Transistor , How it works ?

nice explanation......

Vittal Talawar

We are planning for a video on MOSEFT. It will take 3-4 months to get it materialized.

Lesics

if yes plz send?

simran raut

have any video about body effect in mosfet?

simran raut

Sure, we will start working on it in 2 months.

Lesics

plz post the video for "TRANSISTOR AS A SWITCH"

emmanuel

I should emphasise, that's total speculation on my part. :D But it's exactly the sort of custom tech SGI used to provide to the likes of Lockheed, DoD, etc. The Lockheed guy told me the normal geometry board in SGI's IR2 gfx contained 4 GE chips, but Lockheed had SGI build a version that had 10 (I've no reason to doubt him, the boss of Lockheed came to visit him at one point when I knew him at a research lab in 1994). In the 1990s, high-end expensive gfx tech was SGI's forte, offering future performance 'now' for a big price ("What can we build you guys with a budget of 50K?", was a line one SGI guy told me). A good Onyx2 IR setup would costs hundreds of thousands. I have an old Onyx RE2 rack (was 24-CPU) in my garage from the early 1990s which originally cost 1M (ie. about $1.5M); I bought it in the early 2000s for 200 UKP. :D In 2008 I bought a 36-CPU Onyx3800 IR3 from SPI in CA for 3K (1K + shipping!) which was originally probably about $1M. Given such mega budgets that govts and big corps are happy to play around with, it makes perfect sense that these institutions today would say to NVIDIA, if we paid you 250K (or whatever), can you build us a custom Quadro that combines a bucket load of GPUs and has 1TB VRAM (or whatever). It's all perfectly doable. SGI already had parallel gfx pipe scalability to 16 pipes with IR4 (dates from 2002, 1GB TRAM and 10GB VRAM per pipe) and they were about to increase it to 256 pipes before their gfx business went bellyup (yet people think their PCs are so fancy when using 4 GPUs in CF/SLI! :D). Today the bandwidths available are way higher with PCIe and other methods, memory is much faster and cheaper, so no reason why such a thing couldn't be done. One can already stuff a bunch of Quadros into an SGI UV system or whatever, so it's not really a stretch to imagine NVIDIA could build a custom edition with a lot more GPUs and RAM, even if it's something that literally connects to more than one PCIe slot and has NVLink built-in already, or perhaps an external box just like the PCIe splitter boxes which people use to exploit a lot more than 4 GPUs for rendering (the top spot on the Octane Render benchmark page uses ten GTX 980 Tis). SGI switched to XEONs and Linux for its later systems, but retained the architecture from the older MIPS/IRIX systems (shared memory, cache coherent, massive I/O, low latency, with added MPI offload and other neat stuff), so today the UV 3000 supports up to 256 sockets (ie. XEON E5, max 4096 cores, 8192 threads) and 64TB RAM (the maximum possible due to the XEON's 46bit address bus, SGI is merely waiting for Intel to increase this in future XEON designs), with multiple internal NUMAlink7 I/O links each offering 15GB/sec, giving a total system bandwidth of several TB/sec: <a href="http://www.sgi.com/pdfs/4555.pdf" rel="nofollow noopener" target="_blank">http://www.sgi.com/pdfs/4555.pdf</a> ...but SGI doesn't make gfx tech anymore. Nevertheless, the same application demands from govt, sciences, defense, oil/gas, automotive, medical, aerospace, etc. are still there and if anything way stronger than they were 20 years ago, probably several orders of magnitude more intense. Plus there's data mining and visualisation which has