If you need to sell graphics cards now, follow that link and fill out a form for a free quote.
In this article, we’ll cover the rich GPU history, as well as how graphics processing units have advanced over time up into the recent days of rapid Nvidia data center growth. Specifically, we’ll cover the beginnings of the GPU, then the explosion of ATI (then AMD) and nvidia GPU growth. We’ll go over Nvidia history with the Nvidia GPU timeline and the AMD GPU timeline. Then we’ll discuss the advent of the modern data center GPU with the recent Nvidia data center GPU offerings and AMD data center GPU offerings. Finally we’ll go over GPU price history with a list of GPU prices over time.
Sell your graphics card fast and easily!
Early GPU History: 1949-1985
The “Geometry Processing Unit”
The very first electronics capable of processing code in real time to display graphics also happened to be what is likely the father of all modern computers: MIT’s whirlwind flight simulator developed for the US Navy. It was the first computer that processed in parallel as opposed to simple linear batch computing. The technical phrasing would be bit-parallel as opposed to single-bit. While it was not finished until 1951, by 1949 the machine could be operated as the very first interactive graphic computer game.
The second system to process graphics in a digital way may well have been the flight simulator developed by Curtis-Wright in 1954 depicted below. 3D graphics and other gpu technology ahead of its time was in fact available as early as the 60s, but it was highly secretive and exclusive to the government, university labs, aviation companies, and automotive companies.
Then James Clark at Stanford in 1980 coined the first usage of a “VLSI geometry processor for graphics” which might be the first term ever used that roughly equates to a graphics processing unit. It ran at about 4 million floating point operations per second, or FLOPS, which is much more fun to say. That equated to 900 polygons every 30th of a second. This appears to be the first graphics chip capable of massive parallelism that roughly did the basic functions of modern GPUs, though it certainly wasn’t built with the same basic architecture, and lacked a great number of capabilities.
GPU History: The Arcade Era (‘70s to 80s)
Then the first consumer applications of a graphics unit were in retro arcade games, and had very limited capabilities. They essentially just carried graphics info from the processor to the display. In those years, GPU technology over time moved very slowly.
The graphics were very primitive, if you can remember some of the original arcade games. E.G.
After those first arcade games, in 1976, RCA made its video chip, the pixie, but it only supported monotone, and in a tiny 62×128 resolution, at that.
In 1979, the first Graphic User Interface was developed by Xerox at the Palo Alto Research Center as a large collaboration project. It had windows, icons, drop down menus, and many other familiar features. Steve Jobs eventually toured their facilities,
Three years later, Namco Galaxian advanced graphics chips to support color – namely sprites with multiple colors and tilemaps in the background. That was just before 1980.
IBM made what could be called the first video card with the IBM Monochrome Display Adapter
Then IBM released their 8 color supporting card in 1983, the ISBX 270, which was innovative at the time, if fairly expensive at $1000. For reference, in today’s dollars, that would be $2633.
Sony then coined GPU in reference to their PlayStation in 1984 (though it was designed by Toshiba).
Then, what became a titan of gpu history, ATI was founded in 1985 by the Lau brothers and Kwok Yuen Ho, Hong Kong immigrants living in Canada. ATI eventually was bought out by AMD. More on that later.
In 1986 they released the first in their Wonder series of GPUs. These cards were dominant at the time because they supported many monitors and graphics standards in one card, while others did not.
In 1987, the video graphics array connector, or VGA, was released. VGA became the dominant standard in graphics.
In 1989, to remedy the lack of standardization in the computer graphics industry, the Visual Electronics Standards Association was founded by ATI and seven other companies. Today more than 300 companies are members.
In 1991, S3 Graphics introduced their S3 911. It was named after the Porsche for some reason, and was fairly dominant. S3 truly became a leader in the space after the release of their Trio line, which led the pack for some time.
The development of graphics technology was greatly supported by the release of two notable APIs. Possibly the most ubiquitous API for graphics rendering, OpenGL, was released in June 1992. Many competitors came and went, but OpenGL remains the surviving victor today.
The other, Direct3d, was released in 1996, and remains a standard in the industry today (though it’s evidently fractions of a millisecond slower than OpenGL, for what that’s worth).
The S3 Virge chipset, launched in 1995, was actually the fastest DRAM accelerator of the era on Windows. OEMs purchased the Virge in large quantities for its value and 2D performance, but it was certainly not compelling for its 3D performance.
S3 later sold off its graphics division.
The 3dfx Voodoo add-on card was wildly popular, and spurred on greater development of 3D technology, for gaming in particular.
