This article is more than 1 year old

EV7 AlphaServers unleashed as chip line heads into sunset

Oomph up to 2004 and that's it, folks

ComputerWire: IT Industry Intelligence

Customers of Hewlett-Packard Co's AlphaServer RISC machines got a look at the probable future machines they will use over the next few years as HP rolled out the "Marvel" line of servers,

writes Timothy Prickett-Morgan.

The Marvels, which are based on the Alpha EV7 processor, are kickers to the current "Wildfire" series of AlphaServer machines, which are based on the Alpha EV68 processors.

The Marvel servers are the last generation of Alpha machines that HP will create, and eventually HP will try to move Alpha customers using Tru64 Unix and OpenVMS to HP-UX and OpenVMS operating systems running on its Itanium-based server line. The Marvel machines will supply more than enough oomph and expansion capability for Alpha customers over the next few years, whether they run OpenVMS or Tru64 Unix.

The Marvel machines are designed around a two-way cell board that houses two EV7 processors and a Rambus main memory subsystem. The EV7 processor is based on the EV68 core, which is actually available at higher clock speeds than the new EV7 processor. HP has been able to get a chip with a lower clock speed to deliver a lot more performance, said John Bennett, group marketing manager at HP's Business Critical Systems unit, because it has designed the Marvel machines around a higher-bandwidth Rambus memory subsystem and a faster NUMA-like SMP interconnect that significantly reduces I/O and memory latencies.

Moreover, the EV7 chip pulls the electronics devoted to SMP clustering and that interconnect on to the EV7 chip, as well as bringing in a 1.75MB seven-way set associate L2 cache memory that Bennett said performs like an 8MB to 12MB L2 cache in most benchmarks. The EV7 chip also includes the Rambus main memory controller on the chip as well, which allows the Rambus RIMMs to connect directly to the processor.

The EV7 runs at 1GHz and 1.15GHz and only includes one core per chip. It was designed by the old Digital Alpha chip team in Shrewsbury, Massachusetts, which will eventually be moved over to Intel. The chip maker last year bought the Alpha team and its intellectual property as part of the Alpha sunsetting that Compaq announced prior to being acquired by HP last spring.

However, the EV7 is manufactured by IBM's Microelectronics Division. Bennett said the first samples of the EV79 chip - the last one planned for the Alpha machines - have arrived from IBM Microelectronics, and that this future and last Alpha processor will ship in AlphaServer systems in late 2004. Bennett said the EV7 chip is expected to provide about 30% more performance than the EV7, and will run at 1.3GHz to 1.5GHz.

The Marvel two-way EV7 module is the basis of a processor drawer that is used in two smaller AlphaServers, the ES47, which is shipping now, and the ES80, which will begin shipping in March. The high-end Marvel GS1280 is a 64-way machine that is comprised of eight eight-way drawers, with those drawers being comprised of four two-way EV7 cell boards connected together.

The central feature of the Marvel design is what HP calls a switch-less mesh architecture that is based on the on-chip IO7 interconnect electronics on each EV7 chip. Each IO7 port offers a 32GBps full-duplex pipe directly into the EV7 processor. HP said that the IO7's bandwidth scales linearly as processors are added to a Marvel machine, and that it is even seeing better performance improvements on large configurations than one might expect based on prior SMP servers.

Each IO7 port supports three PCI/PCI-X hoses and one AGP hose, and each hose supports multiple buses. The two-way processor drawer that is the basis of the Marvel machines supports 4GB of main memory per processor today, and in the second quarter of 2003, that memory capacity will be boosted to 8GB per processor. With the introduction of the EV79 processor sometime in 2004, main memory will be boosted again to 16GB per processor as Rambus memories are made denser.

