Larrabee matches a GTX285?

[thilanliyan]

http://www.tomshardware.com/ne...idia-geforce,7944.html

Not bad for a first go...unless G300/RV870 are like double the performance.

 

[Idontcare]

Kind of impressive if that is where they are at today, some 6 months or more from actual launch date with more time to improve drivers and clockspeeds, etc.

 

[BFG10K]

I doubt this is anywhere near true. Even if the hardware is that fast (which I doubt), the GTX285 is also fast because of good drivers. Based on Intel?s track record with the GMA, I don?t see good drivers coming into the picture for a long time.

 

[IlllI]

i'll believe it when i see it.
i still think this could have more success in the next xbox or something, as opposed to desktop system. but we'll see

 

[ArchAngel777]

I agree with BFG. GMA still has pathetic drivers despite Intel telling us for the last 2 generations (X3100 and X4500) that they are really going to improve their drivers. BTW, BFG - Nice work on your Anti Aliasing Comparison. I thought it was well written, easy to understand. I learned a few things that other reviewers didn't explain very well in the past.

 

[OCGuy]

Originally posted by: Bob Dylan

:music: the times, they are a changin' :music:

 

[SickBeast]

If that performance estimate is true, then the Larrabee will be faster than I thought it would be. That said, it's not good enough. IMO intel is releasing this thing in a desperate attempt to keep x86 relevant. IMO it will fail.

 

[ilkhan]

If they can match the performance of a GTX285 at 45nm, I'll be impressed when they transition to 32nm and double the number of cores.

 

[ronnn]

If true, likely the end of at least one company.

 

[Zstream]

Originally posted by: BFG10K
I doubt this is anywhere near true. Even if the hardware is that fast (which I doubt), the GTX285 is also fast because of good drivers. Based on Intel?s track record with the GMA, I don?t see good drivers coming into the picture for a long time.

If we are discussing "game performance", then I say who cares! The meaning of this chip is not to dominate games or have real any impact in that segment.

Everyone realizes the CPU business will be inherently decreased in the future. The cpu will change completely from what it is today. The CPU of tomorrow will be small, quick in multi-path decisions and have huge cache sizes to store the decisions. Its job is to take commands and process them to determine the proper route. In some cases you can say route the traffic to the proper core. Imagine a cpu taken from the cell phone and using that as a base CPU. No more hard calculations performed on the cpu.

Then we have Larabee, it is a design process in which this device in and all components equal the new CPU/GPU/SOUND, go far as possible with Southbridge added.

In a few more years you can expect to see similar AMD cards and chips. They will have to rely upon OpenCL for that. It will be the battle of OpenCL/Cuda vs High level programming language.


My two .cents

 

[Fox5]

Which Larabee is expected to perform on par with a GTX285? There's more than one SKU, and the highest end one is using way more silicon than the gtx 285. It'd be almost more appropriate to compare it to a dual gpu setup.

 

[Sylvanas]

Intel have been on a development extravaganza lately. A year ago they were recruiting driver devs and just recently they setup that institute for visual computing or whatever- I wouldn't be at all surprised if it launches equal to a GTX 285.

 

[Scali]

Originally posted by: SickBeast
If that performance estimate is true, then the Larrabee will be faster than I thought it would be. That said, it's not good enough. IMO intel is releasing this thing in a desperate attempt to keep x86 relevant. IMO it will fail.

Intel have always said that they aim at midrange performance.
If the card is introduced in about 6 months, GTX285 performance may be midrange by then, with the new DX11 generation cards above it.

As for their GMA drivers... I have a laptop with X3100, and although it initially had tons of problems (no DX10 support at all, weird 'tearing' of the desktop at times (probably when powersave mode changed), Flash video acceleration broken, not being able to run most games/DX SDK stuff etc), Intel has really come a long way in the past 12 months. They've fixed pretty much all the issues I've encountered, and I now actually use my laptop for DX10 development from time to time.

Sadly they seem to have broken OpenGL completely though, a few driver releases ago. But I rarely need OpenGL, so it doesn't bother me much. It's strange though, they had a reasonable implementation with OGL 2.0 support and about 63 supported extensions. Now it seems that it's all gone, I get OGL 1.1 with 2 extensions, seems to be a standard software driver from Microsoft?

 

[zagood]

Originally posted by: Zstream
If we are discussing "game performance", then I say who cares!

that's the big question though. I wouldn't doubt that larabee could match a GTX285 in cuda/stream/gpgpu tasks, but that stupid question pops up, "can it play crysis?"

i'd just like to see something concrete to figure out what the hell this chip is supposed to do. with all new technologies it just comes down to two questions...

1. Will it make my life better?
2. Can I afford it?

That's all I care about.

 

[Scali]

Originally posted by: zagood
that's the big question though. I wouldn't doubt that larabee could match a GTX285 in cuda/stream/gpgpu tasks, but that stupid question pops up, "can it play crysis?"