The voodoo line continued to be a dominant player in the market until Nvidia acquired them later on.
Possibly the first formal usage of the acronym GPU was by TriTech in 1996, with their Geometry Processor Unit.
It was one of many similar projects that never quite took off, though it did have interesting features like dedicated bump mapping hardware and displacement mapping capabilities.
Microsoft licensed it from TriTech a couple years later.
Modern GPU History – the AMD Vs Nvidia wars (‘90s to today)
Funnily enough, both Nvidia and ATI got off to a rough start in the ‘90s. Nvidia’s NV1 was hindered by the release of DirectX 1.0 shortly after its release, which it wasn’t compatible with.
ATI’s Rage 1 also struggled due to DirectX 1.0 compatibility, though it was a good 2D performer.
While ATI had already seen some success years prior, Nvidia really first came to prominence with the release of the Riva in 1997, which sold a million units within only four months. Its popularity came in large part due to the fact that it was fairly, dare I say, general purpose. It supported 2D, 3D, and video acceleration, and they weren’t placeholder functions either, as may have been the case with many GPU makers.
With that said, its success was hampered by its lack of driver support.
Their reign truly began with the Riva TNT 2. The 3dfx API, Glide, was losing to DirectX and OpenGL, which began their downfall. GeForce cemented it.
In 1999, they made the term GPU widely used with “the world’s first GPU,” the GeForce 256. It wasn’t really the first GPU, of course. From the Clark era, GPU continued to be used by academics to refer to geometry processing.
The value proposition of the GeForce 256 was based on its inclusion of transform and lighting (or T&L) hardware on the graphics chip itself instead of relying on the CPU. As a result, with a fast CPU to handle T&L satisfactorily, its value prop was negligible. Its circuitry was also fairly criticized. That, in addition to the high price tag, meant that it wasn’t as popular as later cards, though it did have its niche with games like Quake.
It also actually performed worse than the 3dfx Voodoo if the cards were paired with a fast CPU (not that this was a terribly common scenario, however.)
Nonetheless, the DRAM version stoked a lot of excitement and so they crushed 3dfx into bankruptcy/acquisition.
A nice info graphic of the first years below was made by Jon Peddie in his seminal text, “the history of visual magic in computers,” where he goes into the level of detail a book can allow.
Nvidia Timeline/AMD Timeline: Their Reign Begins
Nvidia found itself in a very unique position at the turn of the millennium. While companies like 3D Labs made multi chip units designed for the workstation market such as the Glint, Nvidia continued to capitalize on the rapidly growing video game market, which gained a much larger demographic of buyers.
As a result, Nvidia found itself not only with the gaming market, but also situated to dominate the workstation / enterprise market as the gaming market propelled its earnings and left it with a massive R & D budget.
Some of the notable gaming releases they grew from were the PlayStation 3, World of Warcraft, and the Xbox.
Nvidia released its second generation GeForce in 2000, which did very well despite its slow 166 Mhz DDR, as it was still the fastest card out until ATI released their Radeon 7200.
(As a side note, Nvidia released their first integrated graphics product in 2001 with nForce. )
The Radeon 7200 featured better memory speed, a new bandwidth optimization technology called HyperZ, and the most complete bump mapping technology to date. Its impressive capabilities were showcased with the following Ark demo:
Nvidia answered with their GeForce 2 GTS which offered nearly a half again percentage improvement, and won the OpenGL gaming niche and certainly 16 bit in Direct3D. Its dominance was really only hindered by its poor memory bandwidth optimization.
Around this time, Nvidia began to capitalize on the workstation segment with the quadro, which was essentially just the GeForce 2 architecture with a greater emphasis on precision and reliability (through the use of ECC memory). By repackaging the GeForce with more bells and whistles and segmenting features only as needed between the cards, Nvidia could charge a premium for the Quadro, remain competitive with gaming card pricing, yet prevent workstations from using the cheaper GeForce.
Although the Radeon addressed memory bandwidth issues with HyperZ among other features, it still did not compare very favorably to the Voodoo 5 5500 or the GeForce 2 GTS, though it did well enough in 32 bit color and still sold reasonably well.
Nvidia continued their lead with the GeForce 3:
As you can see, they massively improved the rendering process with the newly improved architecture.
Then ATI answered with their Radeon 9700 Pro. It supported 64 and 128 bit color, DirectX 9, AGP 8X, and had impressive .15 micron chip specs.
Nvidia didn’t really have a competitor until 2004, with their GeForce 6800.
