These Chips Are Better Than CPUs (ASICs and FPGAs)

Learn about ASICs and FPGAs, and why they're often more powerful than regular processors.

FOLLOW US ELSEWHERE
---------------------------------------------------

[luisernestofdez]

FPGAs so great at doing everything, they dont do anything out of the box

[SirChaddington]

I work in a cyber security company that uses FPGAs 😃, we use them to monitor network traffic at rates of hundreds of GB/s.

[theairaccumulator]

The chips inside cheap toys are usually tiny micro controllers that are burned with a program from the factory.

[damonstr]

As an FPGA engineer, this video was surprisingly good.

[BenjaminWheeler]

One major benefit to using an FPGA that Linus didn't mention is that they consume FAR less power than CPUs/GPUs. Accelerating a task like a ML core on an FPGA can meet/outperform a GPU at 1/10th the power draw. Also, depending on your design, they can perform orders of magnitude faster than software equivalent algorithms.

Heterogenous systems are the future!

[dennisfahey]

ASICs are fully optimized and dedicated logic. You get exactly what you want and not a gate more. It's not changeable but the fastest performance that can be achieved. (GPU's are ASICs.) FPGA are made of arrays of programmable blocks of logic, some of which you may not need for a specific function. They are slower and wasteful but for prototyping very nice. Many vendors have macro functions in their chips as well so you can just connect up to say a multiplier and not consume logic to make your own. Note that so much has been done on FPGAs that the libraries of completed designs are really valuable. This is why Intel bought Altera and AMD Xilinx. If you are doing System on a Chip - you just bought all anyone could ever ask for.
When large ASICs are developed some teams take the course of buying or making an Asic Emulator. This is an array of FPGA's interconnected. You then take the ASIC's logic and partition it across the FPGA's. This gives you an accurate representation to run the design on to prove it functions before you spend millions building it. Note the emulator rarely runs at full speed but its a hell of a lot better than a running software simulation of a massive chip which can take a very long time. (Which is lost market opportunity.) You tend to find the tough bugs far faster. And of course the emulator can be used over and over again with fresh reloads of other unrelated efforts.

[ralfpardatscher]

"You can think of them as Lego blocks, once you put them together they stay that way but you can always take them apart and put them back together and make something completely different"
Linus never had these flat small lego blocks that were impossible to take apart after u stacked them.

[ElToroLibre]

As an FPGA Engineer it's incredibly exciting to hear FPGAs being talked about! Also, in some cases for ML applications FPGAs are being skipped over in favour of AI Accelerators (TPU, GroqChip, Cerebras, SambaNova, etc) in order to replace clusters of GPUs.

[Kelble]

As an FPGA Engineer myself specializing in VHDL, I always find it soo cool to be a topic of discussion in your videos!

[awgmax]

A good analogy would be, general CPUs are a toolbox, they can make many different products, and ASIC's are a production line, they make 1 product, but they make a lot of it, fast, and accurately.

[savagepro]

Linus: These chips are better than CPUs.
Pringles: Of course! They're made of potatoes!

[NE]

I've been an FPGA Engineer for 10 years. They really are great chips. You can pretty much do whatever you want in one, it just depends on how much time and money you want to spend 😉

[LundBrandon]

Currently making a generic cpu on an FPGA. Meaning it can be 16-, 32-, 64-, 128-, etc- bit. So I could theoretically have a 1 terabit or more cpu if the chip was big enough.

[AttilaAsztalos]

Well yes but actually no. FPGAs are simultaneously between ASICs and CPUs and also on the other side, past CPUs - yes, an FPGA might be used as simply a task-specific circuit (like an ASIC except reconfigurable to do something else), but the larger FPGAs are also complex enough to behave like a CPU, ANY CPU that fits within the resources they have; and in that sense, they are MORE general than a specific CPU. Not only can they run software like a CPU does, you get to choose WHICH CPU you want them to emulate.

There is also new types of gates to calculate AI with analog circuit. There are bunch of resistors and ADC convertors inside ASIC.

[NightMotorcyclist]

Chips that do one thing really really well... brings back the promises of PhysX dedicated cards prior to nVidia's purchase.

[jajssblue]

Great video, but would've been awesome to mention the time or frequency domains that each of these are capable of working in because that's a big difference.

[AgentOrange]

3:35 With the caveat that they're volatile. If they lose power, they lose their programming. Generally they'll be paired with a ROM chip to program them on power-up. Since an FPGA doesn't emulate, but becomes the logic circuit you've built, you can turn an FPGA into a CPU if you want. (I've designed CPUs for FPGAs) But it's surreal because it's like, I have a CPU here! *unplugs it* and now that CPU is gone!

[TomsLife]

3:30 the only thing I'd add is that FPGAs are not persistent, aka, they are volatile in that they need to be re-flashed if power is lost. Many FPGA boards will have a local non-volatile flash memory that automatically reprograms them on startup, or something more fancy like an MPSoC will use commands in the OS to reprogram

[SyazaniNazarudin]

As FPGA Engineer, I glad FPGA is getting more attention in tech media.