SiFive Is Launching The Most Compelling RISC-V Development Board Yet

[brucehoult]

Frequency is not an independent variable. The same CPU core design decisions that result in higher IPC will often at the same time result in a lower frequency possible in a given process node, as well as higher power consumption and heat. There is no free lunch.

I said the Unleashed performs very similar to the original Pi 3, which I have. The original Pi 3 runs at 1.2 GHz and has limited dual issue. Those are simple facts. The existence of a later model that runs at 1.4 GHz is completely irrelevant to my comparison.

U84 was announced exactly one year ago. It's absolutely normal that there is no news now. It takes everyone, including ARM and their licencees, two to three years to go from announcement of a core to actual chips shipping using it.

The A72, for example was announced in February 2015. It took a couple of years for it to appear in high end smartphones, and 4.5 years to get into a cheap SBC (the Pi 4).

The same goes for the A53. It was announced in October 2012, was in SoCs such as the Snapdragon 41x and Exynos 75xx in 2015 and was in the Pi 3 in February 2016.

This is normal and expected.

[PerformanceExpert]

Frequency is not an independent variable. The same CPU core design decisions that result in higher IPC will often at the same time result in a lower frequency possible in a given process node, as well as higher power consumption and heat. There is no free lunch.

There is no reason to believe A53/A55's higher IPC means they must run at a lower frequency, eg. Cortex-A53 ran at 2.2GHz many years ago on 28nm (eg. Helio P15). The Pi (and many similar boards) simply use the cheapest parts available, not the best or the fastest. And that's fine.

I said the Unleashed performs very similar to the original Pi 3, which I have. The original Pi 3 runs at 1.2 GHz and has limited dual issue. Those are simple facts. The existence of a later model that runs at 1.4 GHz is completely irrelevant to my comparison.

Sure but my link shows that despite having a big frequency advantage, running 64-bit code, having faster memory and larger caches, the Unleashed is still not close to that old Pi 3. My point is that an old A53 is significantly more advanced and claims by Sifive that U540 is equivalent to A53 are simply marketing BS.

U84 was announced exactly one year ago. It's absolutely normal that there is no news now. It takes everyone, including ARM and their licencees, two to three years to go from announcement of a core to actual chips shipping using it.

The A72, for example was announced in February 2015. It took a couple of years for it to appear in high end smartphones, and 4.5 years to get into a cheap SBC (the Pi 4).

The same goes for the A53. It was announced in October 2012, was in SoCs such as the Snapdragon 41x and Exynos 75xx in 2015 and was in the Pi 3 in February 2016.

This is normal and expected.

No, it takes 1 year between Arm announcing a new core to actual products on the market. That has been the case even for the first 64-bit cores. Claiming otherwise is trying to rewrite history.

For Cortex-A53 it took 14 months from the original announcement (October 2012) to Snapdragon 410 (December 2013). That's extremely fast given all the work required for Arm's first 64-bit core. Cortex-A72 was announced in February 2015 with first phones being sold in March 2016 (eg. Xiaomi Redmi Note 3).

The U74 was announced in October 2018, and 2 years later you can finally pre-order a devboard that you might get sometime next year. You see the difference here?

[_ONH_]

Unfortunately because of iGPU progress the market for small low power dGPUs is in that awkward "really small but not zero" spot and looks to stay there for quite a while, so it's hard to justify a lot of ongoing investment. We already make Polaris-based cards (RX 540 and RX 550 aka Lexa) that run without a power connector near laptop power levels.

That market is already at the point where supply chain challenges are dominating, so I doubt that anyone would want to license an existing design and take on that pain when it's easier to just buy the chips in volume for a good price.

I do miss the good old days when humans set prices instead of computers and older stock gradually had its price reduced to move it out rather than having scarcity-based algorithms charging $700 for what used to be a sub-$100 card in the hope that someone would buy it.

