Only stupid people frown - Dave Evans
|
|
|
Preview of 2010 discrete PC GPU futures
WHAT A LOVELY YEAR for GPU watchers - the tables have turned, ATI took over the lead from Nvidia, and Intel is delaying the graphics side of Larrabee for, let's say, quite a while, as it'll test the waters first with those GPUs in the CPU package coming along with the Core i5 soon after you read this.
The 40nm process migration is pretty much completed for ATI, while Nvidia at the very least managed to have an entry level product, GT240, running nicely from those same 40nm TSMC fabs. Global Foundries will probably also see its first GPU 'customer' this year, and I think both ATI and Nvidia will probably use it for selected designs.
At the end of the year right now, and let's put aside the dual-GPU card variants for the moment, we see the following rough performance leadership ranking: the HD5870 leads, followed by the GTX285, HD5770 and GTX275, in that order. You could also insert HD4890 together with the HD5770, but I think it makes no sense to consider the older, hotter card anymore unless you're getting it on sharply discounted clearance sale terms. The mid-range and low-end chaos, with a myriad of often repeatedly renamed entries from both vendors, has no clear leader right now, except for the AMD ATI chip process advantage from its early 40nm adoption.
So, the low- to mid-range ATI HD4770 in the 40nm process gave something of a headache to what was the GeForce 8800 then 9800 then GTS250. The HD4600 versus GT9600 family contest was more equal, though.
DX11 is the buzzword. It was talked about well before the Windows 7 launch, but how many DX11 games are out there by now? Heck, just look even at the number of true DX10 titles by now. In any case, raw performance and ability to handle Full HD and beyond resolution in real time 3D with all the effects thrown in might still become the key strength.
The SLI and Crossfire parallel GPU support has kind of matured now, and even quad-GPU solutions can offer noticeable scaling speedups now, as every additional GPU is used more effectively. The new mix of graphics, physics and computational loads being thrown at those parallel GPUs should improve the useability even further, until we start hitting the PCI-Express bandwidth contention wall. Will it be time for HTX or QPI graphics at that point? The extra latency from the cards' additional PCIe bridge overhead suggests that might become the case.
What to expect in 2010 then? Well, there will be more DX11 games out there for sure, helping justify the higher end cards. I also expect the first 8-10 Megapixel class LCD displays to surface. Resolutions like 3840x2400 or even 4096x2400, bringing in 4K cinema quality pictures, will push the requirement for as fast as possible GPUs to run them. For those on a budget, seamless multi-display setups like the ATI EyeFinity will do the same. On top of that, high refresh rate 100 or 120Hz displays will become normal at the high end, putting further load onto the frame buffer refresh portion of GPU workloads.
In the first half of the year, I expect ATI to mature the HD5870 design and, as the 40nm process is fine-tuned, possibly have a major speed-up stepping in the same format, something like a HD5890. It should enable about 1GHz GPU speed and above GDDR5-5000 memory performance. On a dual-GPU card, though, such speed would at least require a two by 8-pin PCIe power connector set, as well as a very different cooling solution.
By early March, Nvidia's "Fermi" GT300 should also arrive, finally. It will have a wider memory path than the ATI offering, at 384-bit GDDR5 versus the 256-bit per GPU path width in the HD5870, and is expected to have higher double-precision floating-point performance, not to mention easier general-purpose C or Fortran programming promised by Nvidia.
How high the performance of GT300 will really be should be known closer to the launch, hopefully before the CeBIT show in March. While Nvidia did focus a lot on computational GPU usage, gaming DX11 and professional OpenGL 3D graphics performance will still be the main yardsticks to assess the new graphics processors.
Either way, Nvidia is expected to lay out a more firm roadmap for the GT300 family by March, including the dual GPU versions - hopefully on a single PCB this time right from the start - and wider 512-bit memory bus versions for large memory setups.
While DVI slowly gets replaced by the HDCP-malware compliant HDMI, I don't expect the big chubby DVI connector to disappear at least 'til year-end 2010, by which time DisplayPort should be fully matured. Keep in mind that those new ultra-high resolution displays will need multi-channel links to feed theme at full resolution with high frame rates.
Integrated versus discrete? Up to now, the AMD and Nvidia integrated graphics chipsets ran rings around the Intel ones when it came to 3D performance. That is changing with Intel's new "GPU-inside-CPU-pack" approach. Not only will this GPU be faster by itself, but it will also talk directly to the CPU and memory resources and, yes, be overclockable together with them. This will pose a threat to the low end of discrete graphics as well.
