@all

It has taken some time - much longer than I had ever thought it would take...

Today (21 jan 2015) we have booted Fedora 20 on our own T4240 uServer board.

- Two years from idea to reality.
- using rev 1 board
- using NFS mounted rootFS

and...

it is confirmed that the energy efficiency of the T4240 is really great. I ran a simple bash script to run 24 threads flat out, and T4240 consumes 27A at 1.0V


see: twitter post

@RonaldGadget

Congratulations!


What about performance, did you tried some benchmarks?

@pavlor

We have a paper at ISSCC in Feb 2015 describing the benchmark results. I cannot disclose information here and now as the paper would be withdrawn by the conference organizers if I do that.

What I can say: when I compare our node board to an Intel Xeon E3 1230Lv3 (no turbo): we are substantially faster comparing aggregate of all threads while using less energy.

Wait until Feb 24th for details: ISSCC advance program and search for microserver.

Last edited by RonaldGadget on 21-Jan-2015 at 09:21 PM.

@RonaldGadget

From the ISSCC program leaflet:

Energy-Efficient Microserver Based on a 12-Core 1.8GHz
188K-CoreMark 28nm Bulk CMOS 64b SoC for Big-Data
Applications with 159GB/s/L Memory Bandwidth System Density
R. Luijten1, D. Pham2, R. Clauberg1, M. Cossale1, H. N. Nguyen2, M. Pandya2
1 IBM Research, Rüschlikon, Switzerland
2 Freescale Semiconductor, Austin, TX

Last edited by virgolamobile on 22-Jan-2015 at 11:43 AM.
Last edited by virgolamobile on 22-Jan-2015 at 11:42 AM.

@virgolamobile

Thanks.
Quote:

Few questions 1) Is it your latest Microserver based on T4240 able to deal with a GPU on a PCI-E? 2) Did you thought on scaling down your solution to build a desktop? 3) Have you been approached yet by any of the Amiga related companies (e.g. A-Eon or Hyperion Entertainment)?

1 - I do not know this - T4240 does have PCIe interface. The challenge likely will be libraries.
2 - No, we are building this as a server. For a desktop you would want integrated graphics.
3 - No.

Here is a video of booting Fedora on the DOME T4240ZMS.

Ronald

@RonaldGadget

Quote:

What I can say: when I compare our node board to an Intel Xeon E3 1230Lv3 (no turbo): we are substantially faster comparing aggregate of all threads while using less energy.

8 threads on Xeon @25 W, 24 on T4240 @ 60 W, it should be faster... Unfortunately for Freescale, Intel also offers, among others, a 65 W 12 core Xeon.

Last edited by WolfToTheMoon on 22-Jan-2015 at 08:54 PM.

_________________

@RonaldGadget
Is it correct, that you use a standard dimm socket connector?

Are all Serd-Lanes of the T4240 are routed to this connector?
If yes it would be possible to make a Adapter to a standard PCI-E Slot(s) and use a PCIe Graphic card.

What Powersupply is needed for the Microserver?

After watching this video https://www.youtube.com/watch?v=N638UXr2pC0&spfreload=10
I´ve understood why you have choosen the T4240

Rudi

@RonaldGadget

Noone I know want integrated graphics in a desktop. We want an accessible PCIe slot, nothing more

_________________
This weeks pet peeve:
Using "voltage" instead of "potential", which leads to inventing new words like "amperage" instead of "current" (I, measured in A) or possible "charge" (amperehours, Ah or Coulomb, C). Sometimes I don't even know what people mean.

@olegil

Personally? Id prefer if GPUs would work in single lane slots. Then on an 8 lane chip, we could have all 8 sockets available.

Bandwidth?

A single Gen 3 Express lane can handle a 10 Gigabit connection? And youd need a major SSD RAID array ona single lane to even get close to saturating it, on moving large contiguous files. As soo as you start to frag, IO access variations slow the overall speed down.

My favourite one was a graph showing a SATA 3 SSD accessing completely fragged 4k sectors running as slow as 2 Megabytes a Second. And a normal hard drive at well under 1 Meg. Well, the hard drive I can understand, given at most 120 accesses, head moves at 4 k a second, the SSD shouldve been 50 MEg or so, still far less than ATA100

What would I do with 8 PCI Express lanes, if I had the money? put 8 X2GPUs un there, at almost 4000 processing cores apiece. Then use their F/D/GMAC or whatever new patent required name is these days, to implement things like models of Guttenburg Neural Nets.

To recreate the synaptic content of a human brain in todays technology, would require only 2 cubic inches of 256 Gb microSD cards. Which would run roughly 1 million times slower than the brain. And yet this thing can now play poker.

