Drive Thread oficial de novidades e lançamentos SSD

E quando um gajo pensa que já viu de tudo

81791a2f1a867bef2e84a5c1c6c10e6d.gif


Xilinx-Samsung SmartSSD Computational Storage Drive Launched

Samsung-Xilinx-SmartSSD-Internal-Components.png

The basic flow is that commands can be issued to either the SSD or the FPGA portion of the drive and processing can occur at the FPGA instead of going back to the host system.
Clicar para expandir...
Samsung-Xilinx-SmartSSD-Internal-Operation.png


Xilinx-SmartSSD-IP-Development.png


Xilinx-SmartSSD-IP-Xilinx-Storage-Services.png

In terms of examples, we wanted to highlight one from Lewis Rhodes Labs where they are doing NPUSearch using computational storage. Effectively here the SmartSSDs are being used to scale out the number of accelerators with the number of SSDs. An application can send requests to the storage, data can be evaluated at the drives, and only results passed back to the main system.
Clicar para expandir...
Xilinx-SmartSSD-Lewis-Rhodes-Labs-Search.png

https://www.servethehome.com/xilinx-samsung-smartssd-computational-storage-drive-launched/
 
Quanto mais melhor... com o tempo veremos mais capacidade e velocidades no mercado consumidor... primeiro vemos em datacenter... ou seja, nunca é demais....

E eu me lembro que ter um 5GB era algo de loucos, nunca se via o fim para ocupar tanta espaço :D estou a falar de muitos anos atrais mesmo...muito... 1GB dava para sistema, programas pesados e até alguns jogos :D depois um de 5GB para dados...

Nos dias de hoje, para a minha realidade, é claro 1TB é muito pouco... em especial quando pensamos em ter 3 bacups de tudo! e falamos em video e foto... ui 1TB é nada... em especial formatos RAW....
 
A NEO Semiconductor apresentou os detalhes da "X-NAND". Alegadamente tem a velocidade de SLC, com a densidade e preço de QLC.

dVPiPDZ.gif


blMRsVB.gif


dFsy81b.png


qekQi9c.gif


X-NAND promises intriguing performance numbers: The company claims it can do random read and write workloads 3x times faster than QLC flash, and beat it by 27x/14x for sequential read and write workloads, respectively (see above). This is achieved with a far smaller die that's roughly 37% the size of a 16-plane design (see below). There is some flexibility here as speed and die size reduction can be balanced as needed.
Clicar para expandir...

https://www.tomshardware.com/news/new-x-nand-tech-detailed-slc-speed-at-qlc-density-and-cost

Está interessante. :)
 
Não é só tech para controladores, acho. Parece ser uma arquitectura nova, ao que parece, ou pelo menos tem novidades, nomeadamente na cache. E parece-me ter também melhorias na densidade.
 
Sobre o X-NAND:

"It always writes first in SLC mode, and further always at SLC speeds. There's three banks with the first writing incoming data to SLC, a second bank where SLC data is moved to QLC, and a third bank where SLC is erased to make room for more writes. The reason this works so well is because SLC mode tends to have a tPROG (write latency) of ~200µs - this applies to NAND in consumer drives as well, by the way - while QLC in his example is at a total of 6400µs (this is not precise, as newer consumer QLC can manage 1/2 to 1/3 this latency, but he's talking for all pages). Therefore you can write 32 SLC pages (e.g. 32x200µs = 6400µs) while keeping pace with data moving to QLC, therefore avoiding the SLC cache drop-off you usually see. This makes QLC uniquely qualified for this technology but that's also true of other characteristics - namely, if you're splitting 16KB pages into 4KB I/O chunks/subpages (16 / 4) and then the page buffer into 1KB chunks per plane (4 / 4), 4-bit QLC with 1-bit SLC is a match made in heaven.

As for endurance, typically on a consumer drive you'll have static, dynamic, or a hybrid SLC caching scheme. Dynamic SLC shares a wear zone with native flash, you're converting back-and-forth, you can even increase write amplification with it. Static SLC, however, is dedicated and has its own wear zone, usually in the 30K-40K P/E range (vs. 100K with native SLC or 1.6K with dynamic SLC + QLC). If all the writes are done to this sort of static SLC, depending on workload you can increase overall endurance significantly for a variety of reasons. For example, writing/folding out from SLC is done sequentially and not randomly which reduces write amplification. You can also defer writes or erases. I'm not privy to the exact workings of X-NAND in this capacity (I made a lot of assumptions) but I am in contact with the creator and will certainly be investigating this more. However, it basically offers SLC-like performance with respectable endurance.

