In terminology strategists "tick-tock", Broadwell is "tic": the vendor used architecture of Haswell in 14-nanometer execution. "Tic" does not assume revolutionary changes. Analysts consider that in this phase Intel dopilivat known technologies, and improvements will be cosmetic. According to the Broadwell specification provides the lowered energy consumption and overall performance, comparable to processors of the previous generation.
Let's consider three new server Intel Xeon E3-1200 v4 processors on LGA 1150 socket.
|Processor||Xeon E3-1285 v4||Xeon E3-1285L v4||Xeon E3-1265L v4|
|Clock frequency||3,5 GHz||3,4 GHz||2,3 GHz|
|The maximum frequency in turbo mode||3,8 GHz||3,8 GHz||3,3 GHz|
|The unblocked multiplier||No||No||No|
|TDP||95 W||65 W||35 W|
|HD Graphics||Iris Pro P6300||Iris Pro P6300||Iris Pro P6300|
|Frequency of graphic kernel||1150 MHz||1150 MHz||1050 MHz|
|L3 cache||6 Mb||6 Mb||6 Mb|
|The integrated eDRAM||128 Mb||128 Mb||128 Mb|
|Support of DDR3||1333/1600/1866||1333/1600/1866||1333/1600/1866|
|Expansions of set of instructions||AVX 2.0||AVX 2.0||AVX 2.0|
Volume cache memory of the third level is cut down on 2 Mb, that is on quarter in comparison with the previous Xeon models. As notices www.anandtech.com, loss of 2 Mb of L3 is insignificant, considering emergence of eDRAM of 128 MB in L4.
What is eDRAM?
At the beginning of 2007 Intel and IBM together with AMD, Toshiba and Sony have almost at the same time declared technology which will allow to accelerate operation of processors twice. It was for this purpose offered to use memory of eDRAM (embedded DRAM) which will be applied together with the "silicon on insulator" technology (SOI, Silicon-on-Insulator). In IBM have declared that it will allow to accelerate significantly cache memory, three times to increase its density and five times to reduce power consumption in the mode of inaction of the processor. Then it was going to use cache memory with a capacity of 24 Mb.
Technology embedded DRAM — is DRAM which is built in ("embedded") the processor unlike SDRAM. eDRAM demands voltage of 1 V that is much lower than power supply of conventional memory. In practice it means that eDRAM consumes less energy, providing comparable high-speed performance. Cells of eDRAM of the smaller size — 0.108um ² for SRAM and 0.029um ² for eDRAM — on the chip of the same size will get 10 times more slower storage cells.
Intel not only has put known Haswell on 14-nanometer rails. In Broadwell processors the vendor used possibilities of eDRAM so fully as far as it is possible today. The vendor declares about 25% improvement of such parameters, as energy consumption and heat release.
The crystal of eDRAM Crystalwell with a capacity of 128 MB plays role cache memory of the fourth level. Haswell had cache of similar volume, but only in the top E7 models. At this eDRAM it can be used by both graphic kernel, and computing kernels of the processor when processing large volumes of data — for example, textures. According to the statement of the vendor, Crystalwell provides peak throughput at the level of 51,2 Gbytes / with in each party (102,4 Gbytes / with totally).
According to the tests which are carried out 3Dnews.ru, latency based on eDRAM L4 cache makes 55 clock periods, and flow capacity — is approximately twice higher in comparison with two-channel DDR3-1866 SDRAM.
Crystalwell allows the processor to overcome relative restrictions of low flow capacity of system memory and to show the best high-speed performance in tasks of processing of HD video and in mathematical operations.
According to own Intel tests (Supermicro SuperServer was used), when using Broadwell simultaneous processing 4368 HD видеопотоков is possible. The same indicator when using Haswell processors makes 3120. The increase in productivity by 40% is available. All Supermicro motherboards of the X10 series poderzhivat v4 after the BIOS updating.
Video kernel of Iris Pro Graphics
In new line of processors the graphic kernel of Iris Pro Graphics which "facilitated" version could be met only in mobile Haswell processors earlier is used. Now Iris Pro Graphics P6300 has configuration of GT3e and 48 actuation mechanisms (EU, execution units) against 40 EU in graphic kernels of Iris Pro Graphics 5100 and 5200 at Haswell.
Use of Iris Pro Graphics P6300 is also essential difference of new Xeon from Xeon E3-1200 v3 on Broadwell: in last ruler less advanced Intel HD Graphics P4600 and Intel HD Graphics P4700 were used.
Clear advantage of the considered processors is that they are installed in LGA1150 slot: owners of motherboards on the basis of set of system logic of Intel Z97 and Intel H97 can easily apgreydit system. Formally new Xeon are compatible to motherboards on the basis of set of system logic of the eighth series on LGA1150 socket.
The users who are updated with Haswell on Broadwell can make it rather cheap now. So, if the recommended Intel prices of E3-1285L v3 make $774, E3-1285L v4 cost — $445.
The impossibility of acceleration of these processors can appear shortcoming for some users. However, this Xeon series — server solution, and the vendor offers fans of overkloking other processors of new ruler.
In a number of the tests which are carried out by www.anandtech.com, Intel Broadwell Xeon E3-1200 v4 processors have shown inconsistent results.
At archiving of 2867 total files the best amount has shown E3-1285L v4, having overtaken even more powerful E3-1285 v4.
However in other Xeon tests lagged behind competitors. Converting of series from 170 images by means of the FastStone Image Viewer v.4.9 application has appeared for new Xeon quite complex challenge.
Start of the emulator of Dolphin video game consoles has shown that in working off 3D - scenes of Intel Broadwell Xeon E3-1200 v4 advances only weak processors.
At the same time when converting video of the x264 format in mp4 Broadwell has lagged behind only Intel Core i7 6700K.
We sum up the results
The line of Intel Broadwell Xeon E3-1200 v4 processors thanks to new crystal of memory of eDRAM overtakes old models in solution of hard computing tasks and processing of HD video at reduced values of power and heat release.
For what they are necessary:
1) high-performance graphic stations
2) the selected servers the carrying-out tasks of transcoding and stream broadcasting without use of Tesla GPU. New processors Intel show 40% to productivity on such tasks and allow to lower costs for the hardware of transcoders. Here the comparative table of productivity for Wowza online broadcasting products is provided — what processors for what tasks are suitable.
2) the selected servers the graphic applications allowing to virtualize — for this purpose it was necessary to put Tesla or something similar earlier
We in Hostkey offer in rent the selected servers in Moscow and the Netherlands — new type of processors for virtualization of 3D applications and striminga / processings of video will appear at us at the beginning of 2016.
For similar tasks we recommend to clients to use the multi-core processors Xeon of the senior series, for example the machine in configuration 2kh Xeon E5-2670 v2 with 20 kernels on 2,5ggts in on the basis of supercomputer platform of the T-platform and 32gb memories will cost 23000 rubles a month.
This article is a translation of the original post at habrahabr.ru/post/266519/
If you have any questions regarding the material covered in the article above, please, contact the original author of the post.
If you have any complaints about this article or you want this article to be deleted, please, drop an email here: firstname.lastname@example.org.
We believe that the knowledge, which is available at the most popular Russian IT blog habrahabr.ru, should be accessed by everyone, even though it is poorly translated.
Shared knowledge makes the world better.