The ES47 AlphaServer tower system is a two-way server that uses a 1GHz EV7 processor. It has five PCI-X slots on three PCI-X buses and a single AGP slot, which will make it a very fast workstation, too. The ES47 comes in a tower configuration, and supports 8GB of main memory and should support 16GB of memory by the middle of this year. The ES47 rack-mounted machine offers from two to four 1GHz EV7 processors and up to 16GB of main memory, with memory boosting to 32GB later this year as well, and 10 PCI-X slots on six buses and two AGP slots. HP says that the ES47 tower replaces the DS25 AlphaServer in the HP AlphaServer line, and that the four-way rack machine replaces the four-way ES45. The ES47 supports Tru64 Unix v5.18, OpenVMS v7.3-1, and Red Hat Linux 7.2. These Tru64 and OpenVMS releases have been shipping since last November and August, respectively, and unbeknownst to HP customers, include all the tweaks to support the Marvel NUMA architecture. OpenVMS customers will have to apply a last-minute hardware patch to get it to work on their machines.

The rack-based ES80 offers from two to eight 1GHz processors in a single system image. It supports up to 64GB of main memory at announcement, has 20 PCI-X slots on a dozen buses and four AGP slots. The machine supports up to four standard I/O adjunct I/O drawers (which have lots of slots per PCI-X bus) and one high-performance I/O drawer (which puts only one PCI-X slot on each PCI-X bus). The ES80 can be partitioned into two OpenVMS or Tru64 partitions. Technically, Red Hat Linux can be supported within Tru64 or OpenVMS partitions, said Bennett, but this option is not supported by either HP or Red Hat.

The top-end GS1280 server is shipping with between 8 and 16 1.15GHz processors immediately. A 32-way configuration is expected by the middle of 2003 (probably late in the second quarter), and HP is guaranteeing that it will ship the 64-way machine by early 2004, but Bennett said it will probably ship towards the end of 2003. The 8-way and 16-way versions of the GS1280 will support 32GB and 64GB of main memory, and will be doubled on the same schedule as the ES47 and ES80 machines. The GS1280 server offers four standard I/O drawer for the 16-way machine, and will offer eight standard or two highperformance I/O drawers by the second quarter. In the third quarter, HP will ship 16 standard and four I/O drawer, and by the fourth quarter, the machine will support 32 standard and 8 high performance I/O drawers. All GS1280 machines can be partitioned, with a partition mapping to the two-way cell board as a minimum (meaning that partitions can span multiple two-way cell boards). In mid-year, HP will divulge the details of an HPC variant of the GS1280 called the SC1280.

HP is being a bit cagey about pricing, but it can afford to be with a captive customer base. Bennett said that a base GS1280 with two processors, 2GB of main memory, and a single I/O drawer will sell for $117,000. He did say that the GS1280 with eight processors would be 60% less expensive than a GS80 with eight EV68 processors, and that a 16-way GS1280 would be 28% less expensive than the prior GS160 with 16 processors. He said that the ES80 with two processors and 2GB of memory costs $81,000, which is about 20% less expensive than a two-way GS80 machine. The ES47 tower machine with two processors and 2GB of memory costs $39,700 and a two-way rack machine (which can expand to four processors) costs $64,400 with two processors and 2GB of memory.

In terms of performance, customers running ERP workloads under Tru64 should expect performance gains of between 35% and 50% on the Marvel machines with 8 or 16 processors compared to 8-way or 16-way EV68 machines running at 1.24GHz and using the same Tru64 applications. SAP is the gauge for that 35%, and stripped down OLTP is the gauge for that 50% boost. (No, HP is not going to release TPC-C benchmarks on the machine, but has done tests so it can demonstrate performance to customers.)

Because of significant tuning that HP has done with OpenVMS to support NUMA and get scalability of OpenVMS beyond the 8- to 12-processor limit than most customers saw, scalability has been greatly enhanced. On 8-processor machines, moving from Wildfire to Marvel can double performance, and on 16-way machines performance has been boosted 150% to 200%. This will make OpenVMS customers pretty happy, and that is the one thing that HP obviously wants to do over the next several years. timpm@computerwire.com

© ComputerWire

More about

TIP US OFF

Send us news


Other stories you might like