| |||||||||||||
Romanick: The zombies cometh...Romanick: The zombies cometh...Posted Jul 23, 2012 22:34 UTC (Mon) by slashdot (guest, #22014)Parent article: Romanick: The zombies cometh...
What Intel should do is get off their lazy asses and start producing $100-$500+ discrete GPUs (preferably connected via QuickPath and not PCIe) better than the AMD and nVidia ones, and provide the same documentation, support and architectural stability they provide for their x86 CPUs.
Given their leadership in CPU performance and in silicon manufacturing, employment of top engineers, ownership of two highly related product lines (their crappy integrated GPUs, and Larrabee/MIC/Xeon Phi), and having 10x the capitalization of nVidia and AMD combined, they should be able to easily lead the "decent GPU" market as well, so no clue what the heck they are doing.
I mean, it's hard to imagine how one could be better positioned to enter that market.
There's the risk of ending up with a company monopolizing both the CPU and GPU markets, but at least it would be one that releases full documentation of their hardware (except for microcode format), and contributes to the relevant open source projects.
(Log in to post comments)
Romanick: The zombies cometh... Posted Jul 23, 2012 22:53 UTC (Mon) by felixfix (subscriber, #242) [Link]
I suspect Intel is a bit skittish about that monopoly part of the idea, having been thru the government wringer before.
They may also think it more worthwhile to be neutral in the GPU wars, as Google is (used to be?) in the Android hardware arena.
Romanick: The zombies cometh... Posted Jul 23, 2012 23:29 UTC (Mon) by AndreE (subscriber, #60148) [Link]
Except of course that their two "highly related" product lines are barely related to the high performance offerings offered by NVidia and AMD. Larabee was a colossal waste of time and money, which also showed that graphics performance is a specialised problem that needs more than a few parrallelized general-purpose CPU cores for the task.
Yeah, Intel probably has the money to attack the space head on, but it certainly won't be "easy" as you so glibly claim. AMD and NVidia also have great engineers and a wealth of experience in producing such hardware.
Intel's energies (and money) are better spent trying to gain a foothold in the mobile space.
Romanick: The zombies cometh... Posted Jul 25, 2012 0:32 UTC (Wed) by Trelane (subscriber, #56877) [Link]
> Larabee was a colossal waste of time and money, which also showed that graphics performance is a specialised problem that needs more than a few parrallelized general-purpose CPU cores for the task.
Romanick: The zombies cometh... Posted Jul 26, 2012 16:32 UTC (Thu) by gmaxwell (subscriber, #30048) [Link]
Which continues to be a product you can't buy, much to the disappointment of many who have problems which would be well suited to it.
As far as I can tell MIC is primarily vaporware that exists to discourage parties from investing in the development to port applications to CUDA, "If we wait a bit, we can use MIC and have a reduced porting risk/reward curve". Even if not, it's still premature to call it evidence of success.
Romanick: The zombies cometh... Posted Jul 30, 2012 23:19 UTC (Mon) by Trelane (subscriber, #56877) [Link]
> it's still premature to call it evidence of success.
Sure, but OP was claiming,
> Larabee was a colossal waste of time and money, which also showed that graphics performance is a specialised problem that needs more than a few parrallelized general-purpose CPU cores for the task.
Which is equally untrue. Reality is much more interesting.
Romanick: The zombies cometh... Posted Jul 26, 2012 8:07 UTC (Thu) by russell (subscriber, #10458) [Link]
Graphics would be an excellent move for Intel. Why let nVidia, et al, use that lucrative market to bank roll themselves into adjacent markets. Intel should strangle that cash cow before it's too late.
Romanick: The zombies cometh... Posted Jul 24, 2012 5:46 UTC (Tue) by josh (subscriber, #17465) [Link]
Intel provides full GPU documentation: http://intellinuxgraphics.org/documentation.html
And as for making discrete graphics, why? More powerful graphics, sure, but why bother making them discrete? Integrated actually provides *more* opportunities for optimization, and sharing memory between the CPU and GPU provides significant advantages, such as zero-copy transfers between the two.
Romanick: The zombies cometh... Posted Jul 24, 2012 5:56 UTC (Tue) by JoeBuck (subscriber, #2330) [Link] Exactly. It takes a lot of power to drive high-speed signals off-chip, and it's easy to put more cores on a single die than you know what to do with. Discrete GPUs are going away, which leaves nVidia with a big problem.
Romanick: The zombies cometh... Posted Jul 24, 2012 6:07 UTC (Tue) by slashdot (guest, #22014) [Link]
By "discrete" I actually meant "as powerful and big as the Radeon and GeForce chips".
If they can do that by putting it on the CPU die or package, that's even better, but it seems unlikely.
I think a high-performance GPU operating as a cache-coherent NUMA node connected by QuickPath/HyperTransport is a more realistic goal, and that would be almost as good as on-die.