There was a query about endurance during FMS, and this was his reply:

"No, the endurance will not be reduced. Because the SLC pages are not erased immediately. Although their data is re-programmed to QLC page immediately, the SLC pages’ erasure will be hold until all the SLC pages in the bank are programmed, then all the SLC pages of the bank will be erased together. Therefore, the endurance is the same as conventional SLC cache. ""
 
erdnagama disse:
Aqui é que está a questão. Para o dia a dia DLC, TLC ou QLC vai dar tudo ao mesmo SE (muito importante este SE) for acompanhado com DRAM ajustado à dimensão do disco.

Se as marcas não o fazem, cabe às lojas tentarem identificar os produtos como deve de ser, garantindo que o cliente tem noção dos méritos e limitações do produto que está a comprar.
Clicar para expandir...

Eu não vou pelas lojas que nem querem saber disso. Se não for as reviews a dizerem que o produto tem DRAM cache, estamos lixados. Nem os sites dos fabricantes ás vezes, referem esse pormenor.
 
A Intel lançou 2 SSDs QLC hoje.

670p:
1L3uLC0.png

KLa1RHJ.png


Moving on to the consumer SSD business, Intel will be introducing the 670p QLC NVMe SSD in the first quarter of 2021. This follows in the footsteps of the 660p and 665p with another update to newer 3D QLC NAND, and also brings a SSD controller update to this product line. However, it is still only a PCIe gen3 product. Intel will be reintroducing the 512GB capacity that was missing from the 665p, but they are not yet adding capacities beyond 2TB. Detailed performance specs and pricing will be shared closer to the 670p's launch.

Intel is making an adjustment to the dynamically-sized SLC caching behavior with the new 670p. While the maximum and minimum SLC cache sizes are not changing, Intel has managed to improve the cache size that will be available for a partially-filled drive: A half-full 670p will still have almost the maximum SLC cache size available, and the minimum cache size will not be reached until the drive is over 85% full. This probably won't change anything for simple benchmarks that write continuously until the cache runs out and performance drops, but it will help real-world usage where writes come in bursts.
Clicar para expandir...

As novidades parece ser usar NAND QLC 144 Layers e alteração da forma como funciona a SLC Cache.

H20:
12PWtpC.jpg


The client/consumer focused portion of Intel's Optane product family has shrunk considerably. They're no longer doing Optane M.2 SSDs for use as primary storage or cache drives, and there's been no mention yet of an enthusiast-oriented derivative of the P5800X to replace the Optane SSD 900P and 905P (though if Intel plans such a product, they are unlikely to announce it until they have delivered a desktop platform supporting PCIe 4.0). The only client Optane product Intel has been talking about lately is the Optane Memory H10 hybrid drive consisting of a QLC NVMe SSD and an Optane SSD on the same M.2 card. A successor is now on the way: the Optane Memory H20 is planned for Q2 2021, incorporating updates on the QLC side that likely mirror the 670p's updates, and a new controller on the Optane half of the drive. The H20 will still be a PCIe gen3 solution, so we will probably continue to see performance of the NAND side constrained by only having access to two of the four PCIe lanes. Intel is also bumping up the platform requirements: an 11th-generation Core U-series mobile processor and 500-series chipset, and Intel RST driver version 18.1 or later. When the H10 launched, Intel had laid the groundwork for support on their desktop platforms, but this effort was dropped when the H10 became an OEM-only product. The H20 is a mobile-focused OEM-only part from the outset, so the platform compatibility requirements aren't as much of an issue as they would be for a retail SSD.

The Optane Memory H20 will be available with 512GB or 1TB of QLC NAND flash memory, each paired with 32GB of 3D XPoint memory.
Clicar para expandir...

SSD QLC com 512 ou 1 TB + 32 GB Optane a fazer de Cache. Por agora parece ser OEM Only para o mercado de Portáteis.