Which beats me, as a human.

Just what are all those million fold increase in cycles getting used for, compared to non hardware FFT/DCT based A500s?

_________________
The older and more respected a scientist is, the longer it takes to prove him wrong.

@Rudi

For the current version (rev 1 and rev 1b) we use the a DIMM connector, but with our own signal definition. Physically it will fit in a 240 pin DDR3 slot, but turning on power would be a really bad idea.

For signal integrity reasons we will replace the DIMM connector with a high frequency DIMM - like connector to enable going from 10 GbE XAUI to 10GbE KR (rev 2). This reduces carrier board wiring for networking by factor of 4.

We do not bring out the PCI serdes lanes out as there are not enough pins on the connector.

Across the DIMM connector, we supply 3v3, 1v8 and DRAM 1v35/0v67 to the T4240, and the copper plate brings in 12V. This is optimized for a system with many node boards in a carrier with common power modules for the voltages across the connector.

Our Mini base board supplies these voltages and only needs 12V input.

Ronald

@WolfToTheMoon

Quote:

8 threads on Xeon @25 W, 24 on T4240 @ 60 W, it should be faster... Unfortunately for Freescale, Intel also offers, among others, a 65 W 12 core Xeon.

OK: - The T4240 does not use 60W, when run at 85c. Our T4240 node board uses substantial less energy than the E31230Lv3 node board... wait for the paper...

Ronald

Quote:

RonaldGadget wrote: OK: - The T4240 does not use 60W, when run at 85c. Our T4240 node board uses substantial less energy than the E31230Lv3 node board... wait for the paper...

If P=I*E and what you said earlier is true...

Quote:

RonaldGadget wrote: it is confirmed that the energy efficiency of the T4240 is really great. I ran a simple bash script to run 24 threads flat out, and T4240 consumes 27A at 1.0V

P = I*E = 27A*1.0V = 27W near max load

Maybe my electronics from high school are a little rusty though. These are Impressive results if true.

Last edited by matthey on 23-Jan-2015 at 09:11 PM.
Last edited by matthey on 23-Jan-2015 at 09:09 PM.

@matthey

Quote:

P = I*E = 27A*1.0V = 27W near max load

NOTE: 27W at the 1V supply. There are other supplies, but they don't use as much. We can't measure those for the moment, but are expected in handful of Watt range.

We all need to dust off the rust from time to time.

@RonaldGadget

[for some reason I could not submit anything anymore, the system said my address is blacklisted]
seems to work now again.
See EEtimes coverage!

Last edited by RonaldGadget on 27-Feb-2015 at 05:33 PM.
Last edited by RonaldGadget on 27-Feb-2015 at 05:32 PM.

@RonaldGadget

Whoa! Amazing!

@virgolamobile
@RonaldGadget

Thanks!


Edit: For comparison Core i5-2500K (in 4 years old computer od my brother) scores 45.0 (42.8 base) SpecInt2006 (CINT2006) and 133 (126 base) CINT2006 Rates.

Last edited by pavlor on 26-Feb-2015 at 07:19 PM.
Last edited by pavlor on 26-Feb-2015 at 07:17 PM.

@RonaldGadget, virgolamobile: you're comparing apples with oranges. Such Xeon is tailored toward high single core/thread performance. That's why it shows triple the performance in CINT-base 1-thread.

An Intel micro-architecture which can be similar to the T4240 is Avaton (and successors) or Xeon Phi for HPC scenarios (e.g.: massive double precision performance). They embed simple cores, with low power-consumption .

Ronald interviewed at the Freescale YouTube channel.

Dear Ronald, talking about the Xeon systems, like you did at the end of the video, maybe you have to take a look at the last born: Intel® Xeon® Processor D

Following there are some numbers, so you can compare it with the T4240.

Intel® Xeon® Processor D (8 cores, 1.9 GHz, 45 W, ES2): SPECint*_rate_base2006 (16 copy) = 268

But I report also some data about the Avoton family, which is comparable to the T4240 one, as I state before (sorry for the typo).
Intel® Atom™ processor C2750 (8 cores, 2.4 GHz, 20 W): SPECint_rate_base2006 (8 copy) = 102

Pay attention to the number of cores and threads used to achieve the results: far below the T4240. Which means that the latter's microarchitetture isn't that efficient (which is reflected also by the lower IPC).

P.S. Xeon-D's tests are preliminary and subject to increase.
P.P.S. Avotons have no AVX & AVX2 SIMD unit. Only AES extensions are available.