Romanick: The zombies cometh... Posted Jul 24, 2012 6:34 UTC (Tue) by kragil (subscriber, #34373) [Link]
Intel is certainly trying, but ATI and Nvidia have a big headstart. John Carmack (of Doom fame) said that Intel chips will eventually have really good performance for games. He had a good reason, but I can't find that interview anymore. So maybe in 2014 AMD won't rule the integrated gfx marked anymore (with regard to performance)
And just for the record: Intel did discrete graphics once, AND THEY SUCKED BIGTIME. http://en.wikipedia.org/wiki/Intel740
Romanick: The zombies cometh... Posted Jul 24, 2012 17:20 UTC (Tue) by bhassel (subscriber, #68210) [Link]
Might be this interview? http://www.pcper.com/reviews/Editorial/John-Carmack-Inter...
John Carmack:
Romanick: The zombies cometh... Posted Jul 24, 2012 7:58 UTC (Tue) by drag (subscriber, #31333) [Link]
I don't really see much of a point for Intel doing anything other then integrated graphics. They already have well over 80% of the market wrapped up in terms of graphics chips.
As integrated graphics push up into mid-range discrete graphics card performance then that will eliminates the majority of what is left of the discrete market sales.
Romanick: The zombies cometh... Posted Jul 24, 2012 8:38 UTC (Tue) by epa (subscriber, #39769) [Link]
Yes, I wonder about the yield on their Core CPUs with integrated graphics, and whether the ones where the CPU part is defective but the GPU is okay might be cheaply stuck onto a PCI Express card with some RAM and driven with essentially the same drivers they have now.
Romanick: The zombies cometh... Posted Jul 27, 2012 4:34 UTC (Fri) by BenHutchings (subscriber, #37955) [Link] While it would theoretically be possible to design the chip that way, I very much doubt that it would be worth the effort. Current Intel IGPs are not real PCIe endpoints (notice how they don't work with PCI IOMMUs). They would need to have that functionality added as a strap option, and the chip's PCIe link would need to be switchable between the internal root complex (CPU mode) and the IGP (discrete GPU mode). And then, why would anyone buy it anyway?
Intel graphics - it's the SW not the HW Posted Jul 25, 2012 13:16 UTC (Wed) by pflugstad (subscriber, #224) [Link]
So, I think the thing everyone seems to forget about graphics hardware is that a HUGE part of performance and features is based on the software stack they have running over the top of them. And these software stacks are extremely complex, including things like sophisticated memory management and JIT compilers for shaders.
NVidia and AMD have been evolving their software for probably over a decade now. They support multiple versions of both DirectX and OpenGL, all on the same basic hardware.
Intel on the other hand, doesn't have any of that software (that I'm aware of). They are essentially open sourcing their hardware and pushing the development of the open source software stack (Mesa, Gallium/LLVM, etc) as the preferred software solution for their hardware (at least on Linux).
So, while I expect Intel could build a piece of hardware that performs better at a hardware level than AMD/Nvidia could, by dint of their process superiority, the software side of things is still several (3-5?) years behind. As a result, they'd probably still lag behind significantly in performance. And all the spiffy hardware in the world won't fix that.
Intel graphics - it's the SW not the HW Posted Jul 25, 2012 13:29 UTC (Wed) by Cyberax (✭ supporter ✭, #52523) [Link]
Intel doesn't use Gallium3D or LLVM. They're developing shader compiler essentially from scratch.
Intel graphics - it's the SW not the HW Posted Jul 25, 2012 13:41 UTC (Wed) by gioele (subscriber, #61675) [Link]
> Intel doesn't use Gallium3D or LLVM. They're developing shader compiler essentially from scratch.
Ah, irony. The biggest open source contributor to the graphic subsystem and the one with the most updated drivers is developing against the older (hopefully one day deprecated) graphic stack.
Intel graphics - it's the SW not the HW Posted Jul 25, 2012 20:10 UTC (Wed) by pflugstad (subscriber, #224) [Link]
> Intel doesn't use Gallium3D or LLVM. They're developing shader compiler essentially from scratch.
Which makes my point even more valid. The SW stack running over the top of the Intel hardware is basically 3-5 years (at least) behind Nvidia/AMD. So regardless of how good a graphics card Intel actually does make, the SW is just not there to support it.
Intel graphics - it's the SW not the HW Posted Jul 25, 2012 20:55 UTC (Wed) by Cyberax (✭ supporter ✭, #52523) [Link]
That's not entirely correct. Intel was the main driving force behind the push to OpenGL 3 compatibility. They've started the shader compiler rewrite, this compiler is actually shared between the "classic" Mesa and Gallium3D.
So the situation is much more nuanced and complex than you think :)
Intel graphics - it's the SW not the HW Posted Jul 28, 2012 16:43 UTC (Sat) by Jonno (subscriber, #49613) [Link]
> The SW stack running over the top of the Intel hardware is basically 3-5 years (at least) behind Nvidia/AMD.
Actually, Intel uses a stack with an older *base* than AMD/Nouveau, but the Intel stack is still way more *polished*, and therefore currently offers better hw utilization than the AMD and Nouveau drivers. Of course, the Intel hw is so much slower than AMD and Nvidia hw, that they still loose all benchmarks, despite their better drivers.
However, this gap in polish is slowly shrinking as Gallium improves, and all new mesa state trackers (ie OpenVG, OpenCL, d3d1x, xorg, vdpau, etc) requires way more work to implement in Intel drivers, so sooner or later Intel will have to switch to Gallium if they want to keep up...
|
|||||||||||||