Analysis of Tesla’s latest Model S

The new car has a 17-inch vertical screen with a resolution of 2200*1300.

There is also an 8-inch small screen in the middle of the rear row, with an estimated resolution of 960*540.

The picture above shows the GPU part of Tesla's car machine. Tesla once again ignored car regulations and used consumer-grade products. However, Tesla still takes into account the complexity of the car when it comes to the power supply circuit. Tesla uses AMD's Navi 23 GPU, and its power supply circuit is extremely complex.

LM74810 is a vehicle protection chip from Texas Instruments. It is a chip that Texas Instruments just launched in early 2020. It is mainly used in the power supply protection of electric vehicles. Tesla tries to use advanced chips as much as possible. While the industry currently uses filtering and power conditioning solutions, Tesla's can provide additional overcurrent protection and is more integrated.

The LM74810-Q1 utilizes an overvoltage adjustable feature to provide protection during unwanted load dump events. As shown, the OV pin of the LM74810-Q1 can signal an overvoltage event using a comparator. This turns off the HGATE voltage driving the Q2 metal-oxide-semiconductor field-effect transistor (MOSFET). Lower voltage rated downstream components that do not have the voltage range required for transients at the input can be used by adjusting the resistor divider connected to the OV pin to your desired threshold. The LM74810-Q1 device is rated for input voltages up to 65V and can continue to operate through transient events with peak voltages of 35V. The LM74810-Q1, along with appropriate MOSFETs and input transient voltage suppressors (TVS), protects the system from high negative voltage transients such as ISO 7637-2 test pulse 1. If the input voltage is negative, the LM74810-Q1 shuts down and pulls DGATE low. The diode at Q1 in the figure then provides reverse voltage protection to the system and prevents negative current flow. Once the input voltage returns to nominal, the LM74810-Q1 turns on again and allows the MOSFET to operate normally. The LM74810-Q1 features inrush current limiting to control the amount of current during startup. The output capacitor limits the current, ensuring that no current flows beyond the safe operating range through the component. Overcurrent protection and low voltage transient protection are also available. Pricing starts at US$1 for 1,000 pieces.

LT8638s is ADI's 40V step-down switch. It is a product in the third quarter of 2019. It is relatively new. It is characterized by extremely low noise and low EMI interference. ADI calls it Silent Switcher Monolithic Buck. Next up is Texas Instruments' LP87522, which is a 4-phase power management IC with a starting price of $2 for 1,000 pieces. The load switch is from Texas Instruments, model TPS22965, and only costs $0.30. The LDO, or linear voltage regulator, is made by ON Semiconductor, model NCV8710. It was originally used for interior ambient lighting in cars. ADS1015 is a 4-way ADC from Texas Instruments, responsible for monitoring voltage and current changes. Tesla still follows standard vehicle regulations in terms of power supply.

The memory is Samsung's K4ZAF325XM, which is expensive GDDR6 memory. It is estimated that there are 4 modules, which means a total of 8GB of memory. According to the diagram, this memory is 14 Gbps and connected to a 128-bit memory bus. This will provide a maximum bandwidth of 224 GB/s, similar to Xbox Series S.

TMP451 is a temperature sensor from Texas Instruments that monitors the temperature of Navi 23. The crystal oscillator is SiTime's SIT8924, which provides a 100MHz crystal oscillator with a wide temperature range.

There are 3 video output channels, two DisplayPort channels, one of which is High Bit Rate 2 (HBR2), supporting 4K (4096X2160) 60Hz, and the other channel is first-generation HBR, which is enough to support 2K/60P. HDMI 1.4 all the way.

Finally, let’s look at today’s protagonist Navi 23. Let’s talk about the terrifying computing power first. In the case of FP32, it can reach 10TFLOPS. This is different from the computing power of general AI accelerators. AI accelerators only do MAC multiplication and accumulation operations, which are divided into two categories: training and inference. Training is used in data centers, and inference is used in end products. Most of what we usually see are used for inference.

The FP32 data format has two uses. One is for deep learning training, with high accuracy and dynamic range. Another one is to do graphics floating point operations. The two are actually similar, both are large-scale parallel operations. Considering the cost in the reasoning field, there is no need to use such a high number. INT8 is an 8-bit integer. The default computing power values ​​​​in the reasoning field are obtained under INT8 bits. FP32 is a 32-bit floating point with decimals. The unit of integer operations is OPS, and the unit of floating point operations is Flops. The computing power of FP32 can be converted into at least 4 times that of INT8. Usually there are two different computing units in the processor for FP32 and INT8. For example, NVIDIA's Xavier has an AI computing power of 30TOPS INT8 and a GPU computing power of 1.3TFLOPS FP32. Qualcomm’s top-of-the-line Snapdragon GPU is 1.84TFLOPS. Navi 23 is a dedicated graphics chip, which is comparable to the RTX2080 of Nvidia in the United States. It has a naturally high computing power, reaching 10TFLOPS. If it is used for inference, it will have at least 40TOPS. Such strong computing power naturally has its disadvantages, which is high power consumption. The power consumption of Sony PS5 is about 210 watts, and the power consumption of the power supply is 350 watts. The foreign media Eurogamer and its affiliated company Shumao News have conducted research on PS5, PS4 Pro, 2070s, 2080s and other hardware. The power consumption was compared. The test showed that the PS5 power consumption was 1.5W in standby, 36W when idle and downloading games, and 107W when running "Rise of the Tomb Raider", with a peak value of 200W. When Tesla is used as a car machine, no matter how the power consumption is reduced, the average power consumption of the whole machine will not be less than 100 watts when playing games, and the power consumption of a single GPU part is estimated to be 80 watts or higher. It has about the same power as a desktop computer.