come of age. Lots of SGI people moved to NVIDIA (not sure how many are still there though), indeed the guy I knew at SGI who was their global desktop PR manager (Ujesh Desai, my contact at SGI when I was doing post-launch bug-hunting/testing of their O2 system for them) was until recently NVIDIA's VP of product/corporate marketing (he's since moved to Logitech), and most of SGI's low-end gfx team moved to NVIDIA as part of an IP settlement in the mid 2000s, as did many of SGI's high-end gfx team. NVIDIA almost certainly has all the gfx design info SGI had, including the never-released IR5 stuff. Heck, the first ever proper GE-based consumer GPU, the Geforce256, is based on SGI's IR gfx set, just stripped down with all the high-end features like 48bit RGBA removed (anything not needed for games basically), eg. the 32bit shaded GLperf results for IR and the GF256 are identical. And just to mix things up, later on a number of ex-SGI people at NVIDIA moved to AMD, spreading the old SGI knowledge even further. :D SGI used to focus on advanced features and via big budgets started off with a huge leap in performance over everyone else (their early 1990s gfx tech supported 48bit RGBA and full-scene 8x AA many years before PCs even *had* 3D gfx, see my site for the docs about RealityEngine; likewise, their MaxIMPACT gfx in the mid 1990s cost $8000 and a complete Indigo2 system could be $50K, but it was 10X faster than anything else available at the time), but eventually via numerous management mistakes they lost their early performance lead. And the market changed, PC tech improved, people wanted lower quality raw speed at far lower cost ("32bit shaded triangles please! Lots of them! Who cares about geometry precision!"), so high quality imaging and deep colour were ditched in the mainstream., while SGI never tried to catch up; they kept adding new features to IR gfx, but didn't offer anything new that was basic, cheaper but fast. Even modern pro cards still can't do some of the things SGI's old IR gfx could do, such as guaranteed frame rates, guaranteed fill rates, guaranteed hw-based system I/O to ensure uninterupted data flow for critical tasks (real-time kernel performance), hw fade LOD, etc. Modern pro stuff is slowly catching up, but it's always under the same modern umbrella of never being particularly expensive. The Quadro M6000 24GB costs 4250 UKP and the Tesla P100 is 7750 UKP, but these prices are peanuts compared to what SGI tech used to be, eg. Onyx3000 IR4 systems started at around $130K list price, but a complete 5-board IRx gfx set just on its own was about $250K (and a good spec system could be $1M+): <a href="http://www.sgidepot.co.uk/misc/3353.pdf" rel="nofollow noopener" target="_blank">http://www.sgidepot.co.uk/misc/3353.pdf</a> <a href="http://web.archive.org/web/20020802224933/http://www.sgi.com/features/2002/july/ir4/" rel="nofollow noopener" target="_blank">http://web.archive.org/web/20020802224933/http://www.sgi.com/features/2002/july/ir4/</a> <a href="http://web.archive.org/web/20020810161927/http://www.sgi.com/newsroom/press_releases/2002/july/ir4.html" rel="nofollow noopener" target="_blank">http://web.archive.org/web/20020810161927/http://www.sgi.com/newsroom/press_releases/2002/july/ir4.html</a> So imagine what NVIDIA could build if someone with the money (eg. DoD) said here's a huge budget, build us something 10X or 100X better, something the general market won't have for a decade; that's what SGI used to do. NVIDIA doesn't publicly appear to have anything like that going on, but behind the scenes I'm sure they do.

