Project Denver - nVIDIA RISCs with ARM

http://www.theregister.co.uk/2013/12/19/calxeda_shutdown/



Pih, nije im ni "head of hardware supply" iz Facebooka pomogao.

Potpis

DigiCortex (ex. SpikeFun) - Cortical Neural Network Simulator:
http://www.digicortex.net/node/1 Videos: http://www.digicortex.net/node/17 Gallery: http://www.digicortex.net/node/25
PowerMonkey - Redyce CPU Power Waste and gain performance! - https://github.com/psyq321/PowerMonkey

Možda ARM zaista nema šta da traži trenutno na tržištu servera, ali bi tržište jako profitiralo od konkurencije ARM i x86 arhitektura (naravno, pod uslovom da ARM odradi posao kako treba). To je, inače, sasvim moguće, u vreme kad je Sun Microsystems još postojao, bili su konkurentni T serijom UltraSPARC procesora za servere.

ARM trenutno nema sta da trazi na trzistu servera, sto ovo gore i pokazuje.

To ne znaci da ARM ne moze da postane konkurentna platforma - i to bi svakako bilo dobro za trziste.

Ali da bi se to desilo, neko ce morati da poradi na tome da:

a) ARM implementacije koriste standardne interkonekt-e (PCIe)
b) ARM implementacije koriste standardni firmware (npr. UEFI)
c) ARM implementacije imaju konkurentne CPU/FPU performanse
d) ARM implementacije imaju konkurentne I/O performanse

Na mobilnim telefonima nije nikakav problem sto svaki ARM SoC ima svoju arhitekturu za komunikaciju sa periferijama, potpuno razlicit firmware bez ikakvog standarda za inicijalizaciju platforme itd.. - posto telefone niko ne unapredjuje/konfigurise, niti postoji neka specijalna potreba za bilo cim od custom kernel drajvera koje SoC vendori izbace za Linux.

Ali to na serverskom trzistu, jednostavno, nema sanse da prodje - mnogi su probali, pa se nisu lepo proveli.

Potpis

DigiCortex (ex. SpikeFun) - Cortical Neural Network Simulator:
http://www.digicortex.net/node/1 Videos: http://www.digicortex.net/node/17 Gallery: http://www.digicortex.net/node/25
PowerMonkey - Redyce CPU Power Waste and gain performance! - https://github.com/psyq321/PowerMonkey

Karte su konačno na stolu.

Nvidia details 64-bit Denver Tegra K1, claims Haswell-class performance for first 64-bit Android chip
By Sebastian Anthony on August 12, 2014 at 7:38


Nvidia has finally taken the wraps off its 64-bit Denver CPU core — and, well, let’s just say that the 64-bit Android race just got a lot more interesting. The Denver CPU will be available later this year in the Tegra K1 SoC, where you’ll get two Denver cores and a 192-core Kepler GPU. Nvidia says this will be the first 64-bit ARM chip for Android, and that it “completely outpaces” other ARM SoCs (including Apple’s monstrous A7 Cyclone). The Denver CPU is so beastly that it even beats out some PC-class chips (the low-end Haswell Celeron), while consuming much less power. The secret behind this performance boost is a novel technology called Dynamic Code Optimization.

First, let’s run through the Denver/K1 hardware specs. Denver is a 7-way superscalar in-order CPU core (it can execute seven instructions per clock cycle, which is a lot). Each core has 128KB of L1 instruction cache and 64KB of L1 data cache (both very large compared to the usual 32KB/32KB for a 32-bit ARM chip). In the 64-bit Tegra K1, there will be two Denver CPU cores sharing 2MB of L2 cache. The Denver CPU is expected to ship at 2.5GHz, which is rather fast. The Denver Tegra K1 will be pin-compatible with the current Cortex-A15 Tegra K1, making it easier for hardware makers to embrace the new chip when it arrives later this year.

Denver cores are very large, and there won’t be a quad-core variant (at least not in the short term). While the lack of a quad-core part might make for an uphill marketing battle — people still love their higher core counts — in reality a dual-core CPU should be more than capable of handling most workloads (most apps and games are still bottlenecked by single-threaded performance).

