Subject: Processors | May 16, 2017 - 06:49 PM | Sebastian Peak
Tagged: Zen, server, ryzen, processor, EPYC, datacenter, cpu, amd, 64 thread, 32 core
AMD has announced their new datacenter CPU built on the Zen architecture, which the company is calling EPYC. And epic they are, as these server processors will be offered with up to 32 cores and 64 threads, 8 memory channels, and 128 PCI Express lanes per CPU.
Some of the details about the upcoming "Naples" server processors (now EPYC) were revealed by AMD back in March, when the upcoming server chips were previewed:
- A highly scalable, 32-core System on Chip (SoC) design, with support for two high-performance threads per core
- Industry-leading memory bandwidth, with 8-channels of memory per "Naples" device. In a 2-socket server, support for up to 32 DIMMS of DDR4 on 16 memory channels, delivering up to 4 terabytes of total memory capacity.
- The processor is a complete SoC with fully integrated, high-speed I/O supporting 128 lanes of PCIe, negating the need for a separate chip-set
- A highly-optimized cache structure for high-performance, energy efficient compute
- AMD Infinity Fabric coherent interconnect for two "Naples" CPUs in a 2-socket system
- Dedicated security hardware
Compared to Ryzen (or should it be RYZEN?), EPYC offers a huge jump in core count and available performance - though AMD's other CPU announcement (Threadripper) bridges the gap between the desktop and datacenter offerings with an HEDT product. This also serves to bring AMD's CPU offerings to parity with the Intel product stack with desktop/high performance desktop/server CPUs.
EPYC is a large processor. (Image credit: The Tech Report)
While specifications were not offered, there have been leaks (of course) to help fill in the blanks. Wccftech offers these specs for EPYC (on the left):
(Image credit: Wccftech)
We await further information from AMD about the EPYC launch.
Introduction, Specifications and Packaging
Intel launched their Datacenter 'P' Series parts a little over two years ago. Since then, the P3500, P3600, and P3700 lines have seen various expansions and spinoffs. The most recent to date was the P3608, which packed two full P3600's into a single HHHL form factor. With Intel 3D XPoint / Optane parts lurking just around the corner, I had assumed there would be no further branches of the P3xxx line, but Intel had other things in mind. IMFT 3D NAND offers greater die capacities at a reduced cost/GB, apparently even in MLC form, and Intel has infused this flash into their new P3520:
Remember the P3500 series was Intel's lowest end of the P line, and as far as performance goes, the P3520 actually takes a further step back. The play here is to get the proven quality control and reliability of Intel's datacenter parts into a lower cost product. While the P3500 launched at $1.50/GB, the P3520 pushes that cost down *well* below $1/GB for a 2TB HHHL or U.2 SSD.
Subject: Storage | August 10, 2016 - 02:00 PM | Allyn Malventano
Tagged: 2.5, V-NAND, ssd, Samsung, nand, FMS 2016, FMS, flash, 64-Layer, 32TB, SAS, datacenter
..now this picture has been corrected for extreme parallax and was taken in far from ideal conditions, but you get the point. Samsung's keynote is coming up later today, and I have a hunch this will be a big part of what they present. We did know 64-Layer was coming, as it was mentioned in Samsung's last earnings announcement, but confirmation is nice.
*edit* now that the press conference has taken place, here are a few relevant slides:
With 48-Layer V-NAND announced last year (and still rolling out), it's good to see Samsung pushing hard into higher capacity dies. 64-Layer enables 512Gbits (64GB) per die, and 100MB/s per die maximum throughput means even lower capacity SSDs should offer impressive sequentials.
Samsung 48-Layer V-NAND. Pic courtesy of TechInsights.
We will know more shortly, but for now, dream of even higher capacity SSDs :)
*edit* and this just happened:
*additional edit* - here's a better picture taken after the keynote:
The 32TB model in their 2.5" form factor displaces last years 16TB model. The drive itself is essentially identical, but the flash packages now contain 64-layer dies, doubling the available capacity of the device.
Subject: Processors | March 15, 2016 - 12:52 PM | Sebastian Peak
Tagged: TSMC, SoC, servers, process technology, low power, FinFET, datacenter, cpu, arm, 7nm, 7 nm FinFET
ARM and TSMC have announced their collaboration on 7 nm FinFET process technology for future SoCs. A multi-year agreement between the companies, products produces on this 7 nm FinFET process are intended to expand ARM’s reach “beyond mobile and into next-generation networks and data centers”.
TSMC Headquarters (Image credit: AndroidHeadlines)
So when can we expect to see 7nm SoCs on the market? The report from The Inquirer offers this quote from TSMC:
“A TSMC spokesperson told the INQUIRER in a statement: ‘Our 7nm technology development progress is on schedule. TSMC's 7nm technology development leverages our 10nm development very effectively. At the same time, 7nm offers a substantial density improvement, performance improvement and power reduction from 10nm’.”
Full press release after the break.