Sure, the joke is the RDNA successor seems to exist as RX 5300.
To have that PCB not populated with the external supply stuff and the firmware limited to the Radeon PRO 5300M spec could match, and should still be more powerful than a APU can be.
Sure with iGPU there might not be many users for it to justify a separate SKU.

There is just one point, if one need it to get it within a certain footprint, and without the x86 compatibility, but even then one could just take the R1000 APU.

The point for me is mostly power not that much price, I wouldn’t want some PCI 6-pin for this use case.

[bridgman]

Sure, the joke is the RDNA successor seems to exist as RX 5300. To have that PCB not populated with the external supply stuff and the firmware limited to the Radeon PRO 5300M spec could match, and should still be more powerful than a APU can be. Sure with iGPU there might not be many users for it to justify a separate SKU.

There is just one point, if one need it to get it within a certain footprint, and without the x86 compatibility, but even then one could just take the R1000 APU. The point for me is mostly power not that much price, I wouldn’t want some PCI 6-pin for this use case.

Agreed, a downclocked RX 5500 would probably make a nice no-power-connector card. I don't think there has been much interest from our AIB partners in putting it on a card though... my impression is that the "no power connector" market and the "cheap" market have a fair degree of overlap and the part that exists outside the overlap area isn't enough to be sufficiently interesting to justify designing a board yet.

I know you mentioned "mostly power not that much price" but I think that puts you in the minority... although it would be interesting to see how the low power cards would sell if their prices were not being pushed through the roof by mindless pricing algorithms. Unfortunately what we are seeing is the usual side-effect of a self-optimizing supply chain - niche products get pushed aside in favor of the products that more people want to buy.

Not sure how to interpret your next comment about "without the x86 compatibility" - the R1000 is just a small Raven/Picasso so is fully x86 compatible as far as I know.

[brucehoult]

There is no reason to believe A53/A55's higher IPC means they must run at a lower frequency, eg. Cortex-A53 ran at 2.2GHz many years ago on 28nm (eg. Helio P15). The Pi (and many similar boards) simply use the cheapest parts available, not the best or the fastest. And that's fine.



Sure but my link shows that despite having a big frequency advantage, running 64-bit code, having faster memory and larger caches, the Unleashed is still not close to that old Pi 3. My point is that an old A53 is significantly more advanced and claims by Sifive that U540 is equivalent to A53 are simply marketing BS.



No, it takes 1 year between Arm announcing a new core to actual products on the market. That has been the case even for the first 64-bit cores. Claiming otherwise is trying to rewrite history.

For Cortex-A53 it took 14 months from the original announcement (October 2012) to Snapdragon 410 (December 2013). That's extremely fast given all the work required for Arm's first 64-bit core. Cortex-A72 was announced in February 2015 with first phones being sold in March 2016 (eg. Xiaomi Redmi Note 3).

The U74 was announced in October 2018, and 2 years later you can finally pre-order a devboard that you might get sometime next year. You see the difference here?

The Snapdragon 410 was *announced* in December 2013, not *shipped*. There's a huge difference. The FU-740 likwwise has just been *announced*.

I don't know what the first phone actually in stores with a Snapdragon 410 was as it was used in more than 500 designs I think. The Moto G 3rd Gen for example shipped in July 2015. The Xiaomi Redmi 2 was released in January 2015. Feel free to find an earlier example. It's not December 2013, trust me.

[_ONH_]

Agreed, a downclocked RX 5500 would probably make a nice no-power-connector card. I don't think there has been much interest from our AIB partners in putting it on a card though... my impression is that the "no power connector" market and the "cheap" market have a fair degree of overlap and the part that exists outside the overlap area isn't enough to be sufficiently interesting to justify designing a board yet.

I know you mentioned "mostly power not that much price" but I think that puts you in the minority... although it would be interesting to see how the low power cards would sell if their prices were not being pushed through the roof by mindless pricing algorithms. Unfortunately what we are seeing is the usual side-effect of a self-optimizing supply chain - niche products get pushed aside in favor of the products that more people want to buy.

