SiFive HiFive Unmatched Hands-On, Initial RISC-V Performance Benchmarks

[coder]

It's been nearly 4 years since the HiFive Unleashed. I suspect many of us were hoping for a system that was at least "usable" by now, but clearly this isn't it.

It is usable, but I think you mean a competitive PC-alternative. However, SiFive is not competing for that market. Their cores are aimed at embedded and specialized applications, so far. This platform seems merely a developer convenience and possibly aimed at cultivating better RISC-V support for various peripherals (hence the PCIe slot).

[coder]

It's a shame, because unless SiFive and other RISC-V companies can compete in price and performance with ARM and x86 they are not going to succeed. I mean really, just looking at the dismal performance of this board is heartbreaking.

What do you mean by "succeed"? SiFive appears to be doing quite well, actually. They announced a major partnership deal for Intel to fab SoCs based on their cores in a future advanced node @ Intel's new foundry business. Rumor is that they even fended off an acquisition bid by Intel.

There's a lot of commercial business outside the market for high-performance cores. I understand if you're only interested in high-performance cores, but that doesn't mean their strategy won't be very successful for them.


BTW, did you ever look at the performance of Raspberry Pi 3 and conclude that it spelled doom for ARM? Because that's basically what you're saying. This is a dual-issue in-order core, just like the A53. I really wish Michael would've posted the specs on this SoC, instead of blindly benchmarking it as if it was meant to be comparable to a Pi v4.

From the datasheet:

SiFive FU740-C000

* Operating Frequency @ 1.2GHz
* Dual-issue in-order 64-bit execution pipeline
* Quad-core 64-bit SiFive U74

* RV64GC (RV64IMAFDC)
* 32KB I-Cache / 32KB D-Cache per core

* Embedded 64-bit SiFive S7 Core

* RV64IMAC
* 16KB I-Cache / 8KB DTIM

* 2 MB Coherent Banked L2-Cache

[coder]

What in the world. I hope you are kidding, because these look like scam prices.

These are very low-production products. What companies charge for them is usually to cover their costs in designing these boards and providing the SDKs and support for them. They're not profit-centers. By comparing this with high-volume products, you're missing the point.

And the reason they can charge so much is that anyone building a product around these chips is employing a team of software and hardware engineers which costs them much more than the price of these development boards. Compared to all the costs involved in bringing a hardware product to market, the cost of the development kits is peanuts.

It's as if I were to sell water, but I call it enterprise-class water and sell one small bottle for $1000.

If you were selling enterprise computing products, you'd quickly learn why they command such a price premium. But that's a whole different topic. These boards aren't enterprise products.

[SavageX]

I browsed a bit through openbenchmarking.org to assemble a more complete performance picture in comparison to similar ARM designs.

I only could find HiFive Unmatched results using Ubuntu 21.04, so these scores are lower than what Ubuntu 21.10 achieves.









Long story short: That board performs about as can be expected from a dual-issue, in-order design - but be aware that these numbers are from wildly different SoCs and wildly different platforms.

[kiffmet]

Isn't it getting time for someone to try and implement a RISCV core with performance levels comparable to current x86 CPUs or Apple's M1 to prove that the thing can actually scale? In my personal opinion RISCV is way too overhyped in relation to what it currently delivers and progress seems to be made far too slowly.

[SavageX]

Isn't it getting time for someone to try and implement a RISCV core with performance levels comparable to current x86 CPUs or Apple's M1 to prove that the thing can actually scale? In my personal opinion RISCV is way too overhyped in relation to what it currently delivers and progress seems to be made far too slowly.

There's work on "wide", out-of-order, high-performance cores e.g. at SiFive or Esperanto. However, leading-edge performance designs are extremely expensive in both design and fabrication. If you target M1-class hardware, you need to invest M1-class money. For Apple, this is no problem: They know they'll sell millions and millions of devices.

If you were to offer "M1-class" RISC-V hardware right now, brace for hard financial impact, as the software ecosystem and thus demand is currently not yet there. There just isn't a customer yet that'd gobble up millions of such chips.

RISC-V is taking another route. First enter the microcontroller market. This is where we are now, you can buy those and it's relatively easy for developers to target RISC-V devices, as long as you have a RISC-V compiler at hand.