Dynamic Code Optimization

Nvidia says Denver is explicitly designed to have “the highest single-core CPU throughput” — and while its 7-way superscalar design definitely accounts for a lot of that performance, the other trick up its sleeve is Dynamic Code Optimization.

Back in January we speculated that Nvidia might try something like Transmeta’s x86 emulation. As it turns out, Nvidia did try something rather weird and funky with the Denver core — but it has very little to do with x86. While Denver is an in-order chip, most modern CPU cores use out-of-order execution (OoOE) — where code is executed whenever it’s ready, rather than in the exact order laid out by the programmer — to increase the number of instructions executed per cycle. This works well, but it requires a significant chunk of silicon to actually perform the OoOE magic — and that silicon takes up die space and power. Instead, Denver uses Dynamic Code Optimization, which stores the most regularly executed code in a special 128MB main-memory Optimization Cache. This way, instead of having to fetch and decode instructions and data multiple times, the pre-digested and ready-to-go code can be run directly from the Optimization Cache. It’s a similar approach to how hybrid hard drives or SSD accelerators work.

While the actual process of storing code in the cache requires some additional processing power, Nvidia says the long-term performance gains are huge. In the case of Android tablets, where you probably only run a handful of apps, you could be running almost entirely from this optimized cache — in which case, the performance gains will be really quite significant.
In theory…

On paper, and in Nvidia’s internal benchmarks, the 64-bit Denver CPU core sounds rather exciting. The CPU handily beats every other ARM CPU on the market by some margin, and even does pretty well against the dual-core Celeron 2955U (Haswell). The 192-core Kepler GPU will probably be the fastest GPU on the market by some margin, too.

Perhaps most excitingly, though, the Denver Tegra K1 also looks like it will kick ass in reality, too. It’s important to note that the 64-bit Tegra K1 is the SoC of choice for Google’s upcoming Android L. Android L, which is due out towards the end of the year, is being developed specifically for the 64-bit Tegra K1 — and all signs point to the Nexus 9 tablet, which is probably an Android L launch device, having Nvidia’s 64-bit chip inside, too.

In the meantime, we’re still waiting to see what Qualcomm has planned for its ARMv8 64-bit CPU core. Apple caught everyone off guard when it was first to market with its custom-designed ARMv8 Cyclone CPU — and now Nvidia is proudly touting that it’ll be the first to market with a 64-bit Android CPU. It will be interesting to watch Qualcomm play catch-up when its 64-bit chip is finally ready to go (probably in 2015).

http://www.extremetech.com/com...-for-first-64-bit-android-chip

Dijagrami su na linku. Od opšte ofanzive na x86 knjiga spade na android tablete. Primetite da su najave spale sa 8 na 2 jezgra. Sa upokojavanjem Windows-a RT, nećemo čak moći ni da imamo situacije ekvivalentne onim gde Alfa sa Win NT-om ne može da učita Fotošop jer dotični radi samo sa 386 mašinama a ostale aplikacije se po brzini utrkuju sa 286-icama.

Haswell performanse... za Android tablete? Wtf :)

Intel za koju nedelju izbacuje >pravi< Haswell, tj. EP Xeon-e :-) Sa 18 jezgara i 145W TDP-om (8W TDP po >vrlo brzom< jezgru) cenim da je prilicno brzi od onih Y "mobilnih" Haswell gov*nca, sa kojima se NVIDIA "Denver" verovatno poredi :-)))

Sa 4S Haswell EP-ovima ce biti moguce napakovati 72 Xeon jezgra u jedan server - svako ko ima dodira sa tom industrijom zna da je kljucna stvar >density< a ne pojedinacni CPU TDP (koji je, pritom, daleko manja stavka od ostalih troskova koji idu).

Sa druge strane, ARM u CUDA GPU-ovima je daleko opasnija stvar, sa tim dodatkom mogu komotno da eliminisu Intel Xeon-e koji trenutno standardno "drive-uju" Tesla / Quadro GPU-ove za HPC aplikacije i zamene ih nekim daleko jeftinijim CPU-om koji ima brz PCIe. Takodje se eliminse i gomila x86 koda zato sto ce taj ARM moci da "prihvati" serijske task-ove za koje CUDA jezgra nisu dobra, sve u svemu ta NVIDIA-ina strategija je prilicno dobra i moze ostetiti Intel puno na HPC polju.

