IBM Power 7 @ HotChips

I'm not sure if it was mentioned here, but the Power7 seems to have 4 DP FP-units per core. Source: http://arstechnica.com/hardware/new...er7-twice-the-muscle-half-the-transistors.ars
So as it looks like 128 DP-GFlops. Seems a lot!

Ars had a old article about the eDRAM from IBM: http://arstechnica.com/old/content/2007/02/8842.ars
IBM says that it has a 65nm prototype eDRAM running with 1.5ns latency and 2ns random cycle time—speeds that are competitive with current SRAM.

one other point:
The Power7 seems to be a new "breed" of PIM or "Processing in Memory". So maybe IBM will now advance the POWER CPUs in this direction. The large eDRAM cache could point in this direction.
Link to a zettaflops presentation about PIM: http://www.zettaflops.org/PES/3-AdvArch-Kogge.pdf
 
Last edited by a moderator:
I strongly suspect that IBM will use the POWER7 as the basis for the PPE' (<-- note the "prime") in the PowerXCell 32ii and 32iv.

So you think they will even be made?:rolleyes: If so, what do you think will be the differences between PowerXCell 32ii and 32iv?
 
One of the first major application of POWER7 is the NCSA Blue Waters supercomputer, which is expected to have more than 200,000 cores and deliver sustained performance of over 1 PFLOPS.

That's some serious number crunching there. When is the supercomputer due to be completed?
 
That's some serious number crunching there. When is the supercomputer due to be completed?

It's projected to come online in 2011. Blue Waters is for open scientific research and not for nuclear weapons. It'd be interesting to see how efficient it is compared to BlueGene based Roadrunner.
 
So you think they will even be made?:rolleyes: If so, what do you think will be the differences between PowerXCell 32ii and 32iv?

I think that they will be made, and according to IBM's roadmap, the 'ii' has two PPEs, and the 'iv' has 4 PPEs. (All of the PPE-prime variant.)

Cell has already gained a lot of traction in high-performance computing (HPC) and high-performance embedded computing (HPEC). Once the local-store architecture hurdle was jumped, the primary complaint seems to be that the PPE just isn't up to managing eight SPEs. Hence, IBM's decision to use Opteron nodes to coordinate PowerXCell blades in Roadrunner. If the PPE' is significantly better at executing single-threaded code, then the Opterons can go away, and you could have an all-PowerXCell cluster.
 
I think that they will be made, and according to IBM's roadmap, the 'ii' has two PPEs, and the 'iv' has 4 PPEs. (All of the PPE-prime variant.)

Cell has already gained a lot of traction in high-performance computing (HPC) and high-performance embedded computing (HPEC). Once the local-store architecture hurdle was jumped, the primary complaint seems to be that the PPE just isn't up to managing eight SPEs. Hence, IBM's decision to use Opteron nodes to coordinate PowerXCell blades in Roadrunner. If the PPE' is significantly better at executing single-threaded code, then the Opterons can go away, and you could have an all-PowerXCell cluster.

Do you have any roadmaps (pdf's etc., from IBM proper) that you could share with us that say that new chips will be built (from 2009 I mean)? If indeed they planned to make it, they would have been talking about it and people at B3D would have heard.

I am pretty sure that the Cell derivatives line is dead. I dont think that the HPC market (specifically, top500) is big enough to sustain even a single chip, let alone two?
 
What podcast is that? I would like to listen.

The B3D podcast Arwin and I made 2 beta episodes of, and then had to put on standby as I simply didn't have the time for the show anymore, sadly.

You can still grab the episode here, discussion is here.

Hotchips discussion is the first ~30 minutes of the show. Blue Waters comes @ about the 2:11 marker. Rather brief outline of the project, no special insight on my part, I must say.
 
A 3-core adaptation (Even with 3-core, the hardware thread count doubles) plus removal of MCM communication and just 9 MB of eDRAM would keep a POWER7 based XCPU3 under 200 mm2.



Power6 has in-order cores like the current Xbox 360 CPU. They should consume less power and run cooler. The next-Xbox is probably going to have a massive GPU from ATI that will take up most of the power/heat budget.


I doubt Power7 is going to be a part of the next-Xbox.
 
Power6 has in-order cores like the current Xbox 360 CPU. They should consume less power and run cooler. The next-Xbox is probably going to have a massive GPU from ATI that will take up most of the power/heat budget.


I doubt Power7 is going to be a part of the next-Xbox.

I imagine AMD will try to sell MSFT a processor along w/ the GPU. Most likely, it'll be integrated to the GFX core which offers AMD extra margins and MSFT some savings over buying the two separately from different sources.
 
I imagine AMD will try to sell MSFT a processor along w/ the GPU. Most likely, it'll be integrated to the GFX core which offers AMD extra margins and MSFT some savings over buying the two separately from different sources.
It would makes sense, between what MSFT stand for?
 
So, with the process alliance thingie that IBM & AMD are in & this being 45nm SOI, will that mean AMD will get access to the eDRAM tech too?

I imagine even the Phenom II architecture would get a good boost out of 16MB of L3.
I would definitely hope to see it in Bulldozer at the least.
 
I dont think AMD will come out with edram anytime soon. IBM and AMD are competitors in esrver market after all and IBM isn't gonna give up on one of their competitive advantages.
 
I dont think AMD will come out with edram anytime soon. IBM and AMD are competitors in esrver market after all and IBM isn't gonna give up on one of their competitive advantages.

I would argue that the majority of AMD and IBM's highend server markets don't really overlap.

Now if AMD had a chip like Beckton, they would; but as it stands now I don't think they'd ever have the money to invest into an RAS heavy chip that's extremely fast too to be financially viable.

I suppose AMD should be able to license IBM's EDRAM tech, but I think the problem is more of implementation. Power7 is 45 (iirc) and out next year (if it weren't zomghueg and/or a professional product, probably faster). Bulldozer the year after. Refinements and codevelopment?
 
stolen from..ähh seen on aceshardware-forum ( http://aceshardware.freeforums.org/power-7-t891-30.html ) :

Power7 will be introduced in 7 days:

http://www-05.ibm.com/de/events/power7/index.html

The per-core per MHz performance seem to be far greater than Power6 [from page8 of the presentation]:

Chip Int/fp [specint_Rate / specfp_rate]
2 core 5.0Ghz POWER6: 68/68
8 core 4.0Ghz POWER7: 383/442
4 core 2.93Ghz Xeon 5570: 130/100
2x4core 2.93Ghz Xeon 5570: 259/198

Presentation: http://www.power.org/events/powercon09/taiwan09/IBM_Overview_POWER7.pdf
 
Back
Top