DailyTech - Mercury Introduces Cell PCIe Plug-in Accelerator

Submit News

Hardware Mercury Introduces Cell PCIe Plug-in Accelerator
Sven Olsen (Blog) - July 31, 2006 2:07 PM

26 comment(s) - last by SniperWulf.. on Aug 2 at 10:48 AM

The Cell gets ready for the desktop, workstation and server market

Mercury Computer Systems has announced its Cell Accelerator Board, or CAB. The CAB is based on the multi-billion dollar endeavor that is Cell BE, made famous by the Playstation 3. The CAB is a PCIe plug-in adaptor that fits in standard motherboards and cases today.

Mercury claims the CAB provides approximately 180 GFLOPS -- though Mercury claims the CAB is specific to video and render processing. Currently, the hardware is supported via Linux using the Mercury MultiCore Plus Advantage software libraries. The card sports a 2.8GHz Cell BE processor, 1GB of XDRAM and 4GB of SDRAM.

Mercury representatives claim the CAB is slated to retail for around $7,000. Although this price seems high, it should be no surprise as IBM recently claimed that the company is lucky to get 20% yeilds on the Cell processor.

"We have been actively working with customers to migrate data-intensive applications to the Cell BE processor since late last year," said Randy Dean, Vice President, Business and Technology Development at Mercury. "With the CAB, customers can achieve supercomputer-like performance right inside their workstation."

The Cell used in the Mercury is based on the 90nm design process, but the Cell group (IBM, Toshiba and Sony) have already announced the Cell processor is moving to the 65nm process by next year. The idea of plug-in accelerators has certainly picked up some steam in the last three months with AMD's backing. Clearspeed also introduced a co-processor earlier this year for PCI-X acceleration, but Mercury's CAB is the first to use PCIe and the first to use Cell.

Mercury has a long-standing relationship with Cell. The company was also the first to introduce Cell rack mount servers earlier this year.

Comments

Threshold

Username
Password
remember me

This article is over a month old, voting and posting comments is disabled

How On earth?

By creathir on 7/31/2006 2:50:34 PM , Rating: 2

Can Sony afford to practically GIVE the Cell away? At a $7,000 MSRP for a single card... makes ya wonder what the deal is? Is this company totally overcharging for this product? Or is Sony really "takin' one for the team" by selling its PS3 so low?
- Creathir

RE: How On earth?

By Dfere on 7/31/2006 2:55:29 PM , Rating: 2

I don't think this means the Cell processor itself needs to be as expensive.. not sure what is in this monster card. Kinda impressive strategy, Cell is relatively cheap, but a license to stick in a $7,000 video accelerator... Priceless!

BTW, most companies take one for the team so long as they can either keep impressing the stock analysts somehow, or smooth earnings despite cash outflows...... (yes, it takes a cynic to be an accountant).

Parent

RE: How On earth?

By rrsurfer1 on 7/31/2006 3:26:01 PM , Rating: 2

Yea I'm wondering if these are the same chips. At first glance they look to be - both are 8-core Cells @ 2.8 Ghz. It may very well be that the $8,000 price tag is due to the software which actually makes the accelerator board useful.

Parent

RE: How On earth?

By rrsurfer1 on 7/31/2006 3:30:39 PM , Rating: 3

Ahh... mised the on-board memory:

Memory
Memory: 1 GB XDR DRAM, 2 channels each, 512 MB
ECC support: single-bit correct; double-bit detect
Flash: 2x16 MB
DDR2: 4 GB

That's a fair amount of memory with some good ECC. Plus the software:

Mercury MultiCore Plus™ Software
Linux®: Linux with BE-SPU extensions
Windows® XP or Linux

So the price is probably about right if you need the speed for something.

Parent

RE: How On earth?

By Araemo on 8/1/2006 12:33:48 PM , Rating: 2

Well, the PS3 cell is technically a 7-core CPU. 1 core is disabled for 'yield' reasons. If this company is holding out and taking only the full '8 core' cells, that could easily explain the added cost at this point.

Parent

RE: How On earth?

By s12033722 on 7/31/2006 3:34:09 PM , Rating: 2

While Sony is losing money on the PS3, keep in mind the economies of scale involved here. The PS3 will sell millions of units, so the cost of development can be spread out over millions of units. With something like this board, it will be a fairly low volume product. It will likely only sell a few thousand units, so the cost of development is proportionally much higher per board than with something like the PS3. So, take the costs of the hardware (cell, PCB, power regulators, caps, ...), the costs of the development (both hardware and software), marketing costs, and a little profit for the company, and you hit $7k really quick.

Parent

RE: How On earth?

By lemonadesoda on 7/31/2006 6:06:20 PM , Rating: 2

Yes in theory, but no in practice.