Sa druge strane, Intel je obavezan da stavlja standardni PCIe u svoje procesore do 2015-te godine (anti-trust). Posle toga, all bets are off, ne bi me cudilo da u Xeon EP "tock-u" posle Skylake-a izbace standardni PCIe (ostavi neki pateticni PCIe 2.0 na I/O hubu, useless za NVIDIA-ine GPUove) i furaju svoj Xeon Phi za HPC task-ove pa cemo ponovo imati ko-procesore.

Cenim da ce HPC trziste biti jako zanimljivo u sledecih par godina. Obicni serveri, nisam bas siguran, nesto mi se cini da cemo i za 5 godina i dalje imati Intel crne kutije po server room-ovima :-)

Potpis

DigiCortex (ex. SpikeFun) - Cortical Neural Network Simulator:
http://www.digicortex.net/node/1 Videos: http://www.digicortex.net/node/17 Gallery: http://www.digicortex.net/node/25
PowerMonkey - Redyce CPU Power Waste and gain performance! - https://github.com/psyq321/PowerMonkey

Vidim da se dosta potencira ovo "64-bit". Koliki se benefit u realnosti od toga uopste moze ocekivati? Citao sam skoro blog post nekog lika koji je objasnjavao zasto se ne moze (problem je inzenjerske prirode) ocekivati rast RAM memorije u mobile urdjajima istom dinamikom kao u PC svetu koji je ucinio da se 64 bita moze potpuno utilizovati.

Neće bar ne što se tiče Nvidia-e. Ovo mi je promaklo.

Nvidia abandons 64-bit Denver chip for servers

Nvidia has cancelled plans to develop a 64-bit CPU processor for servers, three years after it said it would build such a chip.

“That’s not something we’re doing today,” said Ian Buck, vice president of accelerated computing at Nvidia.

The company is instead focusing its latest 64-bit Tegra chips on mobile and embedded devices, Buck said.

Nvidia in 2011 announced a 64-bit ARM-based chip called Project Denver, which would also go into mobile devices, PCs and servers. The company said it was looking to pair Tegra chips with GPUs in servers, but those plans haven’t seen the light of day. The latest Tegra K1 chip, which includes a 64-bit CPU and will ship later this year, is for smartphones, tablets, cars and other products.

The company has said the mobile K1 chips could make it to microservers, but the company won’t develop a specialized server CPU. The company’s exit leaves a field of only four ARM server chip makers: Advanced Micro Devices, AppliedMicro, Broadcom and Cavium.

Nvidia, known for its graphics cards, is instead focusing on building Tesla high-performance GPUs for ARM servers.

“We should continue to focus on building great GPUs for them,” Buck said.

Interest in ARM servers is growing for web hosting and cloud computing, and also as a low-power alternative to the dominant x86 servers. But there is also skepticism around ARM servers, which are still largely being tested. The platform software stack has not matured yet, and ARM server pioneer Calxeda shut down late last year following lack of funds, raising questions about the viability of such products.

AMD is betting its future in servers on ARM chips, while AppliedMicro is taking a cautious approach, saying its first few chips will help the company gauge interest in ARM servers.

Only a handful of 32-bit ARM servers are available today from vendors like Boston Ltd. and Mitac. The first 64-bit ARM servers from Cirrascale and E4 Engineering will ship later this year, with servers running AppliedMicro’s 64-bit processors and Nvidia’s Tesla graphics cards. The companies did not provide pricing information.

Servers with Nvidia’s Tegra chips have been developed in the past. Barcelona Supercomputing Center in Spain built supercomputers based on Tegra 2 and 3 to test the energy efficiency of ARM chips. However, that effort was supplanted by a supercomputer running ARM-based Exynos smartphone chips from Samsung.

http://www.pcworld.com/article...t-denver-chip-for-servers.html

Tacna CPU.