Skip to content
【☢】 BitHardware.top ▷ Hardware, Reviews, News, Tutorials, Help post

Cost of manufacturing 7nm CPUs, GPUs and SoCs at TSMC

TSMC’s 7nm node will soon be retired by the 5nm node, but it is the most widely used in the industry, as TSMC is the factory used by the vast majority of fabless or no-manufacturing semiconductor companies own. As is the case with AMD, currently TSMC’s largest customer for its 7nm node.

The first step is wafers

The first thing that is necessary to be able to manufacture a processor is the manufacture of the wafers where the different processors are printed together before being laser cut and extracted from the wafers themselves. Which are made from silicon ingots, which have a shape like the one you can see in the image above.

From each silicon ingot, a certain number of wafers is extracted, which correspond to the minimum order per processor. In this state, each wafer of 300 mm in diameter costs an average of 400 dollars, but we have to add a series of extra costs. To this cost we have to add other additional costs such as taxes, the cost of operating the factory, the salary of workers and especially the development of new manufacturing nodes, which are increasingly expensive to develop. So the final cost of the wafers is much higher than that initial $ 400 and can go up to several thousand dollars.

The cost of the TSMC 7nm node

TSMC 7 nm process cost

TSMC has never disclosed costs per wafer, so the source in this case is Sophie Wilson, an engineer at Acorn Computers, who in a presentation by her disclosed costs per wafer manufactured under TSMC’s 7nm node. This presentation was featured on Ian Cutress’s The Tech Potato channel.

If we look at the table we will see that the cost per wafer is € 5,859.28, which translates to a cost of US $ 1.52 per 100 million logic gates. To this must be added two problems regarding the costs of each particular processor.

Problems that raise manufacturing costs

Dark Silicon

The first problem of the new manufacturing nodes is the so-called Dark Silicon, a phenomenon in which the area that the processor has to be inactive is getting larger with each new generation, this means that when implementing a new design or carrying an older one, not only transistors have to be taken into account active, but also inactive. What causes that the different designs no longer scale as before and require larger areas than expected.

The second is the defect rate, which at the 7 nm node of TSMC is 0.09 per cmtwo, so the bigger a chip the fewer good units will come out of it. For example, a Navi 21 GPU, being a much larger chip, will not only mean fewer chips per wafer, but the number of chips without defects that will come out will be in less quantity and therefore will be much more expensive.

In the specific case of AMD, it gets more profitability from the CCDs of its Zen 2 and Zen 3 than by manufacturing AMD Radeon GPUs, and it even gets more profitability from a SoC than its GPUs due to the size of each one. So all this influences when deciding which chips are manufactured at all times.

Why doesn’t TSMC officially disclose the price of its wafers?

Wafer

Depending on when a chip is manufactured, one release date or another can be planned. It must be taken into account that there are dates of the year where consumption shoots up and having your product ready for those dates is important, so companies fight so that their chips are being printed on TSMC wafers in certain periods of the year. anus.

The gossips affirm that although TSMC has an average price to produce in its factories, what it does is a kind of bidding and gives benefits to those products that have more output and therefore for which its marketers are going to earn more money. .

In the case of AMD it is curious, because if we look at their most successful product in recent months it is not a SoC for PC, but the one for PlayStation 5, but the margins for AMD are very low. Instead Apple has huge margins, this allows them to pay more money per wafer, which helps TSMC to consolidate prices in the long term. Hence, not all markets make the leap to a new manufacturing node, but rather those with the highest profitability per chip.