SemiAccurate Forums  

 
Go Back   SemiAccurate Forums > Search Forums
Reload this Page Search Results
Register FAQ Members List Social Groups Calendar Search Today's Posts Mark Forums Read

Page 1 of 5 1 2345
Showing results 1 to 25 of 119
Search took 0.00 seconds. Search: Posts Made By: Lee
Forum: Finance, Legal, Policy 07-23-2012, 08:38 PM
Replies: 115
Views: 17,645
Posted By Lee
Look at that VLIW4 CU another way. What if you...

Look at that VLIW4 CU another way. What if you picture it as a SIMD unit containing 4 vector ALUs with a single, statically structured, instruction stream to feed them. Same memory banking, same...
Forum: Software 04-13-2012, 10:19 AM
Replies: 51
Views: 13,904
Posted By Lee
This isn't really true. Most class methods are...

This isn't really true. Most class methods are fairly easily inlined and you end up with roughly the same code you started with. This is also true of iterators and the like. There is *no* memory...
Forum: GPUs 09-12-2011, 10:43 AM
Replies: 16
Views: 3,350
Posted By Lee
Some of that is true for any vector machine with...

Some of that is true for any vector machine with gather instructions though. You don't need to vectorise loads per lane on AMD chips, but you benefit a lot from it (varying degrees depending on...
Forum: GPUs 09-11-2011, 07:53 PM
Replies: 16
Views: 3,350
Posted By Lee
Well, you say that, but in reality in both cases...

Well, you say that, but in reality in both cases you need to vectorise your loads. It's just that the vector operations are more visible in the AMD case because you can see them in the IL. The vector...
Forum: GPUs 09-10-2011, 07:56 PM
Replies: 16
Views: 3,350
Posted By Lee
I think you're right that that's so close as to...

I think you're right that that's so close as to be meaningless. The effect is the same except maybe in terms of power saving. You wouldn't even notice the difference. The only meaningful mask is the...
Forum: GPUs 08-27-2011, 04:55 PM
Replies: 2,460
Views: 455,076
Posted By Lee
In saying that you make the implicit assumption...

In saying that you make the implicit assumption that compute and games are different. As Eric Demers pointed out in the AFDS keynote that isn't really true.
Forum: GPUs 06-28-2011, 04:35 PM
Replies: 1,523
Views: 276,121
Posted By Lee
But if you talk to NVIDIA's engineers, or even...

But if you talk to NVIDIA's engineers, or even Bill Dally when he's in a non-marketing mood (he was far more responsive to this debate than David Kirk was), they'll tell you it's a vector...
Forum: GPUs 06-27-2011, 06:22 PM
Replies: 1,523
Views: 276,121
Posted By Lee
It really is. AMD describes their architecture...

It really is. AMD describes their architecture the same way. I prefer to think of AMD's architecture as a VLIW issue of 4 16-wide vector instructions rather than vice versa as they describe it. It's...
Forum: OSes 06-06-2011, 02:15 PM
Replies: 95
Views: 16,697
Posted By Lee
Sure, but that's because the UI has been designed...

Sure, but that's because the UI has been designed with a low power CPU in mind. Intel is suggesting a way you can make the windows UI more efficient - basically by making a UI their way, bypassing...
Forum: OSes 06-06-2011, 11:14 AM
Replies: 95
Views: 16,697
Posted By Lee
I don't understand peoples complaints here... the...

I don't understand peoples complaints here... the x86 chips are higher performance than the ARM chips available at the moment. There's little doubt about that. So what makes them slow? It's clearly...
Forum: GPUs 05-25-2011, 11:02 AM
Replies: 2,460
Views: 455,076
Posted By Lee
It's not an implicit memory trick. GPUs are...

It's not an implicit memory trick. GPUs are vector architectures. It's like programming to AVX instead of SSE - you're locking in to the ISA provider maintaining AVX. If the provider wants to narrow...
Forum: GPUs 05-24-2011, 09:13 AM
Replies: 2,460
Views: 455,076
Posted By Lee
Yes, but it's necessary to come remotely close to...

Yes, but it's necessary to come remotely close to peak performance. It's a vector machine so you have to treat it like a vector machine if you want to get performance out of it. Work items aren't...
Forum: GPUs 05-23-2011, 03:50 PM
Replies: 2,460
Views: 455,076
Posted By Lee
It's very much a backward compatibility issue...