As for the GPU chips running in parallel, I don't expect anything beyond quad-GPU support for a while, but rather expect a focus on making parallel GPU operations even more efficent on both graphics and non-graphics tasks thrown at them. If you see more than four GPUs in a system, it'll either mean a very special multi-display setup, since four GPUs now can support up to 24 displays, or a computational GPU 'personal supercomputer' contraption.
In summary, 2010 will be yet another no-Larrabee year, with Nvidia and ATI still sharing the spoils, and Taiwan and Hong Kong GPU board vendors finding it as hard as before to differentiate their products as the two big bosses force them as usual to push the reference designs first. As to whether Nvidia or ATI will be the leader at the end of next year, that will depend upon how good the GT300 proves to be when it finally rolls out. Watch this space. µ
Share this:
Share
We need articles on the next generation to come by now on sites like the inq surely? Not some story on renaming of the old cards, we've been there and seen it, the 5xxx range already is made available in various forms and names and OC'ed and alternative version, that's old news.
You rank the cards from the very fastest single CPU solution, and accidentally ranked the 5770 as faster than the GTX 275. The 5850 would be the card that is faster than the GTX 275 :)
Otherwise, good article.
I reckon either the ht socketed GPU or go with the external GPU solution - and the forerunner to this has to be AMD/ATI's XGP setup - basically external PCI-e.
Sure it'll need some development - but it could give us our tiny, low power desktop systems with the ability to push big graphics loads when required.
My assumption is that AMD ditches the sluggish PCIe for their own speedy HyperTransport on their top-end computation-oriented R900 chips.
Three HyperTransport channels per chip means an easy 8-socket solution according to AMD. I can easily think of one Opteron and seven Radeons for roughly 100TFLOPS per rack unit... by summer.
This goes hand-in-hand with AMD's sharing of documentation to the Linux folks. Linux happens to cover 90% of the top500 supercomputer list, and I think AMD wants to gain some.
WTF, this article is barely comprehensible. What utter garbage is the inquirer pushing out these days?
It is possible to place a gpu in a cpu socket but it would alienate all the 3rd parties who package gpus onto pci-express cards. I don't think you could do a top end gpu in a cpu slot though. You'd need to keep room on the chip for a decent sized cache to reduce how much system memory the gpu would use and of course, I doubt you could really go as high as 256 megs cache without seperating the memory and the gpu silocone which defeats the purpose of using a cpu socket but for netbooks that don't need huge amounts of memory, I don't see why they couldn't put 16-32 megs of on die cache.
It make sense:
if you remove the 16x /8x pci-express slot
you might get enough space to put a gpu socket. however...
However, you would need the extra memory, usually sitting on the card, along with all the extra components. Ati/ nvidia would almost have to make their own Mobo..
Hopefully for gaming, PCI-express bandwith is still plenty...yet.
Can the gpu have it's own socket?
Maybe have 300/600/900 pin socket, a power slot etc. This direct connection to the cpu should improve performance.
Does this sound sensible?
Says the man in second place, lol!
nVidia must think its Christmas.
Lets face AMDATI screwed the pooch once again didnt they? They promised supplies by the 15th, well its the 20th and still no 5870s.
Here they have complete market dominance for three months straight and no frikkin cards to sell the punter. So anyone gagging for a new card can't find an ATI5870 for love nor money.
I tell a lie I did see some ten 5870s going for over £450 GB each and they still sold out. Meanwhile nVidia are getting their cards ready and the market will not be even slightly saturated by the time they get to launch.
AMDATI have lost mega bucks because TSMC have fouled up in the most horrible way imaginable.
Total waste of space. Dump those losers AMD.
Well - if you think about power consumption, but still want to play games with all the eye candy - then I can imagine you would love you have a 5870 Crossfire setup you use for playing - installed on a 780G chipset, where you use the onboard graphics card when you are online surfing forums or watching youtube or watching movies in general.
You only need the power of the Crossfire setup when you are playing games - so why not give us the option to disable them with a single click and switch to the integrated graphics card instead?
I know the picture is not clear, but you can draw some similarities. Go find reviews of the 780g chipset and its power consumption when watching HD movies - (73W according to Toms hardware) and then find any review of any modern graphics card and tell me you can find an idle state this low.(and most likely the card will not be in idle state when watching movies)
I know, I know - pears and apples - but it shows what it is that I want. A system that only uses power when that power is needed
QuadHD res displays would be awesome but they are not likely to come due to the very restricted market size. There are 30" 2560x1920 displays available but they are selling in low numbers which makes that their price is high and there is a vicious circle in number and price but the market is small in the end.