In 2006, we entered the modern era of GPUs with the 8th generation of GeForce cards, the GeForce 8800. It was wildly popular, and we started to see rendering, mapping, shading. lighting, rigging, post processing, etc that are in the same realm of quality as the cards of the last decade. For example, it could play Bethesda’s Skyrim, which is still a popular game today.
Around the same time, Nvidia became the only independent graphics chip maker still in business after the acquisition of ATI by AMD.
They developed the Tesla architecture which supported unified shaders, and did away with the fixed pipeline microarchitectures. It was used all the way until 40nm dies. This was extremely important for the transition to the general purpose GPU.
Instead of having a bunch of separate units, like vertex/pixel shaders, you had the more universal stream processors. They were more efficient in a wide range of use cases, and since they were simple, clock speeds could be ramped up.
In 2007, Nvidia releases Cuda, which allowed software developers/engineers utilize the parallel processing capabilities of their GPUs for more general purpose operations.
The General Purpose GPU (Or, the Rise of the Data Center GPU)
Today, the GPU name is an erroneous remnant of the past, as GPU technology has branched off massively from gaming over the last decade.
They are now systems on chips, or SoCs as they’re commonly called. They have all the circuity and functionalities you might expect from a range of separate components, but as one unified system. Specifically, you have a processor with a great deal of parallel processing, alongside a neural net acceleration engine, digital signal processors to translate analog image inputs and audio, a rasterizer, etc.
GPUs are used today for engineering app acceleration, physics modeling, rocketry, financial analysis and trading, medical imaging, clinical research, and machine learning, to name a few.
For example, possibly the most front-facing application, the GPU is widely employed as an AI inferencing tool in phones and vehicles.
While it isn’t really AI, GPUs are still referred to as “artificial intelligence inferencing engines,” which is really just a fancy way to say that it draws “inferences” or insights from existing data. For example, if you have Google photos or another cloud picture application, you may notice it identify other pictures with the same person in it, then group them together. This is accomplished by GPUs largely through “training” where Google might ask you “is this the same person?”
The reason GPUs lends themselves to this task is that to train a machine learning instance like that, it requires a massive amount of raw quantitative processing ability, where terabytes of images must be scanned one by one. The massive scalability of stream processors lends itself to these sorts of tasks very well.
Another example of recent GPU technology would be in 3D scanning of bodies, as with Magnetic Resonance Imaging, or MRI, which Nvidia also largely innovated.
If you’ve been following the news closely, you may have seen the “folding at home” phenomenon, where supercomputers and crowdsourced computing allowed researchers to better understand the protein mechanics of Sars-Cov-2, or Covid-19. exIT Technologies was actually one of the top contributors to that project’s processing power, and we accomplished it largely by the use of many GPUs used in parallel.
Where a project like that would have taken months in prior years, GPU technology over time has grown enough that we can glean insights like the molecular docking mechanisms of Covid-19’s spike protein in days.
In a more universal sense, general purpose GPUs greatly accelerate processes to reduce the time it takes engineers, researchers, software developers, etc., to solve problems, make new designs, and analyze massive sets of data. Hence the term, application acceleration.
Not only that, but the answers we get are more precise and reliable; when you need to reduce data to find an answer quickly, you sacrifice a great deal of accuracy and precision.
This year, Nvidia took a leap ahead of every other company in the world in terms of raw quantitative processing power with its A100 processor and second generation DGX systems.
It fits 54 billion transistors into just a few hundred millimeters. All that raw computing power can be subdivided into 7 separate GPUs which operate independently, or multi instance GPU (MIG).
The 3rd gen NVlink doubled the connectivity speed between the processors, and they built in sparsity to the hardware itself.
This, along with the 3rd gen tensor cores means that a single A100 GPU server provides 5 peat flops of output. It offered 7x better inferencing and 6x better training performance. In fact, the A100 provided the greatest leap in one generation of any Nvidia release.
In fact, in one swift stroke, Nvidia enabled a single system to do what could only be possible with a monster data center before. A single DGX SuperPOD made of A100 servers competes with the fastest supercomputers in the world.
Where in the past it took months to years to complete a massive supercomputer project, it only took Nvidia weeks to take the title for the fastest supercomputer in the world.
It will be interesting to see how the competition answers.
