How are GPU's manufactured ?

[neoglow]

Ive often wondered how GPU's are actually designed and manufactured by the various brands, so what better way than to ask on here

Take CPU's as an example: Intel and AMD. There is only two mainstream brands which design and manufacture processors for todays computers that are sold to the general public. These processors are sold under the Intel or AMD brand.

With GPU's we have the big players Nvidia and AMD but we also have a huge number of different brands that take an Nvidia or AMD card, box it up, slap a name on it and sell it.

I understand there is far more to it than just 'slapping a name on it' but what is the process journey from Nvidia/AMD's tech labs right up to the final product you hold in your hand?

For example;

Why dont Nvidia or AMD design and sell their GPU's themselves, like Intel and AMD (cpu's).

Is every - GTX 1080 as an example - GPU identical to each other? Are they all made in the same factory?

What does each brand do to their cards to make them unique?

Does each brand 'buy' the blueprints for a GPU from Nvidia or AMD and manufacture it?

So many questions lol but any info would be much appreciated

 

[andybird123]

GPU chips are made on behalf of nvidia or AMD by usually TSMC but also sometimes global foundries or Samsung. AMD and nvidia also design the basic PCB and do have some made up. Nvidia do now sell theirs directly. Other times they just manufacture the main chip and sell that to the AIB's (add in board manufacturers). Some of these cards follow the PCB design given to them by AMD or nvidia, other times they design their own pcb either to save cost or to "improve" the design to make it more desirable to people interested in overclocking (though increasingly the reference design is already powered enough that OC design boards don't actually add that much).
The aib's also attach a heatsink and fan (just as whoever nvidia contracts to make their own brand boards).
The aib's then box, transport and offer warranty on the boards as they are sent out to retailers.

 

[Tythus]

Why dont Nvidia or AMD design and sell their GPU's themselves, like Intel and AMD (cpu's).

They do such as nvidia with their founders card and amd does with it's professional cards they don't like doing it as this usually means they have to deal with returns etc and hand this off to the board partners.

Is every - GTX 1080 as an example - GPU identical to each other? Are they all made in the same factory?

Yes/no the actual making would be farmed off to a chip fab such as TMSC/GLOFO depending on their processes there can be variations between foundry if farmed out to multiple foundries. The GPU's usually have a single big die with GP104 being the die for the GTX 1080 GP102 being the GTX Titan P. Processes are never perfect so during manufacture some defects will occur sometimes this is meaningless and will only affect overclockablity other times they will have to downgrade it to a gtx 1070 etc or bin it altogether. Not all 1080s are born alike.

What does each brand do to their cards to make them unique?

Each brand buys the same gpu core but they have complete control over the PCB such as power configuration ports and cooling solution they use this to differentiate themselves from the competition.

Does each brand 'buy' the blueprints for a GPU from Nvidia or AMD and manufacture it?

Yes/no they buy the core and get the reference design not sure if they pay for that but most will change this slightly for there needs as mentioned above

So many questions lol but any info would be much appreciated

no issue

 

[neoglow]

Thats interesting, cheers guys.

Im surprised each brand or AIB can design their own PCB boards, this alone is undoubtedly a very technical challenge to achieve. Sure they have professional people creating the cooling system, changing the clock speeds and testing but I never knew they designed and potentially altered the PCB spec beyond the original from Nvidia/AMD. Interesting stuff.

Are reference cards considered the best to buy?

 

[andybird123]

Thats interesting, cheers guys.

Im surprised each brand or AIB can design their own PCB boards, this alone is undoubtedly a very technical challenge to achieve. Sure they have professional people creating the cooling system, changing the clock speeds and testing but I never knew they designed and potentially altered the PCB spec beyond the original from Nvidia/AMD. Interesting stuff.

Are reference cards considered the best to buy?

Its not that complicated from an electrical engineering point of view, its basically just capacitors, resistors, VRM's, mosfets. Even serious overclockers will take an AIB board and modify it if they need more power.

The actual GPU silicon is where the really complicated stuff happens. PCB's are pretty basic.

 

[Panos]

Thats interesting, cheers guys.

Im surprised each brand or AIB can design their own PCB boards, this alone is undoubtedly a very technical challenge to achieve. Sure they have professional people creating the cooling system, changing the clock speeds and testing but I never knew they designed and potentially altered the PCB spec beyond the original from Nvidia/AMD. Interesting stuff.