Sharply visible at any useful distance for 18" screens. Not sure gamers will benefit, but it's long overdue to have 8MP screens for business applications, the 1080p or 1920x1200 are simply unacceptable. HD theater also needs higher resolution screens just to get rid of painful pixelation when you view on large screen (technically the data stream is and probably will have way smaller bandwidth, and at 8MP is already quantum noise limited specifically for dark scenes so we will not gain in quality here). Look at the pixelation of 55-65" LCD and plasma TVs - yuck. Look at 28" LCD monitors for business applications - yuuuuucckk. High size, high brightness just made the pixelation effect much much worse.
And i am even not saying how much more applications you can place on larger playing field of 8MP. For now the only we can do is to use fricking virtual resolutions, like mine 3600x2250 and several monitors...At the end of the 12-14 hour day near the screens you feel that after several years you will end up with professional neck injuries... (LOL, but i am not kidding)
The author said that this might be an issue, at some time, but I doubt it. Eventually it will, but I think PCIe 3.0 x16 is going to solve our bandwith needs for a long time yet. Only daul-GPU cards max out a 2.0 x16, so double that bandwith to 1.0 GB/s and remove the 8b/10b coding scheme and change it into 128b/130b, and we'll be good for a while. My issue is with how all that bandwith will get used on the CPU, because HT 4.1 and QPI aren't reaching x16/x16/x16/x16 2.0, so double that and we're screwed. I say cross the LGA 1156 and 1336 to get QPI with controllers onboard. Yay grafiks!
Scaling from 1920 - 4096 would be awfull. What i mean by that is if you try and play your 1920p movies on a resolution of 4096, the pixel width would be 2.13. Pixels would be blended and things would not be as sharp. What you want is 7680x4320, which is 4X the size of 1920p. this will give you 4 pixels too the 1, making the old 1920p movies sharp still.
Why do i call it 1920p instead of 1080p? Simple open up a movie and see if you see its resolution as yyyy X 1080....not normally as it is always something more like this 1920 X 867. That annoys me they use 1080p instead of 1920.
5870 to GTX285?? I thought that ATI had a single GPU card called the 5850 that in most situations whipes the floor with the GTX285.
James, what do you think generates more heat? A Core i7, or an Atom? The Core i7, obviously.
The clock speed isn't the only factor in temperature and heat output.
IGP's don't downclock and start at about 500MHz and go as high as 900MHz with the upcoming 32nm with 45nm gfx (Clarkdale).
So, ATi 160MHz dynamic core clock vs, IGP 500MHz-900MHz fixed core clock...
hmmm...
switching from integrated to discrete to save power...
Um... try looking first at the ATi 4000 series. watching a HD video i got from Gamespot, plus browser, and writing this comment, my ATi 4000 GPU clocks are 160MHz core and 250MHz memory.
Now... wtf would you WANT integrated that limits your mobo choices, and more chipset heat output and cost, vs a single discrete GPU that downclocks itself (55nm with the 4000 series as an example).
Basically it's obvious why your idea is wack, and unnecessary.
PS, the 90C GPU's are already on the decline, so another non issue. Almost every GPU can handle up to 130C and backed by lifetime warranty. So whining about how long the card will last is again unnecessary. 90C sucks, but it was the tradeoff for performance at the time (4800 series and GTX 200 series). I've yet to read an article where a mass amount of cards were dying because of heat.
At the current rate of progress (near nill), my next computer will likely be my last computer.
Where are the 10GHz CPUs & GPUs, and 3GHz memory, that should have reached the market by now? Intel, AMD/ATI, and Nvidia are a bunch of lazy bums.
For the cost, my PS3 & Xbox360 deliver more bang for the buck than the equivelent PC (same was true for my PS2 and N64, but not so much for my Dteamcast). Installing Yellow Dog Linux on the PS3 was simple, and it orthogonally extends the machine's usefulness. I wish all the software I want to use could be purchased for either the Xbox360 or PS3; going forward, I'd never buy more than a netbook.
I can't wait for the PS2^2 and The Next Xbox.
These articles about annual 5% performance improvements, which have been the norm since the late 1980s, are annoyingly supportive for the nearly status quo.
