Ampere Is Designing Their Own Arm Server CPU Cores, Coming In 2022

[coder]

microsoft doesn't want to make an arm personal computer, so nothing threatens x86 there

https://www.microsoft.com/en-us/surf...ro-x/processor

First 3GHz ARM processor on PC with 2 teraflops of graphics processing power

[coder]

Register renaming mostly works around that.

16 general-prupose registers still isn't enough. When the compiler runs out of registers, it generates spills, and the CPU can't optimize those away.

[coder]

Until ARM has at least 20% of enterprise I won't consider it a threat to x86.

Flawless strategy. Wait for the ramparts to be breached before sounding the alarm and rallying the troops, or even stocking the armories. What can go wrong?

Part of the enterprise won't event switch from Xeons to EPYCs due differences in the same ISA implementations from two different vendors

But they're migrating to the cloud, in droves. And don't look now, but the cloud is quickly approaching your threshold!

In case you can't see the slide:

• AWS Graviton2-based EC2 instances make up 14% of the installed base within AWS
• 49% of AWS EC2 instance additions in 2020 are based on AWS Graviton2

And Graviton2 only became publicly available, at the beginning of last year!

Future most likely gonna be heterogeneous. Some part of the share will be ARM (mostly in a form of in-house chips at megacorps), other x86.

x86 won't disappear overnight, but it's going to be a niche product, by the end of the decade.

[coder]

X-Gene/eMag were inherited from Applied Micro when Ampere bought it from them and are legacy. Neoverse/Altra is their first gen product to get them established in the performance market.

It's still a valid question: why did Ampere buy them and what did they do with the IP that was in development? Is that now being released in Siryn? Or did they just take the HW design teams and flush their IP down the toilet?

[coder]

x86 is RISC under the hood since decades. compatibility layer is relatively small part of the die.

This is often repeated and seems only partially true. The expert opinion seems to be that micro-ops aren't strictly RISC. However, that whole debate is basically irrelevant, because it's an implementation detail that's almost completely obscured.

The key points about x86's ISA are that:

• Being variable-length limits how many instructions can be decoded in parallel.
• It's more energy-intensive and uses more die area to decode than AArch64. Probably results in longer pipelines, too.
• Restricted in the number of general-purpose registers (16 vs. ARM's 32).
• Its stronger memory-consistency model acts as a barrier that limits reordering of certain instructions, which can reduce the effective concurrency of some code compiled for x86 vs. ARM.

These are things that put x86 at a natural disadvantage.

Also, the approach that ARM has taken to scaling up SIMD, called Scalable Vector Extensions (now a mandatory part of the ARMv9 ISA), is much better than the way x86 keeps requiring you to rewrite your code, to scale up to ever larger vector widths.

[edwaleni]

It's still a valid question: why did Ampere buy them and what did they do with the IP that was in development? Is that now being released in Siryn? Or did they just take the HW design teams and flush their IP down the toilet?

MACOM which bought Applied Micro didn't really want X-Gene. They wanted AM's network controllers. Carlyle Group who was the primary financier of Ampere needed ready access to the ARM CPU market bought X-Gene and it's fabrication contracts to give it a new start. What was X-Gene3 became the Ampere Emag.

https://www.anandtech.com/show/11978...e-cpu-business

As far as Ampere's roadmap Siryn is the follow up to the Altra. But it will use a custom core design and not the ARM Neoverse reference design. Again Ampere always wanted their own designs but these take time and money. They needed to keep their product pipelines active while they developed their own cores. Ampere is run by a former Intel exec and she brought a lot of senior design talent with her. Just 1 part of the ongoing Intel brain drain. So it appears that the Siryn design has been in flight for 3 years.

https://www.anandtech.com/show/16684...l-custom-cores

So it appears the legacy X-Gene and Neoverse were simply placeholders to keep Ampere going until they could finish development of their own silicon.