Talk:Tesla (microarchitecture)

What are the advantages to using Tesla over the 8800 and Quadro lines? The article only mentions graphics capability is not included -- so why pay extra?


 * Basically the Tesla is a co-processor. It's designed to crunch a lot of numbers. The reason why you don't have the CPU do this is because the CPU is already busy with menial tasks, not to mention CPUs are slower in raw power than GPUs. You also don't want the GPU doing this because it could be doing other things and it's designed to work on graphics rather than large scale calculations. So just tweak with the BIOS or instruction set a little on a already fast performing part, and you have something like a CPU, only much faster. -- XenoL-Type 18:29 23 February, 2008 UTC —Preceding comment was added at 02:29, 24 February 2008 (UTC)
 * Wrong! Tesla is same old Geforce, that is un-crippled with drivers. Nothing else! 93.129.9.216 (talk) 16:13, 26 May 2013 (UTC)
 * Back then there was no big difference with the firtst G80 chip, but then followed additional ECC memory, increased FP64 throughput, and nowadays Transformer Engines and alike. 2A04:4540:6A1B:D00:9A58:621D:C879:77F5 (talk) 08:08, 31 March 2023 (UTC)

It's not that simple, efficient general purpose computing on a gpu requires an architecture redesign. —Preceding unsigned comment added by Vlad Dracula (talk • contribs) 00:15, 13 March 2008 (UTC)

Where is Tesla from?
This article says that Tesla was Croatian, but the article about him says he is Serbian. Which is right? — Preceding unsigned comment added by 24.207.132.139 (talk) 03:48, 19 April 2012 (UTC)


 * Tesla was born in nowadays Croatia but his father was a serbian-orthodox priest. 2A04:4540:6A1B:D00:9A58:621D:C879:77F5 (talk) 08:05, 31 March 2023 (UTC)

The name "Tesla"
Why have they used this? I assume it's named after Nicola Tesla? —Preceding unsigned comment added by 79.66.89.230 (talk) 12:42, 23 January 2009 (UTC)


 * Almost all Nvidia architectures have scientists (mathematics, physics, computer science) as internal code names, starting with Kelvin or Fourier. 2A04:4540:6A1B:D00:9A58:621D:C879:77F5 (talk) 08:03, 31 March 2023 (UTC)
 * Riva architecture starts with "Fahrenheit"...
 * https://nouveau.freedesktop.org/CodeNames.html 2A04:4540:6A17:9000:9600:C390:5F83:56F4 (talk) 18:41, 3 May 2023 (UTC)

I agree, why is this not mentioned? —Preceding unsigned

Fair use rationale for Image:NVIDIA Tesla.png
Image:NVIDIA Tesla.png is being used on this article. I notice the image page specifies that the image is being used under fair use but there is no explanation or rationale as to why its use in this Wikipedia article constitutes fair use. In addition to the boilerplate fair use template, you must also write out on the image description page a specific explanation or rationale for why using this image in each article is consistent with fair use.

Please go to the image description page and edit it to include a fair use rationale. Using one of the templates at Fair use rationale guideline is an easy way to insure that your image is in compliance with Wikipedia policy, but remember that you must complete the template. Do not simply insert a blank template on an image page.

If there is other fair use media, consider checking that you have specified the fair use rationale on the other images used on this page. Note that any fair use images uploaded after 4 May, 2006, and lacking such an explanation will be deleted one week after they have been uploaded, as described on criteria for speedy deletion. If you have any questions please ask them at the Media copyright questions page. Thank you.

BetacommandBot 04:34, 1 October 2007 (UTC)

Floating-point precision
The article needs to differentiate between single-precision and double-precision and present both when discussing performance. I suspect that the FLOPS given are for single-precision numbers. It is biased to present only one of the two. Rilak (talk) 07:00, 11 February 2009 (UTC)

I think it is single precision, as the SPs seem to be using IEEE 754 for single-precision. See http://ieeexplore.ieee.org/xpls/abs_all.jsp?arnumber=4523358 page 46, paragraph on streaming processors. Floating point arithmetic would have to be done on the SFU of which there is only one per four SPs. --155.198.14.201 (talk) 10:36, 2 May 2009 (UTC)

"Fermi-based Nvidia GeForce cards have four times less dual-precision performance." -- So does that mean that the lessness is multiplied by four to get the dual-precision performance of the Fermi-based cards or is this a convoluted way of saying it has one quarter the performance? — Preceding unsigned comment added by 76.84.50.190 (talk) 17:20, 2 January 2014 (UTC)

G200 architecture?
Can someone provide a citation for why the Tesla C1060 specs are being filled in with GTX 280 specs? I have not been able to find anything stating that the Tesla C1060 uses the G200 architecture. —Preceding unsigned comment added by 128.132.173.228 (talk) 16:07, 3 June 2009 (UTC)

C2075
There's now a C2075 but I'm not sure of its characteristics. — Preceding unsigned comment added by 68.183.134.158 (talk) 21:13, 6 December 2011 (UTC)
 * Judging by the spec sheets here: and here, the only difference I can find is that the new card has a slightly lower TDP (225W), and inexplicably only supports up to 1600x1200 out of the Dual Link DVI port.  vlad§inger  tlk  20:29, 27 December 2011 (UTC)

Table is wrong!
The table is totally wrong! The SFUs are not accessible in a peak scenario with MAD calculations and the second MUL is not even available at all from GF100 on up. A sortable table with the correct values would be more useful in the way that if you order per double precision performance, the K10 would result the slowest model of the boards.


