ARM Offers Sage Advice to RISC-V Adopters

[ali_asadzadeh]

Hi :-) I've been a member here longer than I've been at SiFive. I'm not a founder or manager or spokesperson or anything like that, just a pleb programmer. All opinions are most definitely my own. I liked the tech and its possibilities so much I joined the company.

That's great news! what sort of prices should we expect for the RISCV parts?

[janoc]

So I don't believe AMBA is restricted to ARM only.

It is not at the moment, but given the obvious hostility of ARM, why even risk it. What if next version of the specification does add restrictions specifically for RISC-V? And in the mean time you have abandoned your own bus infrastructure.

If the next version adds whatever restriction it can't be applied retroactively to the older specifications. You just may not be able to use the latest spec - which is not necessarily a problem, AMBA is an on-chip bus that isn't exposed to anything else, so if you want/need to change to something else in the future, you can without anyone else but you being affected during the development of the new product.

If anyone could change their terms retroactively and then sue you based on that, any license documents wouldn't have the worth of the paper they are written on. So in that regard you should be safe, IMO.

Anyway, I didn't want to get into the debate whether or not someone should start using AMBA, that's a cost/risk & benefit assessment that everyone needs to make themselves. I only wanted to say that there certainly are non-ARM products using the specification.

[ataradov]

So in that regard you should be safe, IMO.

Technically yes. The same way as first dose is free in drug consumption.

That why I said that once you abandon your own design fro the bus, you give up control. If next version of AMBA is not free to use, but you need to move to the bus architecture with more capabilities, you are stuck designing that bus from scratch while not having any background work done because you relied on someone else.

[janoc]

So in that regard you should be safe, IMO.

Technically yes. The same way as first dose is free in drug consumption.

That why I said that once you abandon your own design fro the bus, you give up control. If next version of AMBA is not free to use, but you need to move to the bus architecture with more capabilities, you are stuck designing that bus from scratch while not having any background work done because you relied on someone else.

And how would that be different from doing that work *now*? It is exactly the same thing. The R&D cost would need to be paid either way.

However, using a pre-existing and known to work design allows you to get the product to the market faster, cheaper and push that redesign for later when you actually are selling a product already and have an established market (and income!). Investing loads of money into R&D upfront and then running the company in the ground because you can't recover it in sales fast enough is not the best business strategy. Especially for a feature that customers don't really care about (few choose a micro or a SoC based on the internal bus architecture as long as it is good enough for the job).

[ataradov]

And how would that be different from doing that work *now*? It is exactly the same thing. The R&D cost would need to be paid either way.

Less of a shock when you already have established tools and workflow.

However, using a pre-existing and known to work design allows you to get the product to the market faster, cheaper and push that redesign for later when you actually are selling a product already and have an established market (and income!). Investing loads of money into R&D upfront and then running the company in the ground because you can't recover it in sales fast enough is not the best business strategy. Especially for a feature that customers don't really care about (few choose a micro or a SoC based on the internal bus architecture as long as it is good enough for the job).

I'm not saying that AMBA is bad in all cases. It is good in some, but it is also good to have an alternative, especially if a company feels like there are resources to do so.

[brucehoult]

So in that regard you should be safe, IMO.

Technically yes. The same way as first dose is free in drug consumption.

That why I said that once you abandon your own design fro the bus, you give up control. If next version of AMBA is not free to use, but you need to move to the bus architecture with more capabilities, you are stuck designing that bus from scratch while not having any background work done because you relied on someone else.

And how would that be different from doing that work *now*? It is exactly the same thing. The R&D cost would need to be paid either way.

However, using a pre-existing and known to work design allows you to get the product to the market faster, cheaper and push that redesign for later when you actually are selling a product already and have an established market (and income!). Investing loads of money into R&D upfront and then running the company in the ground because you can't recover it in sales fast enough is not the best business strategy. Especially for a feature that customers don't really care about (few choose a micro or a SoC based on the internal bus architecture as long as it is good enough for the job).

You don't need to do any work *now* because there already exist several unquestionably open and free to use bus specifications that are good enough.

[janoc]

You don't need to do any work *now* because there already exist several unquestionably open and free to use bus specifications that are good enough.

Even better. But I am sure someone would argue again that what if the license changes in the future or a patent suit hits or whatever. We could have such pointless debate forever.

[brucehoult]

Hi :-) I've been a member here longer than I've been at SiFive. I'm not a founder or manager or spokesperson or anything like that, just a pleb programmer. All opinions are most definitely my own. I liked the tech and its possibilities so much I joined the company.

That's great news! what sort of prices should we expect for the RISCV parts?

Impossible to answer. Anything from tenths of a cent to many tens of dollars. Which core(s) do you want? Configured how? 32 bit or 64 bit? With multiply? How fast? With divide? How fast? How much if any branch prediction? How much return address stack? How many hardware breakpoints? How many performance monitor registers? FP? MMU? How many of each core on the part? How much L1 data and instruction cache or tightly-coupled scratchpad? What about L2? What peripherals do you want from SiFive or its partners, and how many? What process do you want to use?

SiFive isn't a typical chip manufacturer. We don't come up with ten configurations that we think will sell and then build a few million of them and put them in a warehouse. The idea is to provide a chip buying experience that is like going into a web site and ordering pizza. You pick your size, thickness of the base, type of crust, toppings, sauces press a button and some time later the pizza arrives at your door.

Or in the case of SiFive, you get a bitstream and RTL based on your custom choices within a few hours.

Watch this, especially starting at around 7m40s:



SiFive *can* organise to get your chip fabbed -- especially for a sample run of 100 chips -- or even for high volume production runs, but most customers will already be making chips and just want to add our cores to their own chip. SiFive gives them the RTL and they do synthesis and something that falls out the end of that is the area and therefore the cost.

You can right away, for free, download real working (but limited) RTL for a few fixed configurations of various SiFive cores right away, put it through your workflow and get your own PPA estimates. You can download real working FPGA bitstreams of them and write and test software (at limited MHz, obviously).

https://dev.sifive.com/risc-v-core-ip/evaluate/rtl/

If you want to build a typical E31 core (like the HiFive1, but updated and improved) on 180 nm then you'd be looking at a few cents per processor. If you take an E20 and configure it with minimal scratchpad RAM then the cost per core could be fractions of a cent.

[ali_asadzadeh]

Dear brucehoult thanks for the feedback, it was very helpful, I have seen some complete chips in your website, like the E2 Series ,E5 Series and U5 Series, I was wondering if they are available to buy? and what sort of prices to expect and where to buy?

[brucehoult]

The only place you can currently buy chips by themselves at retail is the FE310-G000 on CrowdSupply for $25 for a pack of five.

https://www.crowdsupply.com/sifive/hifive1

That's a chip that taped out almost two years ago and is in the HiFive1 evaluation board. It's an early version of the E31.