two lines of Intel processors, new GPUs from AMD, platforms and Russian servers
Our regular server hardware review column continues — now in 2024. In this issue, we will look at the devices introduced in December. Let’s dive into new platforms, learn new disks and data storage systems. Already taking care of a new iron for yours
? Then welcome to the hangman! Details below!
- 1 Processors
- 2 GPU
- 3 Platforms
- 3.1 Platforms for Instinct MI300X
- 3.2 Platforms for Intel Xeon Scalable 5th Emerald Rapids
- 3.3 Platforms for Intel Xeon 2400
- 3.4 Aquarius platforms for Intel Xeon Scalable 3rd IceLake-SP
- 3.5 Delta Sprut extension module
- 4 Discs
- 5 SGD
In December, I devoted an article to two lines of server processors from Intel. You can read the details in it, and I will remind you of the short characteristics in the tables.
Intel Xeon Scalable 5th Emerald Rapids
AMD Instinct MI300X
AMD company presented its top solution for tasks of artificial intelligence and high-performance computing.
built on the CDNA 3 architecture. The device is equipped with 192 GB of HBM3 video memory with a throughput of 5.3 TB/s and 304 executive units CUs.
The MI300X uses the fourth-generation AMD Infinity Fabric bus to connect multiple devices. The maximum number of devices in a cluster is 8.
Compared to the Nvidia H100, the main competitor in the market, the MI300X has more than twice the video memory (192 vs. 80 GB) and 1.5 times the bandwidth (5.3 vs. 3.35 TB/s). The manufacturer also claimed AMD’s advantage in inference tasks – the result is 1.6 times better.
AMD’s eight MI300X platform delivers 10.4 petaflops of performance in FP16 and BF16 operations, beating Nvidia’s HGX platform with H100 by 7.9 petaflops. However, in the near future Nvidia will introduce the H200 model – I’m looking forward to the benchmarks and comparison of product characteristics.
AMD Instinct MI300A
AMD was also introduced
— Accelerated Processing Unit for data centers, which combines a processor and a GPU. The device will directly compete with Nvidia’s Grace Hopper superchips.
The APU Instinct MI300A is equipped with an x86 “Zen 4” processor with 24 cores and 48 threads, a GPU module c 128 GB HBM3 video memory with a bandwidth of 5.3 TB/s and 228 executive units CUs. Nvidia’s Grace Hopper uses ARM processors. AMD claims that the Instinct MI300A processor achieves 61 teraflops in FP64 operations and 61 teraflops in FP32 operations.
Platforms for Instinct MI300X
Giga Computing (Gigabyte)
presented three models of servers
on the AMD hardware platform, designed to work with the new video accelerators Instinct MI300X and APU Instinct MI300A.
Supermicro, which is also not far behind
offered three models of servers
on the AMD hardware platform, designed to work with GPU Instinct MI300X and APU Instinct MI300A.
The final price comparison of Gigabyte and Supermicro devices with similar solutions from Nvidia is interesting, but the number of potential users of such devices in Russia, in my opinion, is small.
Dell and AMD also announced the release of Dell PowerEdge XE9680 servers with eight AMD Instinct MI300X accelerators in 2024, but the final product has not yet been announced.
Platforms for Intel Xeon Scalable 5th Emerald Rapids
TYAN, part of MiTAC Computing Technology Corporation (MCT), has introduced
their first products
based on the new Intel Xeon Scalable 5th Emerald Rapids processors. The announced products include platforms in various form factors, including tower systems and motherboards.
Asus unexpectedly presented
16 platform models
for Emerald Rapids. Starting from EDGE servers with compact dimensions and low power consumption, ending with monstrous ones
8 GPU servers
in a 7U form factor with six 3000 W power supplies.
Two models were presented
servers for Intel Xeon Scalable 5th
nine server models and 11 motherboard models for Intel Xeon Scalable 5th. Servers
single-socket models and two-socket platforms in the form factor from 1U to 4U with
support for up to 10 GPUs
Platforms for Intel Xeon 2400
Giga Computing (Gigabyte) presented four models of single-unit servers for Intel Xeon E-2400 processors, which differ in size and disk configuration. All models use 550 or 800 W power supplies. They also support up to 128 GB of DDR5 memory with 4800 MT/s ECC error correction.
Lenovo also released platforms for E-2400 processors. They are represented by a 1U rack platform
Lenovo ThinkSystem SR250 V3
and tower server
Lenovo ThinkSystem ST250 V3
. Both platforms support various configurations of NVMe/SSD/HDD drives and a fourth DDR5-4400 memory slot
Also a solution for the E-2400
announced by ASRock
in the form of single-unit servers
. Similarly to competitors, the servers have 4 DDR RAM slots, support for NVMe drives. The main difference is in the configuration of the power supply units, in the first case it is one at 400 W, in the second two at 450 W.
