Slashdot Mirror
Nvidia Claims Intel's Larrabee Is "a GPU From 2006"
Barence sends this excerpt from PC Pro:
"Nvidia has delivered a scathing criticism of Intel's Larrabee, dismissing the multi-core CPU/GPU as wishful thinking — while admitting it needs to catch up with AMD's current Radeon graphics cards. 'Intel is not a stupid company,' conceded John Mottram, chief architect for the company's GT200 core. 'They've put a lot of people behind this, so clearly they believe it's viable. But the products on our roadmap are competitive to this thing as they've painted it. And the reality is going to fall short of the optimistic way they've painted it. As [blogger and CPU architect] Peter Glaskowsky said, the "large" Larrabee in 2010 will have roughly the same performance as a 2006 GPU from Nvidia or ATI.' Speaking ahead of the opening of the annual NVISION expo on Monday, he also admitted Nvidia 'underestimated ATI with respect to their product.'"
Good, learn from that and don't make that same mistake again!
Larrabee [...] will have roughly the same performance as a 2006 GPU from Nvidia or ATI.'
DOH!
Belief is the currency of delusion.
No wonder it's so slow. He keeps making reference to how it paints things. Can't move on to another frame until the previous one has dried.
So why is NVIDIA on the defensive?
Intel is aiming at number crunchers (note that their chip uses doubles, not floats). They don't want NVIDIA to steal that market with CUDA.
When Intel says "graphics", they mean movie studios, etc.
If Larrabee eventually turns into a competitor for NVIDIA, all well and good, but that's not their goal at the moment.
No sig today...
Lots of people here and analysts have written off AMD. I think AMD is in a great position if they can survive their short term debt problems which is looking increasingly likely.
Consider the following:
AMD is in a great position like no other company to capitalize on the coming CPU / GPU convergence. Everyone jeered when AMD bought ATI but it is looking to be a great strategic move if they can execute on their strategy.
AMD has the best mix of technology, they just have to put it to good use.
G. Washington on Government "it is force. Like fire, it is a dangerous servant and a fearful master."
From the SIGGRAPH paper they need something like 25 cores to run GoW at 60Hz. That's 1Ghz cores for comparison though. LRB will probably run at something like 3Ghz, meaning you only need like 8-9 cores to run GoW at 60, and with benchmarks stretching up to 48 cores you can see that this has the potential of being very fast indeed.
More importantly, the LRB has much better utilization since there aren't any fixed function divisions in the hardware. E.g. most of the time you're not using the blend units. So why have all that hardware for doing floating point maths in the blending units when 99% of the time you're not actually using it? On LRB everything is utilized all the time. Blending, interpolation, stencil/alpha testing etc. is all done using the same functionality, meaning that when you turn something off (like blending) you get better performance rather than just leaving parts of your chip idle.
I'd also like to point out that having a software pipeline means faster iteration, meaning that they have a huge opportunity to simply out-optimize nvidida and amd, even for the D3D/OGL pipelines.
Furthermore, imagine intel suppyling half a dozen "profiles" for their pipeline where they optimize for various scenarios (e.g. deferred rendering, shadow volume heavy rendering, etc. etc.). The user can then try each with their games and run each game with a slightly different profile. More importantly, however, is that new games could just spend 30 minutes figuring out which profile suits them best, set a flag in the registry somewhere, and automatically get a big boost on LRB cards. That's a tiny amount of work to get LRB-specific performance wins.
The next step in LRB-specific optimizations is to allow developers to essentially set up a LRB-config file for their title with lots of variables and tuning (remember that LRB uses a JIT compiled inner-loop that combines the setup, tests, pixel shader etc.). This would again be a very simple thing to do (and intel would probably do it for you if your title is high profile enough), and could potentially give you a massive win.
And then of course the next step after that is LRB-specific code. I.e. you write stuff outside D3D/OGL to leverage the LRB specifically. This probably won't happen for many games, but you only need to convince Tim Sweeney and Carmack to do it, and then most of the high profile games will benefit automatically (through licensing). My guess is that you don't need to do much convincing. I'm a graphcis programmer myself and I'm gagging to get my hands on one of these chips! If/when we do I'll be at work on weekends and holidays coding up cool tech for it. I'd be surprised if Sweeney/Carmack aren't the same.
I think LRB can be plenty competitive with nvidia and amd using the standard pipelines, and there's a very appealing low-fricion path for developers to take to leverage the LRB specifically with varying degrees of effort.
>ATI hanging onto a somewhat insignificant market share.
C'mon, 17 million units shipped in a quarter and ~20% of the market is hardly 'a somewhat insignificant market share' in a market with four major players (Intel, nVidia, VIA).
For comparison, take Matrox, they have insignificant market share with about 100K/q
Belief is the currency of delusion.
Ten years ago, the Riva TNT was yet a few months away. S3 and ATI both had a great marketshare for low to mid-end, and 3dfx dominated the very top segment for gamers.
Although I appreciate the attention from NVIDIA and Slashdot, I can't support that alleged quote from my blog (http://speedsnfeeds.com).
First, what's being described as a quote is actually just John Montrym's summary from my original post, which is here:
http://news.cnet.com/8301-13512_3-10006184-23.html
What I actually described as equating to "the performance of a 2006-vintage... graphics chip" was a performance standard defined by Intel itself-- running the game F.E.A.R. at 60 fps in 1,600 x 1,200-pixel resolution with four-sample antialiasing.
Intel used this figure for some comparisons of rendering performance. If Larrabee ran at 1 GHz, for example, Intel's figures show that it would take somewhere from 7 to 25 Larrabee cores to reach that 60 Hz frame rate.
Larrabee will probably run much faster than that, at least on desktop variants.
Well... rather than writing the whole response here, I think I'd rather write it up for my blog and publish it there. Please surf on over and check it out:
http://news.cnet.com/8301-13512_3-10024280-23.html
Comments are welcome here or there.
. png