Take a look at all the single board computers for embedded devices available on the market. Their cost is typically 3-5x the cost of a high volume mainboard.

I'll accept a 10x cost for a real rare value-add item. So USD 500 would be max. for the PCIe processing board, excluing Cell. However, I think it would be less.

The cell processor, even an 8 core one, is no way near EUR 4500. The http://www.theinquirer.net/default.aspx?article=27... suggests EUR 150 for the cell.

Hence hardware is circa EUR 300-650 tops. I'll leave the rest of the explanations up to you...

Parent

RE: How On earth?

By TomZ on 7/31/2006 10:32:45 PM , Rating: 2

quote:
I'll leave the rest of the explanations up to you...

Product pricing generally depends on the value that the product provides to the customer, relative to other possible solutions, not the cost to produce the product. Therefore, it is not hard to believe that a product with a manufacturing cost less than $1000 would sell for $8000, especially if the product has a high development cost and if it includes valuable software and support.

Parent

Actual Quote

By Phynaz on 7/31/2006 4:20:20 PM , Rating: 1

The quote is that a chip like cell yeilds 10 to 20 percent, while cell yeilds twice that due to redundant logic.

RE: Actual Quote

By rrsurfer1 on 7/31/2006 4:24:10 PM , Rating: 1

I'm not exactly sure where your getting your info or what "redundant logic" means in that context. However, what I would guess, is that they get 20% to 40% yield on the cel with one of the 8 cores disabled. This is NOT redundent logic, unless you also beleive that one of the cores of a dual core CPU is simply redundant. Bottom line is the yield sucks and disabling 1/8 of the chip isn't a good solution regardless of the wording.

Parent

RE: Actual Quote

By segagenesis on 7/31/2006 5:04:15 PM , Rating: 2

While the "redundant logic" claim is a bit out there... in the older days they did actually do what you just described, disabling parts of the chip because of lower yields until processes improved. Lookup anything about undocumented opcodes for old 8-bit processors like the 8080/8085/6502/6809 and especially the Z80, most of these contained logic for extra opcodes but were never officially part of the instruction set as they failed often in early batches being at the edges of the die. Later on those with the know how discovered the existance of these opcodes and they became more usable with improved chipmaking, but still being undocumented it was risky using in a production system.

Sucky yields is not a new thing with any new microprocessor but I would agree that 20% or less in todays market is rather poor, hence the cost.

Parent

RE: Actual Quote

By TomZ on 7/31/2006 10:22:45 PM , Rating: 2

quote:
Lookup anything about undocumented opcodes for old 8-bit processors like the 8080/8085/6502/6809 and especially the Z80, most of these contained logic for extra opcodes but were never officially part of the instruction set as they failed often in early batches being at the edges of the die.

I don't think this is related to yield. How can they place certain opcodes at the edge of the die or wafer? I thought that undocumented opcodes typically happen when the bit pattern, due to side-effects of the design of the opcode decode logic, happens to do something useful. Using google, I couldn't find any web sites discussing undocumented opcodes related to yields - do you have any references for that?

Parent

RE: Actual Quote

By segagenesis on 8/1/2006 8:31:06 AM , Rating: 2

Well I've about read this on and off for the past decade... as far as I know at the time they (competing cpu makers) would try to fit as many opcodes onto the chip as possible to try and one-up the competition, with lesser used stuff near the edges... whatever worked worked and became part of the official instruction set.

This seems to directly reference it at least, you probably had trouble because this and similar sites are not very easy to find http://www.microprocessor.sscc.ru/great/s1.html

Parent

RE: Actual Quote

By masher2

(blog) on 8/1/2006 9:15:59 AM , Rating: 2

This was very common in the past, particularly when manufacturers first began introducing floating point processing. If the extra circuitry worked-- fine. If not, you sold it as an integer-only chip, and the customer used software emulation, or bought a coprocessor. Somewhat later, it was common for large caches. If part of the cache didn't pass testing, you blew an internal fuse, and sold it with a half-cache.

And I have news for you...the technique is going to see a *strong* resurgence in the future. In a decade or two, when processors have 100+ cores on a chip, your average yields for all cores is going to be much less than 10%. Disabling 2-3 cores is barely going to impact performance, and boost yield rates dramatically. I suspect eventually this technique will almost entirely replace the speed-binning in use today. The "best" chips will be the ones with the most working cores, not the ones that run the fastest.

Parent

RE: Actual Quote

By lemonadesoda on 8/1/2006 5:29:40 PM , Rating: 2

I think speed binning will survive. While the technique will work for fusing-out "parallel" structures in the chip, there are certain elements that can't be cut, for example:

1./ The "Master Cell", or main processing element and scheduler

2./ The I/O interconnect with the outside world

3./ Memory Interface controller

4./ Internal bus - linking the master with the subordinate cells

5./ Clock, and interrupt, circuitry

6./ Some cache elements, like primary (level 1) cache

There will always be some speed/volt combinations that work for some and not for others, irrespective of the number of cells in operation.