Nvidia GPU Timeline (Modern)
Note: we purchase all of the following GPUs. Get a free offer if you have spares you’d like to sell.
|Tesla GPUs||Release Date|
|TESLA M60||August 2015|
|TESLA P100||June 2016|
|TESLA M10||June 2016|
|TESLA P4||September 2016|
|TESLA P40||September 2016|
|TESLA P6||March 2017|
|Titan Xp||April 2017|
|TESLA V100||May 2017|
|Titan V||December 2017|
|TESLA T4||September 2018|
|TESLA V100S||Nov 2019|
|TESLA A100||May 2020|
|Quadro GPUs||Release Date|
|Quadro P5000||October 2016|
|Quadro P2000||February 2017|
|Quadro P1000||February 2017|
|Quadro P620||February 2017|
|Quadro P600||February 2017|
|Quadro P400||February 2017|
|Quadro P4000||February 2017|
|Quadro GV100||March 2018|
|Quadro RTX 8000||August 2018|
|Quadro RTX 6000||August 2018|
|RTX 5000||August 2018|
|Quadro RTX 4000||November 2018|
|Quadro P2200||June 2019|
|GeForce RTX GPUs||Release Date|
|GEFORCE RTX 2080 Ti||September 2018|
|GEFORCE RTX 2080||September 2018|
|GEFORCE RTX 2070||October 2018|
|TITAN RTX||December 2018|
|GEFORCE RTX 2060||Jan 2019|
|GEFORCE GTX 1650 Ti||February 2019|
|GEFORCE GTX 1660||March 2019|
|GEFORCE GTX 1660 Ti||April 2019|
|GEFORCE RTX 2080 SUPER||July 2019|
|GEFORCE RTX 2070 SUPER||July 2019|
|GEFORCE RTX 2060 SUPER||July 2019|
|GEFORCE GTX 1660 SUPER||October 2019|
|GEFORCE GTX 1650 SUPER||November 2019|
|GEFORCE GTX 1650||June 2020|
AMD GPU Timeline (Modern)
Note: we purchase all of the following GPUs. Get a free offer if you have spares you’d like to sell.
|Radeon Instinct GPUs||Release Date|
|Radeon Instinct MI25 Accelerator||December 2016|
|Radeon Instinct MI8 Accelerator||December 2016|
|Radeon Instinct MI6 Accelerator||December 2016|
|Radeon Instinct MI50 Accelerator (32GB)||November 2018|
|Radeon Instinct MI50 Accelerator (16GB)||November 2018|
|Radeon Pro GPUs||Release Date|
|Radeon Pro SSG||July 2016|
|Radeon Pro WX 5100||July 2016|
|Radeon Pro WX 4100||July 2016|
|Radeon Pro WX 4150||March 2017|
|Radeon Pro WX 7100||March 2017|
|Radeon Pro WX 4170||March 2017|
|Radeon Pro WX 2100||June 2017|
|Radeon Pro WX 3100||June 2017|
|Radeon Pro WX 9100||September 2017|
|Radeon Pro WX 8200||August 2018|
|Radeon Pro WX 3200||July 2019|
|Radeon Pro WX 3200 (Mobile)||July 2019|
|Radeon Pro W5700||November 2019|
|Radeon Pro W5500||February 2020|
|Radeon RX GPUs||Release Date|
|Radeon RX 540||April 2017|
|Radeon RX 580||April 2017|
|Radeon RX 550||April 2017|
|Radeon RX 570||April 2017|
|Radeon RX 560||May 2017|
|Radeon Vega Frontier Edition||June 2017|
|AMD Radeon VII||February 2019|
|Radeon RX 5700 XT||June 2019|
|Radeon RX 5500 XT||December 2019|
|Radeon RX 590||November 2019|
|Radeon RX 5600 XT||January 2020|
Used GPU Prices
Nvidia V100 Price
The Nvidia V100 costs somewhere in the $8000 range as of August 2020 for a new GPU, while a used card could cost close to that or as low as a few thousand dollars less, depending on condition.
Tesla P100 Price
The Tesla P100 GPU runs upwards of $3000 retail, though used GPUs and GPUs purchased in bulk will cost less than that figure.
Tesla P40 Price
The tesla P40 costs around $3500 used, per sold listings on eBay at time of writing 8-3-2020, though bulk orders and orders with varying condition will have different costs.
Tesla M10 Price
The Tesla M10 goes for anything from $1000-$1500 used per sold eBay units, though bulk orders or orders with worse conditions will go for less per unit.
Tesla M60 Price
The Tesla M60 goes for anything from $1000-$1500 used per sold eBay units, though bulk orders or orders with worse conditions will go for less per unit.
Quadro RTX 8000 Price
The Nvidia Quadro RTX 8000 goes for anywhere from $3000 to upwards of $5000 used, based on sold listings on eBay. Bulk sales and cards with lower condition will go for less per unit, however.
TITAN RTX Price
As of August 2020, the Titan RTX is selling for about $2000 for one unit, though less in wholesale.
