU intervenes. There has to be a balance between what the GPU can process and the amount of place it has available to store whatever it computes. It's no use to put to a 8800Ultra 256MB of ram, it would be far worse than a 8800GT with 512MB. The idea is to know how much the GPU can take, how much can it render. Say it can compute X MB/s of data, you can compute what is the max memory it's reasonable for the GPU to use, and put that amount on the Vram. Things don't really happen this way because there are standards like 256MB, 512MB, 1GB.

Now, in this thread the following GPUs: 3850, 3870, 8800GT are all capable of making use of 512MB. ATI has chosen to put 256MB on the 3850 for $ reasons, and I'm happy they did. It's true that if I up to a certain resolution (the reviews talk about 1680x1050) the card runs out of Vram, even sooner with AA/AF. It is forced to access system memory, which has the following added latency: few ms to search within the Vram, the tenths of ms of latency of the PCIe 1.0 slot, the PCIe to NB latency, NB to CPU, CPU to RAM and back RAM-CPU-NB-PCIe-GPU. This is ONE memory read, that otherwise can be made using the GPU-Vram-GPU loop.

Hope that clears up something. Does it have FAQ material? Cheers!

WhiteCrane · 12-07-07

Okay I see. So basicly, the less VRAM u have the more the GPU would need top work by clearing out the old VRAM and filling it up AGAIN more frequently and performing the same tasks in less VRAM, the same way less system RAM can bottleneck a CPU.

Makes perfect sense... I can;t believce I forgot the fact that higher resolution = higher image size... I make VERY careless mistakes easily lol.

Thanks!

My monitor is capped at 1050p. Its a widescreen. (I forgot the horizontal resolution)

Do I need a 512 card? to me, EVERY frame counts. We're talking about the PC in my sig.

pbpenguins412 · 12-07-07

Awesome explanation. REP+, that helped a lot.

Quote:

Originally Posted by dragosmp

I can't be too techy, but a thumb explanation could sound like this.

Let's say we play a game like Crysis, it's advanced enough to make my point. A certain image/frame has a certain amount of stuff on it: leaves, sky, trees, why not buildings and enemies. In this particular game the quantity of information contained in every frame is enormous when compared to other games. All this info is well compressed and stored in the texture buffers, Vram. The size of the frame that has to be stored is proportional to the resolution, assuming the same detail level. Any vid card renders about 5 frames ahead, so it needs the place to store them.

The VID cards aren't quite like CPUs, they don't need an equal amount of Vram for the instructions, because there aren't really that many things that can be applied to a certain frame. Among those we can find AA and AF. The AF roughly speaking makes the edges in old games look smother, in new games it has next to no real impact. AA on the other hand makes textures more sharp by applying filters. Each run by a certain filter (AA) works on the frame, so it has to store the original frame, the intermediate results and then the final frame before rendering. This makes the VID ram usage to explode...

Now VISTA has a preety feature called VID ram virtualisation, a DX10 demand btw. It makes the VID ram and a part of the system ram (equal to the VID ram) virtual VID ram. The VID card's shaders have direct access to the whole thing (say 256VID ram +256sys ram = 512MB). The VID drivers though tend to force the VID card to use firstly the VID ram, logic since that's extremely low latency and 2GHz for some cards. This is a feature used only by very low end VID cards, where the added memory and PCIe 2.0 interface works wonders, as a 1.8GHz DDR3 ram is as fast if not faster than the onboard ram, 128bit (vs 64bit onboard). The only issue is the added PCIe 2.0 latency in the loop, but hey. This maybe the only real way in which system ram can influence significatively card's performance.

The CPU plays a great role, it runs the AI, physics, the load/store instructions, so it's very important. High speed system memory helps the CPUs with high mem latency (C2D for ex), so by speeding the CPU calculus the RAM becomes somewhat more important. A large amount of memory also helps when playing huge maps with lots of AI (C&C3 for ex) that need to be stored somewhere. The sys ram is also an antechamber for the textures that are loaded from the HDD and shipped by the CPU when they are needed.

The GPU is the key for high res, just as the Vmem. Those frames, that are stored in the Vmem have to be computed, and here's where the GPU intervenes. There has to be a balance between what the GPU can process and the amount of place it has available to store whatever it computes. It's no use to put to a 8800Ultra 256MB of ram, it would be far worse than a 8800GT with 512MB. The idea is to know how much the GPU can take, how much can it render. Say it can compute X MB/s of data, you can compute what is the max memory it's reasonable for the GPU to use, and put that amount on the Vram. Things don't really happen this way because there are standards like 256MB, 512MB, 1GB.

Now, in this thread the following GPUs: 3850, 3870, 8800GT are all capable of making use of 512MB. ATI has chosen to put 256MB on the 3850 for $ reasons, and I'm happy they did. It's true that if I up to a certain resolution (the reviews talk about 1680x1050) the card runs out of Vram, even sooner with AA/AF. It is forced to access system memory, which has the following added latency: few ms to search within the Vram, the tenths of ms of latency of the PCIe 1.0 slot, the PCIe to NB latency, NB to CPU, CPU to RAM and back RAM-CPU-NB-PCIe-GPU. This is ONE memory read, that otherwise can be made using the GPU-Vram-GPU loop.

Hope that clears up something. Does it have FAQ material? Cheers!

View Poll Results: What do you think?
The 3850 is good enough for 12x10	19	37.25%
The 8800GT pawns all	29	56.86%
The prices will fall shortly (50$ or more)	6	11.76%
8600GT/S were too expensive to begin with	7	13.73%
Multiple Choice Poll. Voters: 51. You may not vote on this poll

		Overclock.net - Overclocking.net > Graphics Cards > Graphics Cards - General
Oppinions: 3850 256MB or 8800GT

12-06-07	#11 (permalink)
maroon1 New to Overclock.net Join Date: May 2007 Posts: 299 Rep: 36 Unique Rep: 25 Trader Rating: 0	According to AnandTech, 8800GT performs around 15% faster than HD3870 on average look at these benches. http://www.anandtech.com/video/showdoc.aspx?i=3151&p=6

Currently Active Users Viewing This Thread: 1 (0 members and 1 guests)