Fora de QLC, também apresentaram este monstrinho com Optane de 2 Geração. P5800X:
8B8dOSl.jpg


Pci-Ex Gen 4. Reads sequenciais, aleatórios e Writes aleatórios, saturam as 4 lanes Pci-Ex Gen4. Writes sequenciais um pouco menos a 6,2 GB/s. Entre 1.5 e 1.8 milhões de Iops.
100 DWPD...... :n1qshok:. Fazendo umas contas, na versão 400 GB, dá para escrever 40 TB por dia durante 5 anos. Na versão 3,2 TB, dá para escrever 320 TB por dia durante 5 anos. 320 TB durante 5 anos são 584 PB escritos. :D
I like it. :)

https://www.anandtech.com/show/16318/intel-announces-new-wave-of-optane-and-3d-nand-ssds
 
skandal disse:
Esses P5800X são meninos para que preços? Queria uma coisa dessas como cache para pools ZFS que são só SSDs :)
Clicar para expandir...

u4JD3dX.png


Para o tamanho de Storage, se nada mudar a nível de latências, deve ser a drive ideal para ZIL/SLOG em ZFS. Melhor só mesmo MRAM, mas a maior drive actualmente tem apenas 1 GB e não deve ser nada barata.
 
Parece que vêm aí aumentos de preços na NAND, no primeiro trimestre de 2021, que por sua vez, devem afectar os preços de SSDs e derivados.

EviKyjB.jpg


Prices of NAND Flash Controller ICs Poised to Rise by 15-20% due to Tightening Production Capacity for Foundry Services, Says TrendForce

In the upstream semiconductor industry, the major foundries such as TSMC and UMC are reporting fully loaded capacities, while in the downstream, the available production capacity for OSAT is also lacking, according to TrendForce's latest investigations. Given this situation, suppliers of NAND Flash controller ICs such as Phison and Silicon Motion are now unable to meet upside demand from their clients. Not only have many controller IC suppliers temporarily stopped offering quotes for new orders, but they are also even considering raising prices soon because the negotiations between NAND Flash suppliers and module houses over 1Q21 contracts are now at the critical juncture. The potential increases in prices of controller ICs from outsourced suppliers (IC design houses) are currently estimated to be the range of 15-20%.

With regards to the demand side, demand has risen significantly for eMMC solutions with medium- and low-density specifications (i.e., 64 GB and lower), for which NAND Flash suppliers have mostly stopped updating the NAND Flash process technology, while maintaining support with the legacy 2D NAND or the 64L 3D NAND process. This is on account of strong sales for Chromebook devices and TVs. As older processes gradually account for a lowering portion of bit output proportions from NAND Flash suppliers, these companies are exhibiting a lowered willingness to directly supply such eMMC products to clients. As a result, clients now need to turn to memory module houses, which are able to source NAND Flash components and controllers, to procure eMMC products in substantial quantities.

Hikes in prices of controller ICs will lead to hikes in module prices with mainstream 32 GB and 64 GB solutions for Chromebooks experiencing largest increases
TrendForce points out that even though the whole NAND Flash market is still in oversupply right now, medium- and low-density eMMC solutions will likely experience price hikes as they are in tight supply due to insufficient output of controller ICs. An increase in prices of controller ICs will result in a corresponding increase in the fixed costs of eMMC solutions. This, in turn, will put pressure on OEMs that are procuring components. With this scenario now becoming a reality, prices of those eMMC solutions that are in high demand could rise slightly in 1Q21. Examples include 32 GB and 64 GB solutions for Chromebook devices.

In the SSD market, the major NAND Flash suppliers such as Samsung are also the major device manufacturers. Most SSD manufacturers have in-house controller ICs that are made at foundries under long-term agreements. Hence, there have been no reports of price hikes or shortages for SSD controller ICs. However, TrendForce has observed that lead time has been prolonged for SSD controller ICs as well. Additionally, the share of SSD controller ICs that have been outsourced to IC design houses has increased for SSDs with PCIe 4.0. In the future, there is an increasing possibility that prices of SSD controller ICs will be affected by the statuses of outsourced IC suppliers.
Clicar para expandir...

https://www.techpowerup.com/276340/...capacity-for-foundry-services-says-trendforce
 
Inicie sessão ou registe-se para poder participar.
Back
Topo