Still waiting for a gpu chip with 512Meg memory cache on chip that plugs into a 2nd amd cpu slot utilizing hypertransport to talk directly with the cpu. It doesn't even need 800 stream processors, 200 stream processors would be just fine for what I want.
which techology will you use that is fast enough to deliver GPU power from the external box to the computer - without increased latency?
that both NVidia and ATI will move their cards from the PC case into a graphics sub-system box.
Imagine an external box, with its own power supply and cooling, connected to the a PCIe I/F card with a cable. The box can be generic to each vendor, so you can upgrade the graphics board in it as the GPU improves.
This will substantially improve the thermal and power management of the PC, and will allow the graphics vendor to have more freedom in designing their products - unconstrained by PC mechanical, thermal and power limitations.
Also, one can use a lowly graphics card in the PC for non-gaming apps, and hook up the subsystem for game time - a greener solution.
Occasional and full-time Gamers will certainly spend the $100-$150 added consumer cost.
I'm assuming he means just to have the option to switch from one to the other.
Doesn't even have to be automatically, but give us the option to go from a discrete card to an integrated one, by clicking a button. No matter which version of discrete card and integrated card we have.
We want to save power when we are just online reading stuff, and have the power to watch HD stuff. Have the choice to go back and forth
First Glimpse into Nvidias' GTX 380 Shows score Higher than Ati 5870 by 10%. Same test showed gtx285 being higher than 5870,too. So Suspect Parsenipers. So Smudging Might O' Been More Test, yet do have Pics & GTX 360. Column of Salt to Look Back Upon. Anyway, Dual Ati 5970 Shoely Still Bee Crown of Glory. At Present Till ?Known Fuzier Gets Sampled.
Heres headline, as rest is just forum entry(W/Good Graphics): GeForce GTX 360 and 380 Benchmarks
GeForce GTX 360 and 380 Benchmarks By Hilbert Hagedoorn, December 14, 2009 - 2:55 PM
Do you mean power express but based on load? It would be late to switch back and forth after the app starts running. Yet "Load" Express would sound cool.
At the end, based on the load current discrete ATI chips have power saving functionalities already, switching to a integrated might be nice, yet they would need extreme proxy control for all types of app calls before any real draw submit; would be too much to have.
Gamers. The rest of the crowd will be happy with good Windows Aero compatiblity.
Much more interesting for the crowd are good full HD / blue ray support.
3D graphics currently demands way too much power. Which is fine if you use the PC as a gaming for 2-3h a day.
For a "regular" one-size-fits-all PC a good graphics card simply is a "Watt monster" that just sits there and continuously munches away energy without giving any real benefit.
Maybe Intel Larrabee or the "could chip" and/or other concepts will address that issue. Hopefully.
Not sure if nVidia have this sorted but I know ATI haven't - can someone please allow us to use integrated graphics for the OS and then power up the discrete card for games. No jumpers, no switches, no restarts, just a driver that can sense the load and switch accordingly. Cross vendor, cross ATI/nVidia. And all on the desktop too, not just laptops.
Thank you.
I wonder whether we will see LCD TVs with 3840x2400 or even 4096x2400 resolution on the market soon.
Would be nice to test if a regular blue ray video would even look better on a 50 or 60 inch 4k TV.
More dpi should always be better, but it would be interesting to do a test.
3D is not of interest for regular TV as long as you need to on special glasses. It's nice and fun for an special event with friends, but too cumbersome for everyday us.
The biggest use for 3D TV would be football and other sports.
If one could watch live sports in 3D those screens would sell like crazy. Not to be expected soon, I think.
I agree, I am sick of the heat generated by graphics cards. It is especially a problem here in Miami Florida. I actually had to put my computer in the next room and made a small hole in the wall for the monitor/keyboard/mouse/cat5/usb.
I agree and when I buy a high end card I do not look for one of the overclocked ones, stock does just fine. I am surprised to see my GTX 285 is still up there. It has been around for a while. With all the games being written for the Kiddie Consoles first then ported to the PC I see my gaming rig lasting a long time without needing to upgrade anything.
I am tired from GFX cards running around 90c under load... all the cooling solutions that Nvidia and AMD propose for their high end cards are cheap... makes me really wonder if i should always get the lower clocked card (e.g.: 5850)so that u get less heat and more product life.
Also I have learned from experience that the best 3d cards are the ones that come out after a new consoles are launched... e.g: GF6800 or X800 at the xbox/ps2 time.. we did not have to upgrade untill the GF8800GTX (Which i am still using)
Being ahead of consoles with one next gen graphics really pays off as you always worry not about running most of the latest games..