Introduction, Specifications and Packaging
What's better than an 18-channel NVMe PCIe Datacenter SSD controller in a Half Height Half Length (HHHL) package? *TWO* 18-channel NVMe PCIe Datacenter controllers in a HHHL package! I'm sure words to this effect were uttered in an Intel meeting room some time in the past, because such a device now exists, and is called the SSD DC P3608:
The P3608 is essentially a pair of P3600's glued together on a single PCB, much like how some graphics cards merge a pair of GPUs to act with the performance of a pair of cards combined into a single one:
What is immediately impressive here is that Intel has done this same trick within 1/4 of the space (HHHL compared to a typical graphics card). We can only imagine the potential of a pair of P3600 SSDs, so lets get right into the specs, disassembly, and testing!
Subject: Storage | May 27, 2015 - 10:00 PM | Sebastian Peak
Tagged: storage, SAN, S3200, S2200, Lenovo, datacenter
Lenovo has announced two new high-performance storage products aimed at small and medium business, and the new S2200 and S3200 storage arrays are designed with speed in mind.
The Storage S2200 and S3200 arrays offer dual and single controllers in 2U-12 and 24 drive configurations. The S2200 supports up to 96 drives and the S3200 supports up to 192 drives to easily support storage growth. The S2200 and S3200 make connectivity simple. The S2200 and S3200 support Fibre Channel, iSCSI and SAS, with the S3200 supporting multi-protocol connectivity that can work with Fibre Channel and iSCSI at the same time. This combination of flexibility and scalability makes integration into nearly any environment easy.
Lenovo is also using a technology called "Intelligent Real-Time Tiering" to approximate the performance of flash storage by prioritizing frequently accessed data as it "automatically moves frequently accessed data to higher performing drives every five seconds, significantly increasing storage performance".
With hybrid configurations and Intelligent Real-Time Tiering, the Lenovo Storage S3200 can provide near All-Flash-Array (AFA) performance for up to 120,000 IOPS at a fraction of the cost of today’s Flash only systems.
The Lenovo S2200 and S3200 SANs will be available worldwide starting in June.
Subject: Storage | March 23, 2015 - 03:47 PM | Jeremy Hellstrom
Tagged: toshiba, MG04ACA400A, datacenter, enterprise
Toshiba's new MG04ACA series are Enterprise class HDDs available in increments of 1TB, from 2TB to 6TB and ship with either 4K or 512B emulation depending on your preference. Mad Shrimps just wrapped up a review of the 4TB model which certainly cannot match a SSD for speed but it is rated for 1400000 hours and workloads of 550TB a year, constant usage. You do pay a premium for enterprise level drives but spinning rust is still far more economical in high densities that flash based drives are. If you are looking for reliable HDDs for your servers, check this review out.
"The new MG04ACA series from Toshiba is composed from drives which are meant for enterprise, mission-critical applications, while sporting higher transfer rates and capacities. The tested sample comes with 128MB of cache and comes in two versions, depending on the applications it is needed for: with 512 sector emulation or strictly with 4K sector. Make sure to choose wisely which drive is for you and your setups in order to bypass any incompatibilities which may arise."
Here are some more Storage reviews from around the web:
- ASUSTOR AS5104T 4-bay NAS Review @ Madshrimps
- Asustor AS7004T @ techPowerUp
- Patriot Ignite 480GB SSD @ Kitguru
- Kingston SM2280S3 120GB M.2 SATA SSD @ Bjorn3d
Subject: Processors | November 6, 2012 - 01:15 PM | Tim Verry
Tagged: server, piledriver, opteron, datacenter, cpu, amd
AMD announced new server processors on Monday based on the same Piledriver architecture used in the Trinity APUs and Vishera desktop CPUs we recently reviewed. With the release of the Opteron 6300 series, AMD is bringing Piledriver to the server room.
The new chips – similar to the desktop counterparts – bring several performance improvements over the previous generation 6200 series Opterons based on the Bulldozer architecture. AMD is positioning the chips as a upgrade path to existing servers and on merits of performance-per-dollar efficiency. As is AMD's fashion, the new chips are competitively priced and "good enough" performance-wise. With 6300, AMD has stated the goal is to reduce the TCO, or Total Cost of Ownership for servers used in data centers, supercomputers, and enterprises by being compatible with existing AMD server platforms with a BIOS upgrade and representing efficiency improvements over previous chips.
The Opteron 6300 series CPUs themselves build upon the Vishera desktop parts by adding more cores and more L3 cache. The server parts will have up to 16 cores clocked at 2.8GHz base and 3.2GHz turbo. They will have TDP ratings between 85W and 140W and will feature prices from $500 to $1,400. On the cache front, the chips have a 16KB L1 data cache per core, 64KB L1 instruction cache per module, 1MB L2 cache per core, and a shared 16MB cache per socket. AMD has included a quad channel memory controller that supports DDR3 up to 1866 MHz and 1.5TB per server in 4P configurations. AMD has rounded out the chips with four x16 HyperTransport 3.0 links rated at 6.4 GT/s per link. Up to 4 processors per server will be supported, which means a maximum of 64 cores.
