Product and Pricing Announcement
3 January 2005

New, Fast, "F Card," One-Port Myrinet/PCI-X NIC, &
Price Reductions for "D Card" and "E Card" NICs

Myricom is pleased to announce the addition of a new, fast, one-port member of its family of Myrinet-Fiber/PCI-X NICs. The M3F-PCIXF "F card," pictured to the right, is based on the Lanai-2XP chip and its associated memory operating at a 333MHz clock rate, 48% faster than the 225MHz clock rate of the popular D-card NIC. Tests of the F cards using Myrinet Express Beta software (MX-2G-Beta) for the PCIX series of Myrinet NICs show an MPI latency of 2.6µs with the Pallas benchmarks. The list price of the M3F-PCIXF-2 NIC with the standard 2MB of NIC memory is only $595.

We are also pleased to announce the latest price reductions for the other Myrinet/PCI-X NICs. These price reductions are effective for product shipments on or after 3 January 2005.

The list price of the popular M3F-PCIXD-2 "D card" Myrinet-Fiber/PCI-X NIC is now $495. That's a $100 (17%) reduction. The same tests under MX-2G-Beta as reported above for the F cards above show an MPI latency of 3.5µs for D cards. D-card production will cut over in February 2005 to a form, fit, and function-compatible version that is based on the Lanai-2XP chip, part of Myricom's streamlining production of our Myrinet/PCI-X NICs.

The list price of the high-end M3F2-PCIXE-2 "E card" dual-port Myrinet-Fiber/PCI-X NIC is now $795. That's a $200 (20%) reduction. The same tests under MX-2G-Beta as reported for the F cards above show an MPI latency of 2.7µs for E cards. Although the E and F cards are both based on the Lanai-2XP chip and its associated memory operating at 333MHz, the MPI latency with E cards is ~0.1µs higher due to the MX firmware having to manage the distribution of packet traffic across the two ports.

A word about the MX (Myrinet Express) software: Myricom-supported software has always spanned two generations of Myrinet NICs, e.g., from recent history:

• GM 2 was introduced in May 2003 along with the first of the PCIX series of NICs. However, GM 2 was also "backported" to run on the previous generation of PCI64 NICs.
• GM 1 was introduced in July 1999 along with the first of the PCI64 series of NICs. However, GM 1 was also "backported" to run on the previous generation PCI32 NICs.

MX will go into full release along with the next generation of Myrinet NICs, "Myrinet-10G" products, which we expect to announce in 2Q05. However, MX was developed on and is already "backported" to the PCIX series NICs. Thus, MX-2G-Beta versions of MX are available to Myricom customers now for evaluation simply by sending email to help@myri.com. When MX goes into full release, it will be in two versions, MX-2G for the PCIX series NICs, and MX-10G for the Myrinet-10G NICs. After the Beta-release stage, Myricom plans to charge nominal license and support fees for MX software support.

Myrinet-2000 prices and value. With the new F-card NIC, the price reductions noted above, the performance gains achieved with the new MX software, and the new Myrinet-2000 switches for large clusters, we believe that Myricom's Myrinet-2000 products offer the best performance/cost value ever. For example, the NICs, switch network, and cables for a 256-host cluster now have total list price of only $807.50 per host with D cards, or $907.50 per host with F cards. (Prices to OEMs and Myrinet cluster integrators are discounted from list prices.)

In the high-performance computing space, Myrinet-2000 is a superior alternative to Gigabit Ethernet. Here is one way to gauge the Myrinet-2000 "value proposition" relative to Gigabit Ethernet.

The November-2004 TOP500 supercomputer list includes 193 systems that use Myrinet-2000, and 176 systems that use Gigabit Ethernet. No other "generally available" interconnect was used in more than 20 of these TOP500 systems. According to the TOP500 interconnect tabulation, the Myrinet systems total 317,324 Rmax (Linpack) GigaFlops and 499,842 Rpeak GigaFlops for an average fraction of peak ("efficiency") of 63.5%. For Gigabit Ethernet the Rmax sum is 292,227 GigaFlops, the Rpeak sum is 626,176 GigaFlops, and the average fraction of peak is 46.7%. You might instead compare the best Myrinet results with the best Gigabit Ethernet results, but you would still come to the conclusion that for the Linpack benchmark Myrinet provides a ~36% performance boost over the use of Gigabit Ethernet. For computations that are more sensitive to short-message latency and host-CPU overhead than Linpack, the gains from using Myrinet are yet larger.

Although the NICs for a Gigabit Ethernet cluster may be regarded as "free," the switches are not. Indeed, the price per port for high-degree, high-bisection, Gigabit Ethernet switches may be higher than the total cost for Myrinet components for a large cluster.

If your cluster computations are at all "tightly coupled," the performance gains from Myrinet-2000 interconnect easily pay for the difference in cost, if any, between Myrinet and Gigabit Ethernet.

3 January 2005