I had already written a very similar article almost a year ago about the binning of the GeForce RTX 4070 and the difference between MSRP and OC cards. However, the theoretical part on the second page was somewhat neglected, very wrongly in my opinion. The launch of the super cards is coming up soon and the question is always being asked as to how NVIDIA will manage to use the energy supplied much more efficiently. Of course, I can’t and don’t want to publish a white paper here or completely overwhelm the reader with technical frippery, but if you break everything down to a normal level, I think it’s actually quite exciting to read.
Binning as pre-selection and the role of “Speedo”
Of course, I have to take this opportunity to explain the most important terms to you. First of all, however, let’s stay fairly general. Binning refers to the sorting of GPUs in a model series into different quality classes (“buckets”) within a certain from-to range. For example, we know bin 0 (worse) and bin 1 (better). The individual GPUs from such a “bucket” can therefore also deviate from each other and from each other, but only within the specified tolerances (so-called “GPU lottery” at the customer). What was previously recognized as partially defective ends up one weight class lower in a slimmed-down GPU with the same initial chips and is also sorted into buckets again. However, the unusable remainder is then waste.
When you buy a graphics card, it may be a product with different advertised clock speeds, but they all have the same design and the same manufacturing process. However, the GPUs are separated from each other due to process variation (quality variations). However, they all have something in common: a maximum performance specification called TGP (also known as Total GPU Power). This also allows me to spoil the “virtual binning”, which I will discuss in a moment.
After the buckets comes the next step up. Here, the first pre-selection is generally used to put together various other (refined) “buckets” with almost identical clock rates. As is well known, the AIC and the end customer always attach the greatest importance to the ratio of performance to required energy, so that NVIDIA offers the customers exactly those “buckets” that only include chips that also offer the same performance with a similar TGP. This can be more economical or thirstier, but within the bucket, all GPUs are at least more or less equally fast. The only difference between the “buckets” is the different power consumption.
Depending on the quality level, these buckets can then be operated at very different voltages. The power specification is the maximum virtual performance of such a bucket. In the end, the actual implementation consists solely of operating slower chips at a higher voltage and faster chips at a lower voltage in order to achieve the best performance for a given performance environment. The clock speed of the chip (slow vs. fast) is controlled by a variable called “Speedo”, which is burned into the GPU during the so-called ATE flow as part of the FT (Final Test) phase.
Right now I have to explain a few things. ATE stands for “Automated Test Equipment”. These are systems that automatically test electronic components or printed circuit boards for functionality and quality (to increase production efficiency and reduce the error rate). The “Final Test Phase” is the last test phase in a production or development process, where all previously identified errors are rectified and the product is tested for functionality, performance and quality.
“Speedo” is a term used in the semiconductor industry for circuits that monitor the process, voltage and temperature (PVT) variations in integrated circuits. Such specialized Speedo circuits allow GPUs to dynamically adjust their performance and power consumption by taking into account the chips’ response to changes in process, voltage and temperature. These adjustments can be achieved, for example, by changing the clock frequency and supply voltage. I will come to this on the next page.
It is important to note that the term “Speedo” is not standardized and can have different meanings in different companies. However, a general explanation can be given as to how this term is usually used:
- Measurement of performance and quality:
Speedo systems are used to measure the performance and quality of semiconductor devices. This usually involves evaluating the speed at which a chip can function and determining its maximum performance. - Process variations:
In semiconductor manufacturing, there are natural variations in the manufacturing process. These variations can affect the performance of individual chips. A Speedo system helps to identify and quantify these variations. - Binning process:
After chips are manufactured and tested, they are often classified through a process called “binning”. Here, chips are categorized according to their performance and other parameters. Speedo measurements are an essential part of this process as they determine how well a chip performs and which performance category it falls into. - Optimization and design feedback:
Speedo data can also be used to optimize the manufacturing process and improve future chip designs. By understanding how different design decisions affect chip performance, engineers can develop better designs for future generations. - Temperature and voltage dependence:
In addition to measuring performance, Speedo systems often take into account how a chip’s performance depends on factors such as temperature and supply voltage. This is important for determining the operating conditions and reliability of the chip.
Speedo systems are therefore an important and even indispensable part of the semiconductor industry, as they enable a detailed and precise evaluation of the performance of semiconductor devices, which is essential for quality assurance and product categorization. We’ll see just how important right after we turn the page.
22 Antworten
Kommentar
Lade neue Kommentare
Mitglied
1
Urgestein
Mitglied
Mitglied
Urgestein
1
Neuling
Veteran
1
Urgestein
Urgestein
Urgestein
Urgestein
Alle Kommentare lesen unter igor´sLAB Community →