Arm Announces Cortex-A78, Cortex-X Custom

[stiiixy]

That should be pretty easy considering it's the first, beating nothing by 30% is no goal at all. I'm sure they'll succeed in the design with such goals.

The First Annual Stupid Statement party tonight not at my place!

It is a nonsensical statement, I agree.

[xception]

The first CXC processor is the Cortex-X1 that aims to deliver up to 30% better peak performance over previous Cortex-X CPUs.

That should be pretty easy considering it's the first, beating nothing by 30% is no goal at all. I'm sure they'll succeed in the design with such goals.

[_Alex_]

Anything that a for-profit company ever does is an improvement of some sort. Performance, efficiency or cost. Always; otherwise they wouldn't be doing it.

Obviously, but imagine this scenario: You launch the exact same cpu every 3 months and you alternate between 32kb and 64kb l1 cache.

Q1 2020: We've got the cache to 32kb from 64kb cache and improved our product (efficiency)
Q2 2020: We've got the cache to 64kb from 32kb cache and improved our product (performance)
Q3 2020: We've got the cache to 32kb from 64kb cache and improved our product (efficiency)
Q4 2020: We've got the cache to 64kb from 32kb cache and improved our product (performance)
"2020 has been a year of continuous improvements in terms of both efficiency and performance for our product!!!"

[intelfx]

I'm just mentioning the PR rationale behind these moves. I mean whatever they do, they always tout it as an improvement. Whether the cache goes up or down, well... you guessed it: It's always an improvement

Anything that a for-profit company ever does is an improvement of some sort. Performance, efficiency or cost. Always; otherwise they wouldn't be doing it.

[Jabberwocky]

The only demand for RISC-V in the visible future is in closed proprietary embedded devices, if at all. As much as I'd like to see a completely open RISC-V commodity board in a standard PC form factor (e.g. ATX) I think this is a pipe dream and it will never ever materialize. A laptop based on the same is even further out.

The potential is there, but the market demand isn't. Maybe a one-off amateur grade kickstarter or something like that. The demand is lower than commodity Power boards (aka Talos & AmigaOne), so I would expect the per-board cost to come in quite a bit higher. Would you pay $2750 for a totally open commodity RISC-V board? Yeah, me neither. Turns out its actually quite expensive to design and build a motherboard, regardless of which CPU type goes into it.

Sad but true.

It's awesome for educational purposes at the moment though. I've been looking at https://github.com/black-parrot/black-parrot recently. I would love to see something like a RISC-V based Argon Wi-Fi Development Board https://store.particle.io/products/argon . Baby steps I guess xD

[PerformanceExpert]

ARM now has both efficiency (A78) and performance (X1) flagships. The transistor count increases, but the overall size decreases (for A78 compared to A77 by 15%) while still improving performance because of the 5nm manufacturing process.

On the same process, Cortex-A78 is 7% faster, uses 4% less power and is 5% smaller. So on the same process perf/Watt is ~11% better and perf/area is ~12% better. So with enough effort you can improve all 3 at the same time.

I'd be very sceptical about a performance projection based on PR materials. When X1 comes out in a real product it might have to compete with whatever is after Tiger Lake for Intel, AMD Zen 3 and the next Apple core. Architectural reveals are always exciting, but it takes a while for real benchmarks of the final products

The AnandTech article mentions explicitly that previous performance projections turned out accurate. It should be next year for the first products indeed, and yes it will be very interesting I am sure!

[PerformanceExpert]

Flawed example. You can only replace the branch predictor with a larger one that uses the same amount of power if you give up die space. So it's still either/or.
Or you can do it by using a smaller node, but then you're giving up using the same predictor and saving up die space. Still either/or.

No, it's not either/or. A predictor might use more area and can still be more efficient. If say it required 5% more area but performance improves by 6%, the perf/area efficiency is better.

The only case where you'd win across the board is if you redesign the branch predictor to be more efficient using the same number of transistors and die space (or less). And that happens, too. Just not so often as juggling other, known variables.

No, you don't need the same power or number of transistors to improve efficiency. All you need to do is improve performance by more than the area or power cost. As an example imagine halving the L1 cache in Cortex-A78 giving 2% lower performance, reduce area by 3% and power by 2%. That would be an improvement in perf/area and neutral in perf/Watt. You might be able to make up for the performance loss elsewhere since you can now add 2% extra power. CPU designers make many of these kinds of tradeoffs.

[bug77]

It's never either/or. Like with software, hardware is not 100% optimal, so there is always plenty room for improvement of every aspect. You can improve performance and get better efficiency. For example, replace the branch predictor with a larger one that uses the same amount of power. Your performance improves due to fewer branch mispredictions, and as a result your efficiency improves. Similarly improve efficiency and as a result performance improves in power constrained scenarios.

Flawed example. You can only replace the branch predictor with a larger one that uses the same amount of power if you give up die space. So it's still either/or.
Or you can do it by using a smaller node, but then you're giving up using the same predictor and saving up die space. Still either/or.

The only case where you'd win across the board is if you redesign the branch predictor to be more efficient using the same number of transistors and die space (or less). And that happens, too. Just not so often as juggling other, known variables.

[numacross]

I see they're just going bigger and bigger with these cores in terms of transistors per core. Some size growth is completely natural, but it feels like ARM is now very much aiming at continually growing the cores until they have a real desktop replacement in their hands.

ARM now has both efficiency (A78) and performance (X1) flagships. The transistor count increases, but the overall size decreases (for A78 compared to A77 by 15%) while still improving performance because of the 5nm manufacturing process.

Absolutely, the X1 will be pretty much equivalent in performance to 3950X according to AnandTech. That would allow seriously fast laptops and desktops!

I'd be very sceptical about a performance projection based on PR materials. When X1 comes out in a real product it might have to compete with whatever is after Tiger Lake for Intel, AMD Zen 3 and the next Apple core. Architectural reveals are always exciting, but it takes a while for real benchmarks of the final products

[PerformanceExpert]

I see they're just going bigger and bigger with these cores in terms of transistors per core. Some size growth is completely natural, but it feels like ARM is now very much aiming at continually growing the cores until they have a real desktop replacement in their hands.

Absolutely, the X1 will be pretty much equivalent in performance to 3950X according to AnandTech. That would allow seriously fast laptops and desktops!