This article is more than 1 year old
IBM's first 7nm Power10 chip arrives in E1080 server system with a wealth of shiny features
Boasts of a world-record SAP SD benchmark result, 'transparent' in-memory encryption
IBM's heavy-metal arm has officially brought Power10, its first 7nm chip, to market with the launch of the E1080 – a server system it claims blows x86 rivals out of the water for performance and security.
The E1080 is the first commercial outing for IBM's Power10 chips, unveiled at last year's Hot Chips conference and implementing v3.1 of the Power instruction set architecture (ISA). Built on a 7nm extreme ultraviolet (EUV) lithographic process by Samsung, the first Power10 parts include 15 physical cores – up from 12 on Power9 – and a disabled "spare" core used to increase manufacturing yield, with eight-way symmetric multiprocessing (SMT) for a total of 120 threads per chip and support for four sockets per board.
"The E1080 will actually scale to 240 cores in the entire system itself," said Dylan Boday, IBM vice president for hybrid cloud, during a press briefing. "It's really bringing in a lot of great scalability and flexibility.
"We are introducing with the E1080 a world record performance benchmark: the first system to hit 955,000 SAPS [on the SAP SD standard application benchmark] in an eight-socket system – considerably more than that of an x86 alternative architecture, 2x per socket [and] up to 4x per core more capability with the E1080 [than Intel]."
Boosted per-core performance and more cores in a system means, IBM claimed, a big reduction in footprint and power draw. In a case study looking at an unnamed customer, the company claimed that 126 Intel-based Oracle database servers had been consolidated down to just three Power9-based E980s – and was projected to drop to just two E1080s. As a result, what was 102kW of power draw is projected to drop to 20kW – and the number of licences required drops from 891 with the Intel system to 263 with the E1080.
It's not just about performance, though. IBM claimed the E1080 and its Power10 chip add a wealth of security features – including support for post-quantum cryptography, despite the NSA's uncertainty whether that's necessary.
An easier-to-sell feature: so-called "transparent" memory encryption. "What is great about this is it is encrypting information transparently without any performance overhead of the system," claimed Boday. "It's done through the hardware. And so we can actually scale this encryption to very large memory databases.
"As the information is encrypted, you [can] continue to do computational workload on it and not unencrypting it, with fully homomorphic encryption. This is all achieved through our 2.5x faster [AES cryptography] performance per core."
"Not only is there no performance impact whatsoever," claimed Satya Sharma, IBM Fellow and chief technology officer, "but there is no management setup required either. So this is what I mean by transparent memory encryption: it simply works. There is no user action required. There is no performance penalty and no management overhead."
Another key feature, and an indicator of where IBM sees the future of computation, is the integration of acceleration engines for artificial intelligence workloads: four Matrix Math Accelerators (MMAs) per core. "It provides 5x more inferencing performance than what we did in Power9," Boday claimed.
"This provides an alternative route to using separate GPUs as the in-core capabilities of the Power10, with the MMA engines we've embedded into it, allows our clients to do the computation work directly in [the] stream of data."
GPUs as AI accelerators aren't going anywhere any time soon, however. While the E1080 will happily run a workload, including those on the Open Neural Network Exchange (ONNX), the actual training is likely to take place elsewhere. "Many of the training environments do require GPUs," Sharma admitted, "But once the model is built, we are able to bring that model on Power10 and still provide high security and tight reliability."
There's a disparity in IBM's numbers on the AI front, though. At Hot Chips 2020, it boasted Power10 would offer up to 20x the inference performance of Power9 – but it's launching the E1080 stating a (somewhat more sedate) fivefold increase.
- Semiconductor veterans gather to design customizable, chiplet-based RISC-V server processors
- TSMC to hike some chip prices 'by as much as 20%'
- Taiwan and Arizona economic groups agree to bring more chip industry to desert state
- Samsung splurges $200bn to bring 'new future order' over next few years – meaning chips, pharma, comms
"When we presented Power10 at the Hot Chips event, we talked about a 20x number and the 10x number," Sharma explained by way of addressing the gap. "Those numbers are still holding. We talked about a single chip module and a dual chip module.
"The dual chip module will deliver the 20x capability, like we talked about the Hot Chips event; the single chip module, which is what we are using in E1080 in a low precision [mode], which as most of the you know quite a bit of the AI world is in the low precision mode, there we are going to deliver 10x capability. So this 5x is a high precision [mode] proof point – we are still consistent with what we had talked about at Hot Chips.
"This trend towards on-processor AI is actually a broad trend in the industry," Sharma continued. "[IBM] Z [mainframes] announced it. We, of course, covered it at the Hot Chips event when we did Power10. There's a clear trend in the market, and that's going to increase the pervasiveness of AI from the business application standpoint."
The Power10 chips aren't the only new hardware to be found in the E1080. IBM is also launching a new type of memory, designed to improve reliability by taking advantage of Power10's Open Memory Interface (OMI) architecture. "Instead of using the industry-standard DIMMs, which almost all of the x86 world uses," Sharma said, "in Power10 we are using OMI-attached... we, sort of informally, call them DDIMMS.
"These are buffered DIMMS. And we are able to isolate any DIMM failures within the buffer DIMM instance itself, so it causes fewer system outages compared to the x86 world. This is becoming extremely critical in these in-memory [database] configurations."
On the software side of things, IBM had words of reassurance for its independent software vendor (ISV) ecosystem: everything should work as before. "Power10 has a Power9 and a Power8 compatibility mode," Sharma said. "So you can essentially run a virtual machine in that mode, and then all of your software, whether it's ISV software, or customer software, or operating systems for that matter, all of it can be brought forward.
"At the same time, even in Power10 mode, we have a binary compatibility guarantee. So, the entire ISV ecosystem that we have built over the decades is going to be able to come forward to Power10."
Orders for the E1080 are open now, price-on-application, with shipments expected to begin before month's end. Interested parties can find out more on IBM's website. ®