By the beard of Zeus! Arm takes another tilt at serious servers with SVE-capable 96-core Neoverse V1
(2020/09/22)
- Reference: 1600781773
- News link: https://www.theregister.co.uk/2020/09/22/arm_v1_n2/
- Source link:
Arm is taking another serious tilt at server silicon with new designs that incorporate the Scalable Vector Extensions (SVE) beloved of HPC and machine-learning types.
The Nvidia-stalked chip designer's newest offering is the Neoverse V1 – code-named Zeus – which will offer up to 96 cores per socket, all with [1]SVE support. Arm says the V1 offers its best performance-per-thread, and is available now. Licensees decide the clock speed and process node – 7nm or 5nm, say – so YMMV once silicon goes on sale.
Arm has also [2]updated its Neoverse design to a new generation. The new 5nm N2 series can pack up to 128 cores-per-socket, and is due to land next year.
Chris Bergey, the processor architect's senior veep and general manager for infrastructure line of business, said cores-per-socket matter because IT buyers care about power consumption again.
UK govt urged to bolt tough legal protections onto Arm and protect jobs – or simply veto Nvidia's £31bn acquisition [3]READ MORE
Arm's first tilt at the server market saw it promote its lower power requirements, Bergey explained, but rivals had such a software ecosystem head start that clouds and other users stuck with what they knew.
Today, Arm reckons that hyperscalers are tired of building new bit barns or needing extra power, and want more cores per rack. Power-sipping Arm cores capable of heavier lifting are therefore attractive. 5G also makes the case for more cores compelling, Bergey argued, because running a virtualized network needs lots of compute in lots of tight spots.
[4]
Arm's roadmap. Click to enlarge
In a full rack consuming 12 or 15 kilowatts, Bergey reckoned users can score up to 40 per cent higher core count and double the raw compute power rivals can offer.
That the N2 and V1 are more capable hardly hurts.
The biz claimed the N2 improves single-core performance by up to 40 per cent compared to its predecessors, and that the V1 performs 50 per cent better than first-generation N1.
Arm hopes AWS, FOSS, cloud-native combo will wrestle you onto its architecture [5]READ MORE
Arm being Arm, it's up to licensees to put these cores to work. But Arm thinks it now has a story that will satisfy users from the data center to the edge, all on the same architecture, and that the work it has done to ensure CI/CD teams can work with its platforms means users can now contemplate coding for an all-Arm stack.
Bergey told The Register he thinks access to silicon won't be an issue – he said TSMC is on top of its 5nm process so V1 and N2 units should roll off the production line without duress.
All that's missing now is you. And as Bergey has pointed out, now that AWS can rent you an Arm-powered Graviton for pennies, it's harder to find an excuse not to at least play with Arm in the cloud. Or maybe it'll be best to wait for whatever Nvidia does with the British chip design house. ®
Get our [6]Tech Resources
[1] https://developer.arm.com/documentation/100891/0612/sve-overview/introducing-sve
[2] https://www.arm.com/company/news/2020/09/accelerating-the-next-gen-cloud-to-edge-infrastructure
[3] https://www.theregister.com/2020/09/22/uk_government_arm_nvidia_unite/
[4] https://regmedia.co.uk/2020/09/22/pr_supplied_arm_roadmap.jpg
[5] https://www.theregister.com/2020/07/22/arm_enterprise_plan/
[6] https://whitepapers.theregister.com/
The Nvidia-stalked chip designer's newest offering is the Neoverse V1 – code-named Zeus – which will offer up to 96 cores per socket, all with [1]SVE support. Arm says the V1 offers its best performance-per-thread, and is available now. Licensees decide the clock speed and process node – 7nm or 5nm, say – so YMMV once silicon goes on sale.
Arm has also [2]updated its Neoverse design to a new generation. The new 5nm N2 series can pack up to 128 cores-per-socket, and is due to land next year.
Chris Bergey, the processor architect's senior veep and general manager for infrastructure line of business, said cores-per-socket matter because IT buyers care about power consumption again.
UK govt urged to bolt tough legal protections onto Arm and protect jobs – or simply veto Nvidia's £31bn acquisition [3]READ MORE
Arm's first tilt at the server market saw it promote its lower power requirements, Bergey explained, but rivals had such a software ecosystem head start that clouds and other users stuck with what they knew.
Today, Arm reckons that hyperscalers are tired of building new bit barns or needing extra power, and want more cores per rack. Power-sipping Arm cores capable of heavier lifting are therefore attractive. 5G also makes the case for more cores compelling, Bergey argued, because running a virtualized network needs lots of compute in lots of tight spots.
[4]
Arm's roadmap. Click to enlarge
In a full rack consuming 12 or 15 kilowatts, Bergey reckoned users can score up to 40 per cent higher core count and double the raw compute power rivals can offer.
That the N2 and V1 are more capable hardly hurts.
The biz claimed the N2 improves single-core performance by up to 40 per cent compared to its predecessors, and that the V1 performs 50 per cent better than first-generation N1.
Arm hopes AWS, FOSS, cloud-native combo will wrestle you onto its architecture [5]READ MORE
Arm being Arm, it's up to licensees to put these cores to work. But Arm thinks it now has a story that will satisfy users from the data center to the edge, all on the same architecture, and that the work it has done to ensure CI/CD teams can work with its platforms means users can now contemplate coding for an all-Arm stack.
Bergey told The Register he thinks access to silicon won't be an issue – he said TSMC is on top of its 5nm process so V1 and N2 units should roll off the production line without duress.
All that's missing now is you. And as Bergey has pointed out, now that AWS can rent you an Arm-powered Graviton for pennies, it's harder to find an excuse not to at least play with Arm in the cloud. Or maybe it'll be best to wait for whatever Nvidia does with the British chip design house. ®
Get our [6]Tech Resources
[1] https://developer.arm.com/documentation/100891/0612/sve-overview/introducing-sve
[2] https://www.arm.com/company/news/2020/09/accelerating-the-next-gen-cloud-to-edge-infrastructure
[3] https://www.theregister.com/2020/09/22/uk_government_arm_nvidia_unite/
[4] https://regmedia.co.uk/2020/09/22/pr_supplied_arm_roadmap.jpg
[5] https://www.theregister.com/2020/07/22/arm_enterprise_plan/
[6] https://whitepapers.theregister.com/
Solving the wrong problem
It's not chip feature/function that will make the difference for ARM. It's having a consistent and widely adopted* systems architecture so that software can run anywhere without extensive redesign (see for example the work Fujitsu had to do to get Lustre to be effective on its newest supercomputer).
And if ARM doesn't make that happen, maybe AWS will do so by default...
*Yes, there is an architecture. No, it's not widely adopted.