ATI/NV Policies, Watts, and working DirectCompute with ATI 4xxx

[FastZ28]

Merging (lower electric bill) game PC and desktop PC into one PC. I was surprised it was possible, and the results.

Desktop was uATX low profile S775 Q6600 G0, onboard intel G41, idle ~82W with 1 HDD/1 Optical and CPU ~2900Mhz. Not fastest PC, but fit the bill.

Game PC was Asus M3A78-T AMD 720BE CPU unlocked to Quad and max OC'd for best FPS (NB/L3 at 2800Mhz, etc) and ref XFX 4870's 1GB (I game at 1920x1080 on Plasma with AA) just over $100 each, load watts go to ~500W. PC stayed off mostly. BTW, previous game PC was Q6600 G0 at 3.8Ghz on Asus P5Q Deluxe Pro (great board, 4 sticks DDR2 was 1120mhz) but AMD's 45nm CPU (on all 4 cores) beat my trusty old 65nm G0 Intel good.

Reference 4870 idles at high watts & usually sport a silly fan profile. I BIOS edited/flashed both cards to 160/200 (2D, 1.083v) and 855/4500 (3D). Fan profile was fixed also. This only matters if you try to get my sub-100 watt idle using a 4870.

I wanted low watt PC with DirectCompute to fully utilize GPU & CPU. ATI 9.11 Beta with OpenCL has no DirectCompute for 4xxx (5xxx yes), it's sorta obvious ATI wants to sell 5xxx's now. But NVidia has DirectCompute (& more) working on 8,9,200 GPUs for awhile on DX11 and XP/Win 7 allows mixed cards.

I had spare Asus NV 9600GSO's 384MB DDR3. GPU is 96 shaders, 65nm process, they clocked well even with 1.1ns ram without Vmod. Now that ATI + Physx is sorta popular again I pulled one 4870 to install the 9600GSO without issue (191.07 driver, patched with GenL patch, latest Physx software).

Using 191.07 ("GenL" patched) lets my existing DX10.1 4xxx ATI cards render 3D as primary, but also adds Physx (big deal in some games), CUDA, and DirectCompute (big deal to me with power saving & performance in mind). DirectCompute is native to Win 7 as it ships with DX11).

Wattmeter shows Quad Phenom II (3.6Ghz down to 800Mhz/1v at idle) and single 4870 (160/200 1.083v) idles at 95W. PC is nearly silent, even with moderate loads.

Adding the old 9600GSO 192-bit 384MB (550/1350/1600 stock) OC'd to 800/1890/2000, PC idle at 115W. When GSO is running the DirectCompute Bench V0.25 watts hit 125W. When this benchmark gets to CPU test, watts are ~225W). Running "memtestG80" (NV stability test) I get ~200W with CPU basically idling (super stressing the G92 9600GSO, NVidia people usually stress test their clocks with this).

"DirectCompute Bench v0.25" uses same code path for CPU and GPU in the multi-thread test and shows my 9600GSO around 10x faster than my OC'd Phenom II Quad (at over 100W less, at least in this bench). But I don't need this bench to tell my Win7 PC is much faster/responsive from file browse to internet browsing, nothing slows it down (70 firefox tabs). Games work awesome.

The Unigine DX11 bench is quite fast and says "9600GSO as renderer" even tho obviously 4870 is primary 3D (nothing attached to Nvidia card). Force fake monitor is another Win7 feature. All Physx/CUDA stuff works fine.

DirectCompute (DX11) is awesome and I won't build a PC without it now. CPU + GPCPU works great, even on old DX10 NV card & DX10.1 ATI, with "GenL" driver patch. I may not even require my CF 4870 setup anymore (still testing). Even at full 3D load, PC watts are lower since 9600GSO is taking load, watts (& heat) away from the Quad Phenom II, thanks go to DirectCompute.

Amazing a Quad Phenom II and 4870 (mod BIOS) with two lousy PSU can idle at 95W.

Adding just 20W more idle watts (9600GSO) gives ATI DX10.1 users everything; can run all apps, all API's, & becomes more power efficient, relying less on the power hungry OC'd CPU, and becomes faster just in using it (DirectCompute from DX11, working on Nvidia DX10 hardware, if installed).

BTW, yes (800/1890/2000) is high for 96 shader GSO, but for Physx & Direct Compute you can go high with stability, I doubt this speed works in 3D render. My DirectCompute score is "M17248" (it may go higher, I haven't tried) beats the 9800GT score in the "readme.txt" for DirectCompute Bench v0.25.

ATI might sell more cards if they "activated" DirectCompute on 4xxx cards now, not 1-6 months from now (if they never do, NV has a super win, once people fully realize what they are missing even on older cards). They have it working on brand new chips (5xxx), so they certainly know how to make 4xxx work. Performance?? 5xxx series does bad in DirectCompute bench (early drivers, maybe, but they get twice the shaders).

Nvidia bought Ageia right?, not ATI. NVidia would certainly sell more cards if they stopped breaking the mix with ATI cards (195.39 again), as some people want NV cards of all levels if not for dedicated PPU, DirectCompute for Win7 is disabled for ATI 4xxx (unless 9.11 official, delayed, brings it?). Frustrating game developers; too many API & policy yields more bad console ports.

Win 7 is here, Nvidia was ready on years old DX10 stuff, ATI is messing with DX10.1 customers (existing or not) that might crossfire or switch to NV. Perhaps they feel its worth more 5xxx sales?

Any motherboard usually (even 2 PCIe) can mix NV and ATI together, most PCIe boards have at least 1x slots too (1x to 16x risers are only <$4), and its all you need (and DX11) for Physx, CUDA, and/or DirectCompute to give power hungry CPU a break (& electric bill) with performance boost as gravy.