 * Discuss and get consensus before making major changes. You appear to be vandalizing articles and aren't using edit summaries. I suggest you create an account before proceeding further, or you will just be blocked like yesterday. You're removing thousands of characters from articles without any reasonable explanation. Wikipedia is a collaborative effort, but you are apparently taking it upon yourself to make major changes in the structure and content of articles without involving the editors who created and contributed to the article in any way. That's not how we do things. See our Bold, Revert, Discuss process. You've been reverted multiple times, now it is time to discuss and wait for responses and agreement. Yworo (talk) 17:23, 18 February 2013 (UTC)

I'm sorry. I was just trying to sinthesize the table, making it sortable for each value and also correct the wrong values (that you re-edited back in, BTW). If I can't correct the page, then who will, since, apparently you didn't correct them at all? — Preceding unsigned comment added by 93.38.171.227 (talk) 09:04, 19 February 2013 (UTC)


 * You're making changes to a table of numbers without either edit summaries or sources. That's a common form of vandalism. Without sources for the "correct" figures, how do we know your aren't just vandalizing the article? We don't. That's the problem. Just provide a reliable source that shows that the values you are changing are corrections. And please consider creating an account. Yworo (talk) 20:15, 1 March 2013 (UTC)

ECC
Sorry, but how is it possible? GDDR5 has ECC from factory! How is Nvidia going to "ECC" this memory? Does anyone have any photo proofs that its different type of memory? I have strong feeling Nvidia PR is BS'ing people again. 93.129.9.216 (talk) 16:06, 26 May 2013 (UTC)

K20X GPU Computing Module memory parameters
I'm confused! I'm trying to calculate maximum theoretical throughput for the K20 and K20X GPU Computing Modules. For the K20X, the article table says the memory has a channel width of 384 bits, a clock rate 5200 MHz and bandwidth of 250 gigabytes per second. The cited reference gives the clock rate as 2600 MHz. But sure enough, (250 GB/s) x (8 bits/byte) / (384 bits/channel) = 5200 MHz. So is the nVidia cited reference wrong?

On the other hand, the table gives the K20 bandwidth at 208GB/sec over a 384 bit channel, but the same calculation says the clock should be 4333 MHz, not the 2560 given in the table.

So what's amiss? Am I making a false assumption about memory clock rates and bandwidth? Thanks to anyone who can either educate me better or fix the data! p.r.newman (talk) 15:24, 17 July 2013 (UTC)


 * Ah! Is it simply because it's  DDR memory that the clock rate is half the apparent data rate? p.r.newman (talk) 15:31, 17 July 2013 (UTC)

Confusion between microarchitecture and product line
This article confuses the Tesla microarchitecture with the "Tesla" product line from Nvidia. The former is the architecture powering the GeForce 8 through GeForce 300 series of video cards, while the latter is Nvidia's line of products for the HPC market, which are actually based on GPUs from subsequent architectures too (at least Fermi and Kepler). The intro text correctly describes the former, while the table lists products from the latter. I suggest moving the table to a new page describing the Tesla HPC product line, similarly to how the Quadro line has its own page. Marco 93.37.4.4 (talk) 20:10, 6 March 2014 (UTC)


 * YES! I came here to note that this is all a bit confusing - if there is a "Tesla Card" that has a "Kepler core" then where in that table do I look for it? The article is also getting sort-of out-of-date: it talks about "the latest C-class products" (which Nvidia lists on their web pageas "legacy products") and statements are made "as of 2012" (which is more than one computer generation ago as of Oct 2014)Iron Condor (talk) 18:04, 13 October 2014 (UTC)

manufacture Process 40 nm
according to this site, the smallest manufacturing Process was 55nm, but I'm currently using a GT218 with 40nm Technology (Asus Nvidia Geforce 210 Silent). — Preceding unsigned comment added by 217.232.242.81 (talk) 20:36, 5 August 2016 (UTC)

Scalar Streaming Processors?
Maybe the definition of "Scalar Streaming Processor" is a bit misleading, AFAIK Nvidia never mentioned SIMD as hardware architecture in their white papers but referred to Tesla as SIMT, Single Instruction Multiple Thread, IMHO this is basically a 8 lane wide SIMD8 unit (8 Cuda cores) executing a Warp of 32 threads over 4 waves as SIMT-4. 2A04:4540:6A02:6300:1919:33A0:A06F:D7E7 (talk) 11:39, 29 March 2023 (UTC)