Second, target more complex embedded applications that need a "proper" operating system found e.g. in routers, custom accelerators or IoT control systems. This is what cores such as SiFive's U74 are aimed at - "proper" Linux computers, slotting in wherever e.g. an ARM A53 (or similar) could be considered. This phase will yield stable programming tools, some platform standardization, adapted operating systems and a first batch of properly ported applications. Once this is done, RISC-V is ready to handle "behind the scenes" computing tasks and a demand for actual high-performance designs might emerge from that.

Once that is done, the question if RISC-V will enter the high-performance server or client market will depend on whether some market opportunity arises and a venturing party follows that path.

[coder]

Isn't it getting time for someone to try and implement a RISCV core with performance levels comparable to current x86 CPUs or Apple's M1 to prove that the thing can actually scale? In my personal opinion RISCV is way too overhyped in relation to what it currently delivers and progress seems to be made far too slowly.

Basically, what SavageX just said.

However, I'd like to point out that RISC-V is basically (de facto) following the same playbook that ARM did, over the past 2 decades. 10+ years ago, I don't recall people dismissing ARM because they didn't have a core that could challenge x86. ARM was busy carving out a nice niche for itself in low-power and embedded applications. They only made a real server push starting about 5 years ago. And yet, their main market is still low-power and embedded.

Furthermore, ARM still doesn't have a truly comparable core to Intel or AMD. It's only when you take into account power or area that ARM becomes competitive. Running 1 thread/core, Intel and AMD cores are still faster than anything ARM, themselves make. Apple stands alone in proving what you can achieve with the ARM ISA.

BTW, when either the ARM or RISC-V server markets become large enough, don't expect Intel and AMD to sit this one out. They both have everything it takes to build world-class implementations of either ISA. However, they're not going to jump before x86 starts to enter a death spiral, in that market. Moving too soon would instill doubt among their customers in their commitment to their x86 product lines and basically force them into a still immature market that can't yet replace their x86 server revenues.

[stalkerg]

I did not consider it polite to ask on mailing lists for funding. From all LinkedIn contacts I found only Bridgman responded. From SiFive nobody reacted. If there are means within AMD to sponsor some hours of developers time I'd love to finish this patchwork sooner than later ;-)

it's not working like this, Open Source in 90% time working differently. After finishing your work and after merging all patches I believe you can expect to talk about work in AMD and a good chance.

PS I also saw a few good positions in Tokyo that are bound with OpenSource drivers but for business customers. Probably I should try...

[PerformanceExpert]

Long story short: That board performs about as can be expected from a dual-issue, in-order design - but be aware that these numbers are from wildly different SoCs and wildly different platforms.

The U74 performs worse than the PI 3B, which uses a 1.2GHz Cortex-A53. SiFive likes to claim that U74 is equivalent to a Cortex-A55, but it clearly falls far short of that, despite having a lot more cache and modern DRAM. Also note many (perhaps most) of the boards you listed use a 32-bit OS and software. Using 64-bit AArch64 gives a significant speedup.

So all in all, it's not a good showing, and the absence of any OoO RISC-V cores (in actual silicon rather than announced) is telling.

[PerformanceExpert]

However, I'd like to point out that RISC-V is basically (de facto) following the same playbook that ARM did, over the past 2 decades. 10+ years ago, I don't recall people dismissing ARM because they didn't have a core that could challenge x86. ARM was busy carving out a nice niche for itself in low-power and embedded applications. They only made a real server push starting about 5 years ago. And yet, their main market is still low-power and embedded.

RISC-V may be following the same playbook, but remember it took Arm almost 30 years to get into servers and high-end cores. And that was despite having most of the embedded market and getting 10s of billions in licensing fees and royalties. I think it will be hard to emulate - high-end OoO designs are incredibly hard and expensive.

Furthermore, ARM still doesn't have a truly comparable core to Intel or AMD. It's only when you take into account power or area that ARM becomes competitive. Running 1 thread/core, Intel and AMD cores are still faster than anything ARM, themselves make. Apple stands alone in proving what you can achieve with the ARM ISA..

That's not true, Arm's server cores are already equivalent to x86, even single threaded. Nobody but Apple has made a high-end laptop core so far, but a high clocked Cortex-X should be competitive.