Ian Mapleson

1TB VRAM... that's mindboggling.

David

David writes: > Yes, I noticed that modern gaming GPUs are becoming quite good > at CAD software. Or I guess its more accurate to say that CAD > software is getting quite good on gaming GPUs. It's a combination of apps being recoded and the drivers being deliberately more friendly. I know someone at NVIDIA who helps write the drivers, and I talked to a guy who knows a lot about how the SPEC tests operate (he was pretty scathing actually). Of course what gamer cards can't offer is viewport precision, bit depth and other advanced features like heavily AA lines (plus the same level of reliability, warranty, support, etc.), but this is a sacrifice a heck of a lot of users and companies were increasingly willing to make in recent years, so no surprise both app vendors and GPU vendors moved to exploit the demand for a lower common demoninator. NV has bounced back & forth between catering to such demands (original Titan) or holding them off (no Titan has ECC which rules out numerous pro uses, later Titans have crippled FP64, no gamer card has a full speed PCIe return path, advanced caching features are only in the pro cards, etc.) > I ran W7000/R9 390 hybrid until about 2 weeks ago, when I bought > a 1070 to replace both. I made the call after seeing this: Looks about right. :D > Though there are a few apps that still bomb on gaming GPUs. ... Very true. SNX, SW, etc., and of course apps no longer present in Viewperf such as ProE (now I wonder why?) Hence the scathing comments from the SPEC-knowledgeable person I mentioned, he didn't like the way the suite has changed to favour non-pro cards much more, as he felt the switch is misleading buyers into thinking they gain nothing from using a pro card, ie. people are more likely to base a decision purely on performance which is often a bad idea. > ... Coollinks. I'll bookmark em for future reference. Thanks! Plenty more on my site, but I wouldn't want to make your head explode. :D Btw, the Energy benchmark in Viewperf 12 really makes me laugh, it's so unrealistic. Another comment from the guy I talked to, he said the datasets in Viewperf are typically terrible, not like real-world data at all in many cases. That page describes it as being, "representative of a typical volume rendering application in the seismic and oil and gas fields." Yeah right, representative of workloads from about 20 years ago! :D It says the, "medium-res refers to a 1GB dataset; large-res refers to a 3.2GB dataset." Large?? What total hokum. Medical datasets have evolved in the same way, moving from thick slice to thin slice at higher res as the tech has improved (and of course, a 10X increase in scanning resolution results in datasets that are three orders of magnitude larger). I talked to a guy at Chevron Nigeria in the late 1990s who told me they were already working with "typical" desktop datasets of around 750MB (rendering very quickly back then using SGI Octane MGRAS systems), while larger datasets were dealt with by a 24-CPU POWER Onyx. A decade later (but still ten years ago), SGI's PR for Onyx4/Prism described a typical full GIS set as being around 300GB. So why does the Energy benchmark test something that's much less than 4GB? Probably because that's exactly what's friendly to gamer cards aswell as pro cards. :D If they were serious about these tests, a low-res test would be 3.2GB, a medium-res test would be 10GB, and a high-res test would be 100GB. I talked to an oil company guy at a conference in 1998 who told me their complete database for an entire oilrig was already >1TB. It's become worse over the years, but today SPEC really only shows how fast SPEC runs, not real-world data. SGI was designing tech to handle volume sets in the tens of GB more than a decade ago, but most likely the customers who do that sort of thing today are using custom NV tech which isn't public (someone at Lockheed I knew told me they did all sorts of meddling to build custom SGI setups). I highly doubt the DoD is using a commercial Quadro for defense imaging, far more likely to be a custom multi-chip version with a lot more RAM, something that could properly replace the Group Station SGI used to provide to the DoD in the early 2000s (load/display a 67GB 2D sat image in less than 2 seconds on a 64-CPU system using IR2): <a href="http://www.sgidepot.co.uk/onyx2/groupstation.pdf" rel="nofollow noopener" target="_blank">http://www.sgidepot.co.uk/onyx2/groupstation.pdf</a> Such images btw can be typically 100K pixels across at high bit depth (max res within the imaging libs and ELT back then was 115K across IIRC; grud knows what modern tech & libs can handle). They probably have a multi-GPU Quadro with 1TB VRAM or something. Note that data for newer products (Onyx3/4, IR3/4) was not released after 9/11, and data for the followon UV series paired with modern Quadros was never released either. Btw, I know a carpet manufacturer in Turkey which has to deal with large 2D images, namely pattern data for enormous carpets many metres wide. I helped them upgrade to more powerful SGI systems some years ago to cope with more complex orders, a typical file being several GB. Completely different end task, but the same high I/O demands. Ian.

Ian Mapleson

Ok cool. Yes, I noticed that modern gaming GPUs are becoming quite good at CAD software. Or I guess its more accurate to say that CAD software is getting quite good on gaming GPUs. I ran W7000/R9 390 hybrid until about 2 weeks ago, when I bought a 1070 to replace both. I made the call after seeing this: <a href="http://pcfoo.com/2016/07/gtx-1070-as-a-cg-workstation-gpu/" rel="nofollow noopener" target="_blank">http://pcfoo.com/2016/07/gtx-1070-as-a-cg-workstation-gpu/</a>#more-3776 Though there are a few apps that still bomb on gaming GPUs. Cool links. I'll bookmark em for future reference.