[IPlayNaked]

wat.

[nascasho]

Originally Posted by IPlayNaked

wat.

+1 I'm tring to figure this out because it goes from ATi to Nvidia to PPU to drivers...

[FastZ28]

Originally Posted by IPlayNaked

wat.

lol

http://forums.guru3d.com/showthread.php?p=3338707

This thread on this forum shows some bench results. Download the latest v0.25 DirectCompute, you have 5xxx series, you should get a high number (this new version gives diff results).

I'm just pointing out that people with older ATI primary 3D hardware can add Nvidia 8/9/200 series to get DirectCompute (and other stuff) in Win 7, makes a nice performance boost even at a lower wattage at the wall.

Most 5xxx users add the NV card to get Physx/CUDA...

4xxx users get that plus DirectCompute, a "co-op" between the CPU and GPU handled by Win 7 natively.

[FastZ28]

If it would help, I can post screengrabs of popular performance/benchmark programs.

Plain and simple, Win 7 allows ATI and NV to co-exist; I and many others have had NO troubles. XP does too. But Win 7 has DX11 (Vista SP2 can get DX11, but don't let ATI/NV play together).

ATI DX10.1 hardware is getting the shaft on DirectCompute, a performance boosting and wattage saving feature of DX11 allowing some work to be done on GPU instead of CPU at the operating system level. At least until ATI provides drivers, if they do.

ATI DX11 hardware lacks Nvidia Physx, obviously, and adding at least a 64 shader Nvidia card (96-128 is better) can provide a nice boost in the latest games that even serious choke on the ATI 5870 and NV 285 with Nvidia "Physx" (even with original 2006 Ageia PPU card)

ATI Physx is different, Havoc right?, different games I believe...

DirectCompute was my interest, Nvidia supports it on old 8-series cards, ATI doesn't on 4890's even.... just 5xxx...

[IPlayNaked]

Oh...I guess I understand...kind of. What's with the wattage? It doesn't save you any power to use physics and I don't know of any mainstream useful program that uses direct compute..?

[FastZ28]

Originally Posted by IPlayNaked

Oh...I guess I understand...kind of. What's with the wattage? It doesn't save you any power to use physics and I don't know of any mainstream useful program that uses direct compute..?

Depends on the physics, some actually save me power/heat now since the NV GPU is using almost no power unlike my OC'd CPU.

I just wanted a low heat/wattage PC (when surfing/downloading/etc) that was also powerful enough to take on the latest games at 1920x1080xAA. So I got the Phenom 2 Quad CPU OC'd and one OC'd 4870 past 4890 status and the whole thing idles at 95w with 4GB 1066, burner, etc...

But my mobo has 3 PCIe slots (8x8x4). So I added the Nvidia 9600GSO 192-bit to the 4x PCIe slot with a fake Win 7 monitor. Now idle watt is 115w, still good because now Win 7 is using DirectCompute (and also NV Physx and CUDA, plus the ATI stuff). It's so nice I might not require my 2nd 4870 in crossfire, still benching/gaming.

Nvidia's Physx just comes along for the ride as I see it. If it helps the game you play, then great. If it dies, it'll get replaced by OpenCL or whatever. Although I do know Nvidia extended Nintendo's, EA's, Sony's PS3 license to use the "tech" so I dunno when it will die.

Having ATI's Stream, NV's CUDA/Physx, etc, in the same PC obviously means you don't have much to worry about software wise. Hopefully it all gets sorted (in a year?). But ATI DirectX 10.1 people (4000 series) are wondering when they get the DirectCompute driver (ATI gives it to 5000 series so far) so the CPU can rest more, saving power & being more efficient. Extensive game physics/AI can zap the the CPU (even i7, working it hard, raising temps, watts, etc). There are some charts in another thread.

But the watts comes in here: Some things lend themselves better to the GPU, not CPU, and not only scientific calcs or video transcoding. I've noticed, with my wattmeter (which also measure joules per hour), my PC now uses less power to do typical tasks I've been doing in Win 7, & the kicker is it does them faster. Basic stuff even, switching programs, file/website browsing, etc... Win 7 with full eye-candy uses DX11 of which DirectCompute is one major component.

When my CPU would normally come out of "idle" or power save "1v" mode it jumps depending on the CPU load. Whatever the GPU can do, CPU doesn't have to do, thus staying idle way more often (saving watts). Win 7 (or DX11, I can't tell) must be "smart", because I wouldn't notice the dramatic user difference I do having DirectCompute now if it wasn't. Games, demos, benches, normal program load times, etc; it's definately snappier.

Not just subjective, if I pull the Nvidia card, my "joules per hour" rises quite a bit showing how well Win 7 (or DX 11) balances the CPU/GPU with various tasks IF you have Nvidia or ATI 5000 series.

SLI and Crossfire seem to be "render" tech, and don't seem to fully assist the CPU like a dedicated PPU does. Could be why even high end SLI people add Physx card and get better frames on even i7 (there's graph somewhere).

But my NVidia card was in my closet, & I got tired of waiting for ATI to release the DirectCompute for DX10.1 cards (4000 series). I'm so happy I'll probably use the other 4870 in another system, so they "lost" a sale?

Consider "Windows 7" the first useful program (that allows ATI and NV), that sports DirectX 11, of which DirectCompute is one of several key features to make computing/gaming better. i.e. Dirt2 will use DX11 which will use DirectCompute and Shader 5.0 so your 5000 card is fully used. 4000 card will use slightly watered down DirectCompute and Shader 4.1, but if my NV card does the calcs, my "4890" is actually free to render faster.

This is also why NV people seem to be adopting dedicated PPU with the release of DX11, better FPS...