I think that where we have, lets say for example, 100 cells on chip, then with a "90%" yield per cell, there will be on average 90 working cells, and by application of the binomial theorem, there range of working cells will be so close to 90, e.g. 85-95, and the number of chips outside this range, so rare, that it will hardly be worth binning. Clock/volt will however still be a major differtiation factor.

Parent

RE: Actual Quote

By lemonadesoda on 8/1/2006 5:38:33 PM , Rating: 2

Oh my goodness, I hope you readers could make sense of that! Aweful typos didn't help. Let me provide an explanation by example:

If you toss a coin 10 times, then on average there will be 5 heads, 5 tails. (no bias). There is quite a large possibility that you would get 4, 5 or 6 heads. A smaller possibility that you would get only 3 or 7 heads, and very unlikely, but still possible that you would get only 1 or 2, or 8 or 9 heads. You can forget about 0 or 10 heads. The probabilty of you obtaining 10 heads is 0.5^10=0.098%.

Now consider tossing the coin not 10 times but 100 times. The average is 50. But the likelyhood of getting only "40" or "60" is much much lower than "4" or "6" in the 10 toss example before.

This is probability theory for you. The same applies to cell manufacture process with, e.g., 90% yield per cell. If you only have 3 cells, then the possibility of only 1 or 2 cells working is a relevant calculation, and the performance different quite significant. But if you have 300 cells, the possibility of only 100 or 200 cells working is an irrelevant calculation. It just won't happen.

Parent

Emulation

By Alphafox78 on 7/31/2006 4:04:51 PM , Rating: 2

I can imagine that lots of software will be popping up for this allowing PS3 emulation... lol, ya right!

RE: Emulation

By Master Kenobi

(blog) on 7/31/2006 4:08:35 PM , Rating: 2

Lucky to get 20% Yields? Jeeze, from a manufacturing perspective, that absolutely blows. From a technology perspective, this pretty much guarantees nobody else will try to use Cell other than the PS3, atleast not anytime soon.

Parent

RE: Emulation

By rrsurfer1 on 7/31/2006 4:20:42 PM , Rating: 2

I know. It's a dead end. Sony is just trying to lose money on the PS3. It almost seems that way...

Parent

In the future, Cell will be cheaper

By wingless on 8/1/2006 1:40:09 AM , Rating: 2

In a few years the Cell processor will be easily manufactured (theyll have the bugs worked out by then) and a lot cheaper. We may actually have Cell-powered physics accelerator cards for our home PCs in a few years folks. I think this is a good step to showcase the raw power of this type of processing unit. I remember back in the 80's when you had to buy a MATH CO-PROCESSOR for a few hundred dollars. I woouldnt mind spending a few hundred on a Cell co-proc in 2010. 150Gflops is good anytime.

RE: In the future, Cell will be cheaper

By threepac3 on 8/1/2006 6:41:45 AM , Rating: 2

I guess no buddy noticed that this monster has something like 5~6GB of memory some with error correction. Its clear that most the price comes from something other then the CELL BE.

Parent

RE: In the future, Cell will be cheaper

By Master Kenobi

(blog) on 8/1/2006 7:34:15 AM , Rating: 2

We're not disputing the price of the item, it seems reasonable given the large amount of memory and the processor. We're pointing out that a ~20% yield is terrible for any processor by modern standards. Not to sound like a Fanboy or anything, but I doubt we will see tremendous process improvements until Intel (if they ever do) tries to manufacture a Cell processor.

Given that Intel has as far as I can remember, been the front runner for manufacturing process improvements, its only logical to bet on them to do it again. Granted IBM might be able to get the yields correct if they have enough time and money to throw at it.

Parent

Fix the first link, please...

By blckgrffn on 7/31/2006 3:38:01 PM , Rating: 2

"The Press Release you requested could not be found"

RE: Fix the first link, please...

By rrsurfer1 on 7/31/2006 4:17:44 PM , Rating: 2

Yes. Also, the price is set at $7,999 per board, not $7K.

Parent

Small market

By mindless1 on 8/1/2006 12:48:06 AM , Rating: 2

There may easily be another reason for the lofty price, that it's a MUCH smaller anticipated market. Put together a board like this and the software and see how much you'll have to charge if you only sell 10,000 total.

heh

By SniperWulf on 8/2/2006 10:48:10 AM , Rating: 2

Imagine the folding you could get done with this thing.... too bad it's gonna cost 7 Grand

"Spreading the rumors, it's very easy because the people who write about Apple want that story, and you can claim its credible because you spoke to someone at Apple." -- Investment guru Jim Cramer