With Piledriver, AMD added a number of new instructions including FMA3, BMI, and F16c. The company has also implemented server tweaks to the Bulldozer design to improve branch prediction, instructions per clock, scheduling, and reduced the power draw at higher clockspeeds allowing for the chps to clock higher while staying within the same power envelope of the Bulldozer-based Opteron 6200 series.
AMD is using the same socket as the 6200 series processors, and the new chips can be deployed as an upgrade to the old servers without needing a new motherboard.
When pitting the new Opteron 6380 to the previous-generation 6278, AMD is claiming a number of performance increases, including a 24-percent and 40-percent improvement in SPECjob2005 and SPECpower_ssj2008 respectively.
Further, the company is claiming competitive performance in server workloads with the Intel competition. AMD offers up benchmarks showing the Opteron 6380 and Xeon E5-2690 trading wins, with the AMD part being slower in the STREAM benchmark, but being slightly faster in LAMPS and NAMD. The allure of the Opteron, according to AMD is that the AMD part is almost half the price of the Intel processor, and is hoping the lower priced parts will encourage adoption. AMD argues that the money saved could easily go towards more RAM or more storage (or simply be saved of course).
The company has announced that its first major design win is Big Red II supercomputer at Indiana University. Built by Cray, the Big Red II will feature 21,000+ Opteron 6300-series CPU cores paired with NVIDIA GPUs. It represents a massive increase in computing power over IU’s previous Big Red supercomputer with 4,100 CPU cores, and will be used for medical, physics, chemistry, and climate research. Beyond that, AMD has stated more that 30 hardware vendors are slated to introduce servers based on the new Piledriver-based Opteron processors including HP, Dell, Cray, SGI, Supermicro, Sugon, and (of course) SeaMicro. On the software side of things, AMD is working with Microsoft, VMware, Xen, Red Hat, and Openstack. The company also stated that it is leaning on the experience and knowledge gained from the HSA Foundation to improve software support and guide the future direction of Opteron development.
The Opteron 6300 series is an interesting release that brings several improvements to the company’s server chip offerings. At launch, there are 10 processors to choose from, ranging from the quad core 6308 clocked at 3.5GHz for $501 to the top-end 6386 SE with 16 cores (2.8GHz base, 3.5GHz max turbo) and a $1,392 price tag. The 6366HE is an interesting part as well. It is the same price as the 12-core, 115W TDP Opteron 6348, but its has 16 lower-clocked cores and an 85W TDP. With the non-HE edition processors with 16 cores starting at $703, the 6366HE for $575 is a decent deal if you need multi-threading more than a fewer number of higher clocked cores.
Another bit that I found intriguing is that in a few years, AMD will (likely, if all goes according to plan) be offering processors for just about every type of server. They will have low cost, low power ARM Cortex-A57 based chips, Accelerated Processing Units (APUs) well suited to mixed workloads including GPU-accelerated tasks, and CPU-only chips with lots of traditional x86-64 cores. It seems that Intel will continue to hold the high end on pure performance, but AMD and its SeaMicro server division have not given up competing in the server room by a long shot.
The Piledrive architecture and Vishera desktop CPU review and The future of AMD: Vishera and Beyond at PC Perspective.
Subject: Storage | November 5, 2012 - 12:39 PM | Allyn Malventano
Tagged: ssd, s3700, enterprise, datacenter
Today Intel officially launched a new line of enterprise-oriented SSDs. Dubbed the DC S3700 ('DC for Data Center', 'S' for SATA), this new line fills the large interface speed void left by the older 710 Series, which was limited to SATA 3Gb/sec speeds.
The S3700 makes some big promises and we are expecting samples shortly. Here's the tally of what's to come:
- Intel designed 8-channel controller ASIC and firmware
- SATA 6Gb/sec interface
- 7mm x 2.5" form factor
- Random 4k writes 15x faster and reads 2x faster than SSD 710
- 75,000 4k random read IOPS (all models)
- 19/32/36/36k 4k random write IOPS (for 100/200/400/800GB capacity)
- 500MB/sec sequential reads (all models)
- 200/365/460/460 MB/sec sequential writes (for 100/200/400/800GB capacity)
- 25nm MLC-HET IMFT flash
- Rated for 10 Drive Writes per Day (DWPD) over a 5-year lifetime
- Solid-State-Capacitor backed power-loss protection
- Shipping in volume ~Q1 2013
- 1k qty pricing: $235/$470/$940/$1880 ea. for 100/200/400/800GB capacities
The cost of just over $2/GB should be very enticing for an enterprise-grade SSD, but the most interesting tidbit I got from the briefing was that Intel claims this drive will achieve a <500us response time for 4k random writes, 99.9% of the time. Most SSDs will begin to show intermittent peaks in latency when hit with sustained 4k random access. The S3700 Series should mostly eliminate that issue. More to follow on that front once we can log some hours on a sample.