David

Thanks for the Win7 key offer! As mentioned though I've already installed Win7/Pro/64bit. Re posting stuff, it's actually not so bad, but the trick is to use the right packaging. :D A friend in CA recently sent me a couple of RAID PCIX cards, very low weight of course, but he used the wrong type of box which resulted in a crazy high $66 shipping cost (oops!); I'm in the process of building an X79/3930K PC for him, so I'm glad I secretly upgraded the spec from a 500GB 850 EVO to a 512GB Pro for free. :D For USPS, be careful not to use one of their premade boxes which enforces the use of an expensive shipping option like Priority (I mention this just as general info, not re whether you might want to send anything to me). Sent via economy mail, it's pretty cheap to send most things, though in the context of my build for LE you're right that beyond a certain point, often the shipping cost might be kinda high compared to the value of the item, or how much it might cost to obtain something 2nd hand on eBay or something (depends on the item). However, a US company has sent me a PCIe storage unit to sell which should help out a fair bit. Btw, I have a posting/packing advice page here: <a href="http://www.sgidepot.co.uk/postingadvice.html" rel="nofollow noopener" target="_blank">http://www.sgidepot.co.uk/postingadvice.html</a> It focuses on UK services of course, but the general concepts for sensible packaging are universal, ditto info on how to obtain free packaging materials, etc. Ian.

Ian Mapleson

Blender does not need a pro card at all. Like many modern apps, it's writen for D3D, not OGL. Viewport performance simply works better with a gamer card because that's how it's been coded. Years ago, most pro apps did run much faster on pro cards, as my Viewperf data shows (check the data for Viewperf 7 to Viewperf 11): <a href="http://www.sgidepot.co.uk/misc/viewperf.txt" rel="nofollow noopener" target="_blank">http://www.sgidepot.co.uk/misc/viewperf.txt</a> but in recent years this has changed, companies have been rewriting their code (and AMD/NV have changed their drivers) to be far more favourable to D3D and consumer cards. This is why the behaviour of Maya has completely reversed with Viewperf 12. Some pro apps do still favour pro cards, for various reasons, but Blender is not one of them. Also, pro cards are nowhere near as good as consumer cards for providing cost effective GPU accelerated rendering, something for whch cards like the GTX 580 are exceptionally good. I should mention I own dozens of GPUs, including numerous pro cards, I've been testing these issues for a long time. Here's my GPU shelf which shows mayb 2/3rds of my GPU collection (the rest are on test boards or in test systems): <a href="http://www.sgidepot.co.uk/misc/gpushelves16.jpg" rel="nofollow noopener" target="_blank">http://www.sgidepot.co.uk/misc/gpushelves16.jpg</a> Sometimes it makes sense to combine pro and gamer cards, though this isn't something AMD/NV would care to publicise too much, for obvious reasons. Pro cards have better colour precision and other features, gamer cards are great for extra GPU rendering oomph. Thus, for example, my video editing system has a Quadro 6000 for the primary display card and two GTX 580 3GB for CUDA (which I'll probably replace soon with a K5000 + two 780 Tis). I've already installed Win7/Pro/64bit btw, but thanks for the offer. :) I bought a whole stack of unused OEM Dell lics way back, plenty going cheap these days. Oh, more performance data on my site here: <a href="http://www.sgidepot.co.uk/sgi.html" rel="nofollow noopener" target="_blank">http://www.sgidepot.co.uk/sgi.html</a>#PC Ian.

Ian Mapleson

I guess more to Ian but sure, you can answer it too. Would a windows 7 ultimate key be of use? I'm across the pond so the cost of shipping kills the idea of sending any hardware.

David

Why Geforce and not quadro/firepro? Is this going to be used for CAD at all? What OS? I have a lifetime key for Malwarebytes and a Windows 7 Ultimate key I'm pretty sure isn't being used.

David

Is this question to me ? We use both Windows and Ubuntu.

Lesics

David

My page is up for the PC build for the LE guys! See: <a href="http://www.sgidepot.co.uk/misc/charitypc1.html" rel="nofollow noopener" target="_blank">http://www.sgidepot.co.uk/misc/charitypc1.html</a> All assistance is welcome, whether via straight donations, parts to use, or anything I can sell to help fund the build. Any questions, etc. feel free to PM, call, email, etc. Ian.

Ian Mapleson

Getting there slowly! 8)

Ian Mapleson

Thanks Ian for your great support. I hope your efforts will soon be fulfilled.

Lesics

I'm building the LE guys a better PC to help out their channel. If anyone wants to assist with parts, funds, or items I can sell to raise funds, please let me know. I have the case, mbd, CPU/cooler/fans, PSU; currently sorting out storage, GPU and some other things. I've been posting related adverts on forums.nekochan.net in the for-sale section. Will add an info page to my SGI site when I can. Oh, I'm in the Patron list at the end of the vid, wearing the eBid T-shirt.

Ian Mapleson