RTX 2080 TI Price
The Nvidia RTX 2080 TI is currently selling used for around $1500; units bought in bulk will of course be sold for less.
Nvidia Titan V Price
The Titan V is selling used for anything from $1300 all the way up to $5000, currently. The market will likely stabilize on the lower end soon.
Nvidia Titan XP Price
The Titan XP is currently selling used for around $600 to $1500 on the upper end.
Nvidia Tesla T4 Price
The Tesla T4 is currently selling used for around $1300, though less for bulk transactions.
Quadro RTX 6000 Price
Currently the Quadro RTX 6000 is selling for around $3000 used. The price is lower for bulk sales, of course.
Tesla P4 Price
The Tesla P4 is currently selling for around $1000 to $1200 used. New units will sell for more, while bulk sales will be less per unit.
Quadro RTX 5000 Price
The Nvidia Quadro RTX 5000 sells for about $1300 to $1800 used. New units sell for more, while bulk sales are a lower price per unit.
Nvidia Quadro RTX 4000 Price
The Quadro RTX 4000 sells currently for around $600 to $1000, depending on condition. Bulk sales will be priced lower per unit, however.
Quadro GV100 Price
The Nvidia Quadro GV100 GPU sells for anywhere between $3000 and $5000, with $5000 being the upper end for a single unit in ideal condition, and $3000 being the lower end. Bulk sales of course are subject to lower pricing.
Nvidia Quadro P6000 Price
The Quadro P6000 sells for about $1500 pre-owned in ideal condition, with newer units worth slightly more, and bulk sales subject to discounted pricing.
Nvidia Quadro P5000 Price
The Nvidia Quadro P5000 is worth about $800-$900, with brand new units selling for slightly more, and units sold in bulk at a moderately lower price per unit.
Quadro P4000 Price
The Nvidia Quadro P4000 sells for roughly $500 depending on quantity and condition.
GeForce RTX 2080 Super Price
The GeForce RTX 2080 Super sells for about $600 to $800 dollars, depending on condition and quantity. Bulk sales will go for lower.
Other GPU FAQs
No. Nvidia merely depends on ASUS to meet their manufacturing needs. ASUS manufactures the GPUs that Nvidia designs, as would be designated by an ASUS Geforce, for example.
AMD has managed to succeed as a competitor against Intel, so it’s certainly not unfeasible for them to come out ahead in the GPU wars. That being said, GPUs continue to grow more heavily entwined with machine learning applications, and Nvidia’s R & D department has accomplished great things in that realm. It’s hard to say if the ML application side of things will shift how GPU hardware is developed, but Nvidia is well positioned in that sense.
No, but a massive chunk of Nvidia’s sales growth has come from the Chinese market, and subsequently Nvidia shareholders are beholden to China in a sense. If China wants something from Nvidia that wouldn’t lose them massive sales in other areas, Nvidia will likely do it.
Virtually every single major cloud company and artificial intelligence company uses Nvidia, as they have the lead in terms of enterprise grade GPU technology. Examples include Google, Amazon, Facebook, and even Tesla.
Yes, the Intel Xe GPU is set to be announced this month in August. Details are scant. Find more details here.
The cost of GPUs has increased partially because of an increased demand for DRAM, but also due to the increase in the R & D expenses it requires to fabricate new GPU systems. For example, Nvidia spent 2.8 billion dollars on R&D in the most recent fiscal year.
A GPU can die for a number of reasons. The earliest reason a GPU will die is faulty manufacturing, where components of the circuit board are loose or the integrity of the circuitboard is already compromised. You can cause a GPU to die earlier by overclocking it in excess of its tolerance levels, which puts strain on the system. The GPU can also be damaged during installation if static is allowed to build up and is transferred to the card. Finally, if the GPU you installed is incompatible with other components, that can cause damage to the GPU, though often it simply won’t work to begin with.
- Ray Tracing
“RTX” simply refers to the ray tracing abilities of the newer graphics card model. It isn’t an entirely different category of card so much as it is the successor to the previous GTX generation. There isn’t really a reason to buy the older GTX cards aside from cost if you’re looking to get a budget GPU option.
Yes, you can gradually wear out a GPU, causing its performance to decrease. The electrolytic capacitors on a board can dry out more quickly if your GPU heats up excessively during operation. Additionally, electromigration, where electrons move around the metal atoms, can create gaps in conductors, eventually causing failure. With that said, most GPUs don’t significantly degrade in performance over their typical lifespan.
Have something to add? Let us know your thoughts in the comments below!