It's very much a backward compatibility issue because a lot of code does not quite following the OpenCL spec of treating each lane as a "thread" because the reality is that that is horrendously...
Forum: GPUs 05-19-2011, 07:41 PM
Replies: 2,460
Views: 455,076
Posted By Lee
I don't think that's necessarily true. NVIDIA may...

I don't think that's necessarily true. NVIDIA may not have a performance benefit because there isn't as much of a performance benefit from this scheduling "improvement" as some people here suggest....
Forum: GPUs 05-19-2011, 12:52 PM
Replies: 2,460
Views: 455,076
Posted By Lee
Actually, I think of it the other way around much...

Actually, I think of it the other way around much more like the CPU core itself. The Intel core issues two SIMD instructions at once. The Cayman architecture issues four 64-wide SIMD instructions at...
Forum: GPUs 04-29-2011, 08:08 PM
Replies: 49
Views: 6,700
Posted By Lee
4xxx is compliant implementation because the spec...

4xxx is compliant implementation because the spec does not require, for example, __local to be backed into scratchpad. __local on 4xxx is in global memory (the hardware LDS is owner writes which is...
Forum: CPUs 02-21-2011, 03:55 PM
Replies: 133
Views: 29,719
Posted By Lee
That's only true if you ignore the fact that GPU...

That's only true if you ignore the fact that GPU code is vectorised by design. You're running a 64-wide vector, I'm sure you can write a 2-wide vector for the same application.
Forum: Software 02-13-2011, 03:48 PM
Replies: 23
Views: 8,344
Posted By Lee
Define "next version" and how big a change you...

Define "next version" and how big a change you expect to see. I think Khronos has been arguing that OpenCL is on an 18 month version cadence, which should give a good clue.
Forum: CPUs 02-10-2011, 08:15 PM
Replies: 13
Views: 3,019
Posted By Lee
Yes, but as it doesn't give any significant...

Yes, but as it doesn't give any significant advantage in most situations, is the necessary effort really on marketing or engineering?
Forum: CPUs 02-10-2011, 07:23 PM
Replies: 13
Views: 3,019
Posted By Lee
Well, Bullet has cloth acceleration to a degree...

Well, Bullet has cloth acceleration to a degree now. 3DMark 11 used it.

On the other hand, in general it's hard to see that GPU acceleration of physics has real benefits until communication...
Forum: Article discussion 02-10-2011, 05:05 PM
Replies: 52
Views: 6,455
Posted By Lee
You're right. It will definitely be interesting...

You're right. It will definitely be interesting to see who is hired as a full time replacement.
Forum: Article discussion 02-10-2011, 03:57 PM
Replies: 52
Views: 6,455
Posted By Lee
I don't see any reason to think AMD's public...

I don't see any reason to think AMD's public fusion roadmap is less competent than, say, Intel's. It's certainly more thorough than nvidia's public roadmap. Excluding Tegra comparisons, which we know...
Forum: Software 02-08-2011, 12:41 PM
Replies: 23
Views: 8,344
Posted By Lee
Writing a specialised tool like that is a...

Writing a specialised tool like that is a challenge, but it clearly is the right option for a lot of companies. A high level library with efficient, future proof, code gen for multiple targets is a...
Forum: GPUs 12-12-2010, 05:27 PM
Replies: 25
Views: 5,298
Posted By Lee
Scatter/gather etc is too hard in an OoO...

Scatter/gather etc is too hard in an OoO pipeline, apart from general efficiency losses doing dependence analysis. I'm not sure we'll ever see that, having a few threads running is just more...
Forum: GPUs 12-09-2010, 07:14 PM
Replies: 2,690
Views: 457,484
Posted By Lee
Only if you look at raw numbers. Cypress is FPU...

Only if you look at raw numbers. Cypress is FPU heavy, it can easily lose a few because except in very rare cases it couldn't utilise them anyway. If Cayman uses the die area to boost primitive rate...
Showing results 1 to 25 of 119
Page 1 of 5 1 2345

 
Forum Jump

All times are GMT -5. The time now is 11:59 AM.

Contact Us - SemiAccurate Forums - Archive - Top

Powered by vBulletin® Version 3.8.7
Copyright ©2000 - 2013, vBulletin Solutions, Inc.
SemiAccurate is a division of Stone Arch Networking Services, Inc. Copyright © 2009 Stone Arch Networking Services, Inc, all rights reserved.