The strongest point of RDNA2 is the Infinity Cache, which is located next to the main computing core cluster within the AMD Navi GPU and is essentially a small and efficient memory storage. It sits between the smaller L1 and L2 caches, embedded in the GPU itself. AMD Infinity Cache acts as a massive bandwidth amplifier, achieving an effective bandwidth up to 3.25 times that of 256-bit 16 Gbps GDDR6. The solution using Infinity Cache+ 256bit GDDR6 memory has twice the bandwidth of traditional 384bit GDDR6 memory and consumes less power. Usually the capacity of L2 cache is very small, generally no more than 1MB, while Infinity Cache has 128MB.

What is the power of a typical car CPU? The Tesla Model 3 uses Intel's A3950, with a TDP power consumption of about 9.5 watts, while the A3960 used in BMW's top-of-the-line model is 12.5 watts. . Some explanations are needed here about the chip power consumption. Usually the power consumption of the chip is TDP, which is a word invented by Intel. The full English name of TDP is "Thermal Design Power", which is the thermal design power consumption, or thermal power consumption for short. As the name suggests, it is used to identify how much heat the CPU or GPU is designed to dissipate to ensure that it can operate stably at the nominal frequency. It is true that power consumption and the amount of heat that needs to be dissipated are positively related, but the two are not equal. The current and voltage fluctuate with the amount of calculation, and the power consumption P also changes accordingly, but in general the instantaneous power consumption is greater than the heat that needs to be removed by the radiator. This is because power consumption not only increases the temperature and requires the radiator to cool down, but also generates other dissipations such as thermal radiation. The power consumption of most automobile chips uses TDP standards. Generally speaking, the TDP power consumption of automobile chips is not higher than 15 watts, and is generally below 10 watts.

However, in the field of GPU, the TDP standard is not used at all. It is usually TBP, GCP or TGP. TBP, which is Typical Board Power, is the typical power consumption of the entire board. It is usually used on graphics cards. It nominally refers to the entire board (for graphics cards) It is the power consumption of GPU + video memory + other accessories), which has more reference value. GCP (Grpahics Card Power) is also called graphics card power. TGP, Total Graphics Power, is basically the same as TBP. AMD has not announced the TBP or TDP of NAVI 23, and no actual information about NAVI23 has been released. According to VIDECOCARDZ information, the TGP design target of AMD NAVI23 is 65-90 watts, and the TDP power consumption of a single NAVI23 is estimated to be around 70-80 watts.

Only an almost crazy manufacturer like Tesla would use a design that consumes 8 times more power than most car chips. No one even in emerging car manufacturing would dare to be so crazy. Not only does this design have low reliability, it also consumes more power. . Tesla may remain in the lead for 10 years.

Tesla must prepare at least three fans for this GPU, or use water cooling. Of course, Tesla may be brave because the previous generation Model S frequently blacked out and crashed because it used consumer-grade EMMC and OTA to save costs. Too many times, maxed out. But Tesla fans don't care at all.

Many people will say that Tesla designers must have considered such high power consumption.

Tesla has always been unconventional. Tesla’s idea may be to let you play games while parked. Musk claims that his car can reach Cyberpunk 2077. Not to mention the car, 99% of desktop computers cannot play this game smoothly. game. Musk wants the title of the strongest in the universe, and everything else can be slowed down.

The following table is the recommended computer configuration for Cyberpunk 2077.

The lowest level requires an Nvidia GTX1060 graphics card, and for smooth 2K screen display, you need an RTX2060 or AMD's RX 5600.

As for the CPU, we don’t know. The one that works best with NAVI 23 is AMD’s Zen2. Therefore, it is speculated that Zen2 is used. If a traditional car chip is used, it may be difficult to cooperate with NAVI23, and the development cycle will be very long. NAVI23 will have to be finalized in January 2020 at the earliest. Such a short time will be used for mass production. In the car, I am afraid that only Zen2 as a CPU can do it, and with the experience of PS5 and Xbox, it can also shorten the time. This is good for AMD and TSMC. AMD has a monopoly on the chips for the three major game consoles, and chip foundry is exclusively controlled by TSMC.

*[Copyright Statement: This article is an original article by industry experts. The author is responsible for the opinions and content compliance of the article. If you have any questions or reprint requirements, please contact the author. PS: The pictures are all from the Internet. If there is any infringement, please inform us. 】

-END-