UltraSPARC T2 Plus detailed

Sun will soon be updating it's UltraSPARC line with the T2+, a slight updated to the T2 - also known as Niagara 2.

The original UltraSPARC T2 is a processor aimed at highly threaded workloads since it is capable of up to 8 simultaneous threads running per core. It has eight cores per socket, delivering up to 64 threads per socket at a clockspeed of 1.2 or 1.4GHz. This highly threaded approach is good to hide latency to the main memory and is similar to the one Nvidia uses in their current GPUs - a "SIMT" programming model - which requires thousands of active threads per GPU to achieve a high throughput.
The T2 is highly aimed at Java workloads, especially enterprise applications, where there are frequent pipeline stalls due to memory accesses, a situation where the highly threaded model has very strong arguments.

The new UltraSPARC T2+, codenamed "Victoria Falls", features the following characteristics:

• Added cache coherence to the T2 - it's the most important change.
• Memory control units (MCU) were reduce from 4 to 2.
• The FB-DIMM pins were reused to create a high speed SERDES link between the cores.
• 51.2GB/s of available bandwidth available for the coherence links, which are four.
• The CPU supports up to 8 cores per socket, SMT provides up to 64 threads, 8 per core. The T2+ has the same one FPU per core as the T2.
• No changes at the pipeline level, which was a short one, so no performance loss or improvements from here.

Below are slides from a SUN presentation about the T2+:







There's also a more in-depth presentation available here.

AMD's Opteron 2435 ''Istanbul'' reviewed

AMD released today the first native six-core CPU, Anandtech has a look.

If you recall correctly, Intel had already released a six-core part, the Xeon 7400 series, codenamed "Dunnington". Contrary to Intel's design, "Istanbul" was built natively for this purpose and is not a bunch of dual cores glued together to an L3 cache, coupled with a low bandwidth FSB. AMD pumped the HT3.0 speed to 2.4GHz and added HT assist, aimed at reducing latency and bandwidth usage for cache snooping when using 4 sockets or more. 4 way CPU to CPU bandwidth is massively increased in this case, up 60% from around 25GB/s to 41GB/s.

The die size hasn't increased much and power consumption either, maintaining the L3 cache at 6MB helped restrict both but will hold back some performance in particular applications.
Die size is then 346 mm² compared to 263 mm² of "Shangai" cores. Xeon 55xx series currently are using 265 mm² and "Dunnington" six-core used a massive 504 mm².

Power consumption is well contained at 75W ACP or 115 TDP, the same as the 2.9GHz Opteron 2389, while the clockspeed has been just slightly reduced to 2.6GHz.

Performance is good in some applications, most notably virtualization. It's priced lower than the two top Xeons, but there's a sad lack of HPC benchmarks in the test suite for real comparisions. Most applications benefit the most efficient "Nehalem" architecture but, if priced right, AMD has a good CPU, although it's far from hiting a home run.

AMD has a good CPU in it's hands until Nehalem-EX platforms debut in 2010, and will reign in 4 and 8 socket configurations until that time comes. As for 2 sockets configurations, nothing seems capable of standing up to Intel's products right now but I'm curious to check HPC benchmarks, which typically see no benefit from Hyper Threading.

AMD Opteron 2435 review @ Anandtech

Supermicro brings Intel's Atom to server environments

Atom CPUs are now, officially, everywhere.

Super Micro Computer, Inc., a leader in application-optimized, high performance server solutions, today announced the launch of 4-watt and 8-watt Intel® Atom™ processor-based Server Building Block Solutions®. These extreme low-power and quiet solutions deliver the best expansion and storage features of any Atom-based platform and are designed for embedded industrial PC (IPC) applications.

These quiet, energy saving solutions make ideal storage and network appliances or print and email servers. In addition, Supermicro's mini-tower chassis with advanced user-friendly design features create a very cost-effective home or office workstation for basic computing including Internet usage, office and educational applications.

"Bringing the low-power consumption advantages of Atom processors to the server appliance market empowers our customers with energy-saving, quiet solutions that provide flexible expansion and storage features previously unattainable with Atom solutions," said Charles Liang, president and CEO of Supermicro.

Optimized for the single-core Atom 230 processor, which consumes only 4 watts of power, Supermicro's cost-effective X7SLA-L platform supports up to four SATA ports with RAID 0, 1, 5 and 10, along with seven USB 2.0 headers, 2 GB DDR2 memory, Intel GMA 950 graphics and a Gigabit Ethernet port. For more performance-intensive applications, the high-end X7SLA-H integrates the dual-core Atom 330 processor, which consumes 8 watts of power and expands upon the features of the X7SLA-L with dual Gigabit Ethernet ports, an additional onboard Type A USB 2.0 connector and an extra internal serial port.

Weighing only ten pounds and with a depth of 9.8 inches, Supermicro's mini-sized 5015A SuperServers feature a high-efficiency, quiet power supply, two internal hard drives, and a full-height, half-length expansion card option. The fan-less design of Supermicro's SC502-200B chassis enables silent operation and minimizes system power consumption.

For greater expansion and storage features, Supermicro offers the SuperServer 5035A, which supports four hot-swap drive bays and three add-on cards. This server features a high-efficiency 300-watt power supply that has earned the 80-Plus® Bronze level certification for achieving 85%+ power efficiency. The system's mini-tower chassis (SC731i-300B) also includes a host of user-friendly design features for easy installation. These quick-and-easy features include a toolless side panel, toolless drive bays, 90-degree pivoting hard disk drive cage, and stamped motherboard support (no standoffs required). This small footprint server also supports two external 5.25" bays and Kensington lock security to safeguard the system.

Although Supermicro targets these at IPC applications, they will no doubt make their way into servers. These solutions are based on the desktop variants of the Atom, which support useful things like x86-64, which are nowadays very important for that kind of environment.
The choice is yours, virtualization with very expensive Xeons/Opterons or a Fast Array of Wimpy Nodes?