Not sure how to interpret your next comment about "without the x86 compatibility" - the R1000 is just a small Raven/Picasso so is fully x86 compatible as far as I know.

Sure, no one wants to produce niche products and the market outside off just "no power connector" and cheap is small.

As fas as I know, there could be more people than the ABI think, in the last year I heard quite a few who ended up buying a GT 1030/1050/RX 550, to have video, complaining about the uvd is antiquated. For the Price if the power budget is limited its as high as it is, with a 100 - 150 $ price point (below where the WX 3200 stands) they could sell quite well. Personally I would even consider it at the WX 3200 price point.

If one would want license to a design, it would only make sense for use in sbc, but there they could just use R/V1000 even if they didn't need x86 compatibility.

[PerformanceExpert]

The Snapdragon 410 was *announced* in December 2013, not *shipped*. There's a huge difference. The FU-740 likwwise has just been *announced*.

I don't know what the first phone actually in stores with a Snapdragon 410 was as it was used in more than 500 designs I think. The Moto G 3rd Gen for example shipped in July 2015. The Xiaomi Redmi 2 was released in January 2015. Feel free to find an earlier example. It's not December 2013, trust me.

Generally when an Arm SoC is announced, test chips and dev boards already exist for internal use and key customers. Every year there are benchmark leaks of the dev boards for the next generation Snapdragon and Exynos SoCs, often months before they are announced. Snapdragon 410 went into mass production early 2014 and various phones were for sale in Q3 2014 (eg. HTC desire 510). By that time devboards were publicly sold, eg. Arm's Juno boards. That's ridiculously fast since we're talking about a brand new ISA and first 64-bit cores.

[brucehoult]

Generally when an Arm SoC is announced, test chips and dev boards already exist for internal use and key customers. Every year there are benchmark leaks of the dev boards for the next generation Snapdragon and Exynos SoCs, often months before they are announced. Snapdragon 410 went into mass production early 2014 and various phones were for sale in Q3 2014 (eg. HTC desire 510). By that time devboards were publicly sold, eg. Arm's Juno boards. That's ridiculously fast since we're talking about a brand new ISA and first 64-bit cores.

First chip for a brand new ISA, except of course for the iPhone 5s announced on Sep 10, 2013 and shipped to millions of customers 10 days later :-)

Ok, cool, HTC desire 510 was shipped it seems in Sep 2014, so beat Xiaomi by four months. But a full year behind Apple (with of course their own SoC).

As you say "Generally when an Arm SoC is announced, test chips and dev boards already exist for internal use and key customers". Similarly SiFive has had real chips of the FU-740 since probably around July, demonstrated an actual system using it at the announcement at the Linley Processor Conference on October 29, and I'm sure there is more than only one working board and others are indeed already in the hands of key customers. That is the usual thing at announcement of a chip, as you don't want to announce something that turns out not to work. But those are test chips from a "shuttle run", not volume production chips.

[CommunityMember]

you are very confused. risc-v was not started by big players

It would seem you may not be aware of the complete history. And that is if the big players did not start/fund a foundation risc-v would still just be something out of UC Berkeley that was an interesting research project but gone no where in any real world (there are a number of such research architectures floating at many major universities). That risc-v offered a path to reduce the ARM tax (by potentially threatening a wholesale move to risc-v from ARM) was the reason that many of the big players joined up (a dollar here, a dollar there, and after you ship a few billion units it matters a lot). Risc-V was even explicitly mentioned as a threat to ARM's licensing model in some financial evaluation. And arguably the threat has worked. Since the formation of the foundation, and real development work has happened, ARM has (some would argue) substantially reduced certain types of fees, and opened up their development/extension processes. Risc-v did what it was supposed to do. With nvidia purchasing ARM, risc-v is still something to keep in your back pocket just in case nvidia decides to try to change something irt licensing all over again.

[lyamc]

I want AMD to have more low-profile and some mini-ITX form factor cards. The smallest, lowest power Navi AMD card I could get was a mini ITX Visiontek 5500XT which has a super sketchy cooler on it.