I'm quite sure they were talking about gaming performance, not GPGPU.

Originally posted by: zagood
i'd just like to see something concrete to figure out what the hell this chip is supposed to do.

It shouldn't take too long now. Intel has said that they want to start demoing the chip in Q3/4 and launch it in early 2010. So the first demos shouldn't be that far off now.
It will be especially interesting to see how differently this GPU behaves from the more conventional AMD/nVidia offerings. I bet it will have completely different performance scaling behaviour. It may be much better than conventional GPUs at some tasks, and much worse at others. Question is how it will balance out in practice.

 

[Keysplayr]

Originally posted by: SickBeast
If that performance estimate is true, then the Larrabee will be faster than I thought it would be. That said, it's not good enough. IMO intel is releasing this thing in a desperate attempt to keep x86 relevant. IMO it will fail.

If it wasn't for football, I would not be playing football today.

Whatever Larrabee is, another alternative is always better. Whether it has GTX285, 9500GT performance or anywhere in between. And besides, they didn't really specify HOW Larrabee was "equal" to the performance level of a GTX285. Was it in gaming? GPGPU? Weight? LOL.

 

[Scali]

Originally posted by: Keysplayr
Whatever Larrabee is, another alternative is always better. Whether it has GTX285, 9500GT performance or anywhere in between.

It seems almost like a pre-emptive strike.
For years we've had two trends:
CPU: highly programmable architecture ----> more and more parallelism ----> highly parallel architecture
GPU: highly parallel architecture ----> more and more programmability ---> highly programmable architecture

So basically CPUs and GPUs have opposite starting points, but are moving in eachother's direction.
They're bound to meet halfway at some point.
Intel approaches it from the programmable side, where the entire 'GPU' is just a 'driver' for the underlying hardware, which has little specific functionality for graphics.
AMD and nVidia on the other hand are trying to replace bits and pieces of specific functionality by generic programmable hardware (the unified shaders being the first step here, removing the difference between different types of shaders, and now compute shaders will allow you to do even more with just 'software').

There is a chance that Intel may be ahead of its time right now, but this is just a side-project for Intel. They don't rely on the success of Larrabee as a company. Therefore, they can continue to develop and improve the design over the years, and at some point Intel will 'get it right', and the timing will be optimal for a highly programmable parallel processor/GPU. AMD and nVidia don't really have this luxury. They have to 'get it right' on every architecture or else they'll likely go out of business like all the other great GPU manufacturers of the past.

 

[Rusin]

If Larrabee's performance is only at GTX 285 level..then it's huge dissapointment. Basically Nvidia wouldn't even need to release new cards to counter it!

Larrabee is huge and very expensive chip to make. One 300mm wafer could hold only about 64 Larrabees when for comparison one 300mm wafer could hold 94 Nvidia's 65nm GT200 chips. Which means that Larrabee couldn't fight against GTX 285 price wise. Also Larrabee is said to have HD4870X2..ish.. power consumption which would mean over 100W higher realistic peak consumption compared to GTX 285.
---

Then there's next generation and if any of those rumours are true then Larrabee's first incarnation would be DOA:

Nvidia's GT300 should be cheaper to make..even when it's pretty large. Though they say it would have 2.3 times higher raw power than GTX 285..if there's no advancements in efficiency.
AMD's Evergreen is around 180mm^2..over 3.5 times smaller than Larrabee, and this chip would have better performance than Larrabee?.. AMD could stick four Evergreens on one card and it would still be cheaper than single Larrabee, but they only need one.

And talking about Intel's 45nm -> 32nm transition. Simple die shrink wouldn't be enough to get Larrabee to fighting shape. Specially when Nvidia's and AMD's transition between 40nm -> 28nm would happen really soon after that.

 

[Scali]

Originally posted by: Rusin
Larrabee is huge and very expensive chip to make. One 300mm wafer could hold only about 64 Larrabees when for comparison one 300mm wafer could hold 94 Nvidia's 65nm GT200 chips. Which means that Larrabee couldn't fight against GTX 285 price wise.

You don't know that.
nVidia has to have a third party like TSMC manufacture their GPUs. Intel does everything in-house, and on a huge scale. Intel's production facilities are also more advanced than TSMC's, and their 45 nm process is very successful and mature, where TSMC is struggling with 40 nm, and 55 nm is still the bread-and-butter of nVidia and AMD GPUs.

Aside from that, yes you may get 94 GT200 chips out of a 300mm wafer, but how many of those are GTX285? Probably a minority, as most of them are salvaged as GTX260/275/295.
So it's really hard to make a comparison between nVidia and Intel... There are so many factors involved here.

Intel has proven before that they can beat a competitor with a much larger chip. The Pentium 4/D were much larger than the competing Athlon XP/64/X2 processors. Instinctively you might think that AMD would get the lowest prices and/or the highest profits, but in fact it was Intel undercutting AMD's prices with the Pentium D series, and Intel had the better business results during that era, so they would have had the higher profit margins.
I think Intel will take advantage of their production capabilities in a similar vein here.