Are reference cards considered the best to buy?

There is a rule of thumb for non reference cards.
AMD : Sapphire & Powercolor will improve over the reference design with amazing coolers. Gigabyte & Asus will try to cut down costs using coolers from other GPUs and cut down PCB costs.
Asrock I do not have experience with them and MSI doesn't bother.

Nvidia : For Pascal all cards were the same, regardless the PCB design, as the cooler is only the limitation here since all of them can clock to same speeds. (2190 the 1080 and 2100 the 1080Ti)
To go over you need to plug secondary boards to the cards, short the resistors and generally hardware mod your card out of spec.
For Turing we have no idea. If the behemoth burns 400W+ as is rumoured that going to be fun to cool without water, regardless.

 

[NoNameNoNumber]

Cool thread, I've often wondered why you don't buy the PCB, then GPU and VRAM separately... Would make things more interesting
That way I wouldn't have to buy expensive air coolers

 

[Rroff]

Im surprised each brand or AIB can design their own PCB boards, this alone is undoubtedly a very technical challenge to achieve

A lot of it really isn't - its the insides of the integrated circuit that is the GPU core that is complicated. I could design 9/10th of a GPU PCB in my sleep (few bits I'm not upto speed on - not done much with video output) and I'm not particularly well versed in electrical engineering.

 

[Jeps]

Thats interesting, cheers guys.
Im surprised each brand or AIB can design their own PCB boards, this alone is undoubtedly a very technical challenge to achieve. Sure they have professional people creating the cooling system, changing the clock speeds and testing but I never knew they designed and potentially altered the PCB spec beyond the original from Nvidia/AMD. Interesting stuff.

Think of the PCB as the motherboard. If we use CPUs as an example, intel/AMD make the CPU itself but do not make motherboards or if they do only make very basic ones. Its 3rd party manufacturers that make the motherboards and you get all sorts of different specs and price brackets depending on the features included. The only difference with graphics cards is that the GPU is soldered to the board and you buy them together as a single unit.

Are reference cards considered the best to buy?

Not really. The reference cards usually have worse coolers and VRM although intels new founders edition cards seem to be bucking that trend.

 

[LePhuronn]

Are reference cards considered the best to buy?

Only if you're watercooling.

Every waterblock manufacturer who releases a block for a graphics card will always target the reference PCB first. For instance EK, Bitspower, Watercool, Aquacomputer and Phanteks all have blocks for the reference 1070/1080 (same PCB) and 1080 Ti/Titan XP (same PCB). After that, it's up to the waterblock manufacturer which custom AIB PCBs they want to target. Note though that "reference PCB" doesn't automatically mean "Founders Edition", some AIBs will still use reference PCBs underneath their custom cooler (EVGA Superclocked for example is a reference PCB under their ACX 3 cooler).

So not only do you get better guaranteed block support for reference PCBs, you're probably also not paying a hefty premium for an air cooler you're not going to use (like the big Strix and Aorus coolers).

But if you're not going under water, reference coolers and cooling performance are usually pretty poor compared to the beefy custom stuff, as Jeps said above.

 

[Freddie1980]

Its not that complicated from an electrical engineering point of view, its basically just capacitors, resistors, VRM's, mosfets. Even serious overclockers will take an AIB board and modify it if they need more power.

The actual GPU silicon is where the really complicated stuff happens. PCB's are pretty basic.

I would argue it's the software and drivers that are complicated bit when it comes to GPU's as it can dramatically effect performance.

I was looking a roundup of various SBC's on YouTube the other day and SBC's like the odroid XU which has a powerful ARM SOC got beat by boards that had half of it's throughput in the same tests. The only difference was the O/S it was running on. Now while SBC's are still a relatively new market you can expect software to lag behind but the same can be true for video cards and the latest games which often require driver updates to fix performance.

 

[andybird123]

I would argue it's the software and drivers that are complicated bit when it comes to GPU's as it can dramatically effect performance.

I was looking a roundup of various SBC's on YouTube the other day and SBC's like the odroid XU which has a powerful ARM SOC got beat by boards that had half of it's throughput in the same tests. The only difference was the O/S it was running on. Now while SBC's are still a relatively new market you can expect software to lag behind but the same can be true for video cards and the latest games which often require driver updates to fix performance.

We were talking about hardware - PCB's are simple compared with chip design. Yes drivers are also tricky but you wouldn't even have drivers without a good GPU core.