Asus introduced the server
and motherboards for E-2400. Motherboards
Usually with four DDR5 RAM slots, differ in the configuration of built-in network adapters and SATA-3 ports. A two-port 1GbE adapter and 8 SATA-3 ports are available for the first motherboard, and a two-port 10GbE controller and 6 SATA-3 ports are available for the second.
The server platform is equipped with two 450 W units
Aquarius platforms for Intel Xeon Scalable 3rd IceLake-SP
Russian company Aquarius
presented the servers
based on Intel Xeon Scalable processors of the third generation, intended for use in enterprises and data centers. The servers are entered in the Register of the Ministry of Industry and Trade of the Russian Federation.
The Aquarius T50 D110AC flagship model supports the installation of up to two Intel Xeon Ice Lake-SP processors with a maximum TDP of up to 270 Watts, as well as two GPU accelerators. The server is equipped with 32 slots for RAM modules DDR4-3200, has the ability to install up to ten drives of the SFF form factor, as well as two M.2 drives in the form factors 2280 and 22110 with SATA/NVMe interfaces. Power is provided by two power supply units with a capacity of 1600 or 2000 watts each.
Delta Sprut extension module
Company Delta Computers
the first Russian OCP hardware platform for GPU
. It is a specialized JBOG expansion module that allows you to add additional PCIe devices to servers for AI, high-performance clusters, and 3D VDI applications. The company plans to enter the equipment into the Register of the Ministry of Industry and Trade of the Russian Federation.
The module is made in a 2OU form factor and has four PCIe 4.0 x16 connectors for installing four HHHL or two FHFL cards with a power of up to 350 W. Compatible with Delta Tioga Pass and Delta Bright Lake compute nodes.
The Delta Sprut allows you to cascade up to 8 or 16 double-thickness or standard GPUs, respectively, in a single OCP case. You can also use accelerators from Nvidia (A100, H100) and AMD Instinct. For Nvida devices, pairing is supported by NVLink bridges.
SSSTC SSD M.2
State Storage Technology Corporation (subsidiary structure of Kioxia) has introduced M.2 SSD storage PJ1 series with PCIe 4.0 x4 interface (NVMe specification 1.4c), designed for use in data centers. Disks are presented in two standard sizes:
at 480GB, 960GB and 1.92TB as well
for 960 GB, 1.92 TB and 3.84 TB.
Storage devices are equipped with the TruePLP power loss protection system. It uses backup capacitors to maintain power in the event of failures, allowing the SSD to correctly store data in outage situations. The five-year DWPD indicator is from 1 to 1.2 – depending on the model.
In Selectel’s dedicated servers, our customers can order SSD M.2 up to 2 TB. However, we recommend using drives of this format as secondary drives because they are not hot-swappable.
Seagate Technology has announced a new hard drive
SkyHawk AI with a capacity of 24 TB
for use in data centers. The drive is designed for image processing and analysis using artificial intelligence and deep learning algorithms.
The ST24000VE002 model uses CMR magnetic recording technology and has a helium chamber. The connection is made via the SATA-3 interface, the size of the cache is 512 MB, and the declared MTBF is 2.5 million hours.
Once again, I will tell you about a simple formula for calculating the cost of disks: it is enough to divide the cost by the number of TB and compare it with the previous size model (22 TB). It is good if these indicators are close.
You may also be interested in these texts:
→ What to do on vacation? 5 ideas for pet projects that can be done during the New Year holidays
→ Not washing, but rolling: Nvidia will start supplying China with powerful video cards that are not subject to US sanctions
→ It’s time to turn back the hands of the clock: a message to scientists for 2024
QNAP PCIe JBOD expansion modules
QNAP Systems has announced the release of
three new modules
extension for network attached storage (NAS). They are PCIe JBOD rack mounts and can hold 12 to 24 drives. At the same time, up to 12 such modules can be installed in one system.
New devices support both SSD and HDD with SATA 3 interface. Hot swapping of drives is possible. The modules use the SFF-8644 host interface and require a QXP-3X8PES or QXP-3X4PES card to connect. The devices can operate in temperature ranges from 0°C to 40°C. In addition, for convenient management, the developers implemented the function of remotely turning on and off the modules via NAS.
CES 2024 is happening right now, where key players in the tech industry are unveiling new products. This show is likely to be the main source of new iron news in January.
What do you think about the new products? Share your opinion in the comments!