 

[WelshBloke]

Originally posted by: Scali

Originally posted by: Rusin
Larrabee is huge and very expensive chip to make. One 300mm wafer could hold only about 64 Larrabees when for comparison one 300mm wafer could hold 94 Nvidia's 65nm GT200 chips. Which means that Larrabee couldn't fight against GTX 285 price wise.

You don't know that.
nVidia has to have a third party like TSMC manufacture their GPUs. Intel does everything in-house, and on a huge scale. Intel's production facilities are also more advanced than TSMC's, and their 45 nm process is very successful and mature, where TSMC is struggling with 40 nm, and 55 nm is still the bread-and-butter of nVidia and AMD GPUs.

Aside from that, yes you may get 94 GT200 chips out of a 300mm wafer, but how many of those are GTX285? Probably a minority, as most of them are salvaged as GTX260/275/295.
So it's really hard to make a comparison between nVidia and Intel... There are so many factors involved here.

Intel has proven before that they can beat a competitor with a much larger chip. The Pentium 4/D were much larger than the competing Athlon XP/64/X2 processors. Instinctively you might think that AMD would get the lowest prices and/or the highest profits, but in fact it was Intel undercutting AMD's prices with the Pentium D series, and Intel had the better business results during that era, so they would have had the higher profit margins.
I think Intel will take advantage of their production capabilities in a similar vein here.

Yep. If Intel can compete with a GTX 285 (I personally dont think they will) they have the ability to price it so low that it'll sell very well.

AMD/Nvidia have no way in hell of competing pricewise against Intel (if Intel want to muscle in), they both need to be much faster to stay in the market.

 

[Keysplayr]

You know, I'm still not quite sure what Larrabee is, from a physical standpoint. Is it a discrete PCI-e card? Is it a CPU? Will a specialized motherboard with a new socket be needed? Anyone know? I haven't really been keeping up. It seems like Larrabee has been in development for 30 years now. It got boring.

 

[Scali]

Originally posted by: Keysplayr
You know, I'm still not quite sure what Larrabee is, from a physical standpoint. Is it a discrete PCI-e card? Is it a CPU? Will a specialized motherboard with a new socket be needed? Anyone know? I haven't really been keeping up. It seems like Larrabee has been in development for 30 years now. It got boring.

If I'm not mistaken, Larrabee is the codename for Intel's GPGPU architecture (such as G300 or RV870), which is a massively parallel in-order x86-based architecture, and a derivative of the Terascale project.
However, this GPGPU is going to be placed on a PCI-e card with dedicated video memory, just like your average discrete graphics card from AMD, nVidia and others.
People often refer to the card as a whole as Larrabee aswell, because we lack a better name (just like how the codenames for future AMD and nVidia GPUs float around on the web, but we don't really know if the cards they will be on will be called Radeon HD5870 or GeForce GTX385 or such...)

After the initial launch of Larrabee as a discrete card, Intel will also integrate the technology in their future IGPs, which will be integrated into the CPU package, rather than the chipset (a CPU with integrated GPU will probably be launched later this year, but this is still based on their current grahpics technology, not Larrabee). There will be a new socket introduced for CPUs with integrated graphics, but there isn't going to be a second socket on the motherboard (CPU + GPU), if that's what you were thinking.

 

[Genx87]

Matches the GTX285 in performance doing what?

 

[BFG10K]

Originally posted by: ArchAngel777

GMA still has pathetic drivers despite Intel telling us for the last 2 generations (X3100 and X4500) that they are really going to improve their drivers.

Most definitely, especially with OpenGL. I saw some driver queries being run on a GMA and it failed almost every test it was given.

Intel recruiting driver programmers is one thing, but to expect the initial release of a device that is essentially nothing like what we?ve seen before to match nVidia?s current flagship is ridiculous. In fact I?d be absolutely floored if it could even match nVidia?s three year old 8800 GTX.

Look at how many driver issues the G80 had at launch, and this from nVidia?s veteran driver programmers with years of experience that were programming a device that had simply moved from a split of pixel/vertex shaders to unified shaders.

BTW, BFG - Nice work on your Anti Aliasing Comparison. I thought it was well written, easy to understand. I learned a few things that other reviewers didn't explain very well in the past.

Thanks for the kind words; I can?t wait to do the write-up for nVidia?s hardware.

 

[BFG10K]

Originally posted by: Zstream

If we are discussing "game performance", then I say who cares!

Considering gaming is by far the biggest market these parts compete in, anyone that wants to be a viable discrete GPU competitor should care. Remember that Larrabee is a discrete PCIe part, not some kind of new CPU socket.

The meaning of this chip is not to dominate games or have real any impact in that segment.

Then it can?t really compete with ATi?s or nVidia?s GPUs.