Deneb can't, but it seems Thuban will, to some extend. The 1090T has 6 cores at 3.2 GHz and still stays within a 125W TDP. Thuban seems to be quite a step forward in power efficiency compared to Deneb. In applications that scale well with the 2 additional cores, it could actually match at least Bloomfield in perf./W.
Fusion die-shot - 2009 Analyst Day
- Thread starter AnarchX
- Start date
Would it be possible to drop a fusion thing on a PCIe card?
Low profile? (To put 20 of them on a 2U chassis)
Would TDP be a problem?
Would it make sense in a HPC environment compared to sandy bridge + GPUs?
Is fusion double or triple channel?
Low profile? (To put 20 of them on a 2U chassis)
Would TDP be a problem?
Would it make sense in a HPC environment compared to sandy bridge + GPUs?
- On perf/W?
- On software flexibility (x86 + OpenCL nodelets/bladelets)?
- On perf/$?
Is fusion double or triple channel?
Why you want this?
You can simply put fusions into 4P motherbroad if the chip has enough HT links...
Last edited by a moderator:
It's just a silly idea that crossed my mind.
Price: 4P motherboards are expensive. Professional grade blade chassis are quite expensive. A PCIe backplane (or an enthusiast board with excessive PCIe slots) should be cheaper. A PCIe card with just NIC+fusion+DRAMs could be cheaper, booting off the mainboard, or from a small amount of flash on the card itself.
Density: I remember 20xPCI backplanes, back in the turn of the century. I assume similar PCIe products are available today. A 4U el-cheapo chassis with such a backplane could provide a slightly higher density than 4X 1U 4P racks.
Flexibility: a 1xPCIe product could be a simple add-on HPC product for desk-side compute nodes, add 4 cards to your workstation PC, increase your compute power.
And: It's cool!
(but not cold, at guesstimated minimum 100W per card it's 2kW/4U)How much could such a card cost? Cheapest version: 1xPCIe PCB, dual 1 Gbit ethernet, unpopulated dual DIMM slots, one fusion socket, boot off windows/linux host or NIC. €200 for a 10k run?
Bobcat will be bigger and faster than Atom; It will be competitor for both Atom and low voltage core2's. (And will not be competitor to cell phone versions of Atom)
So it will be a very good cpu for real 13" notebooks, not just for netbooks.
Clock speeds should increase to well over 4 GHz when "turbo" mode is used.
Llano will propably have 240 shaders(based on the die images). Current AMD integrated chipsets have 40 shaders, and the performance is about equal, or maybe slightly worse than the new intel clarkdale's GPU.
So even if intel doubles it's GFX performance for sandy bridge, AMD should have some 2.5x advantage.
Last edited by a moderator:
Isn't it way too optimistic to expect 5x increase over current fastest IGPs with only 240SPs? I mean look at this:
5450 is ~50% faster than IGPs
http://anandtech.com/show/2972/the-...-pentium-g6950-core-i5-650-660-670-reviewed/2
5570 is ~3x over 5450
http://anandtech.com/show/2935/5
Which would correspond to 4.5x, but that's with 400SPs, and somewhat more bandwidth than Dual Channel DDR3-1600.
And I don't know if anyone else noticed, but Intel have said on their Investor Meeting that Sandy Bridge's GPU will be able to show "order of magnitude" difference over their 65nm products, which is the GMA 4500.
If they can really manage 10x over the GMA 4500, I can see Intel's GPU being competitive with Llano.
I'd think so but the number of SPs probably is the least of the problems (that is, if you're compute bound you could well see an increase of more than 5x). Memory bandwidth is probably going to be the limiting factor (for graphics that is).
Somehow, I've got some feeling this is more "binary orders of magnitude"
.I think a 10x increase would be only possible if the chip drastically improves bandwidth saving tricks. Ironlake has started on this but compared to the competition it's still weak. A 10x improvement over GMA 4500 would be a huge step up from Ironlake, even larger than what Ironlake did over 4500.
Jawed
Legend
AMD's OpenCL stuff recently mentioned L4 cache:
http://forum.beyond3d.com/showthread.php?p=1426880#post1426880
Is that a generic term for system memory, or could it imply a wodge of eDRAM on-die as an L4 cache?
Jawed
http://forum.beyond3d.com/showthread.php?p=1426880#post1426880
Is that a generic term for system memory, or could it imply a wodge of eDRAM on-die as an L4 cache?
Jawed
5570 was terrible limited by memory bandwidth, this is about the performance I expect from Llano.
But not what I expect for the future of Fusion, there are some ways to reduce memory bandwidth requirements like TBR.
Why wouldn't it make sense to share the L2 (not the entire cache, just a subset)? I don't know much about Llano's L2.
Last edited:
Silent_Buddha
Legend
You know, now that I'm doing my yearly stint of recycling computers and harvesting the precious metals contained within, I've once again run into some really ancient machines that still contain Cirrus Logic and Trident on board graphics (among others).
Thinking about that in conjunction with the state of computing these days, there are no sucky onboard graphics solutions anymore.
It kind of makes me appreciate how good (at least with regards to OS and application display and acceleration) on board is now. Even Intel's worst is still very good at the things it needs to be good at. Something that just couldn't be said even 10 years ago.
Oh, and I just salvaged a state of the art 10 MB 1.5 inch high HDD. I think I may just keep the sucker.
End of minor sidetrack.
On topic, I still highly doubt even a 5x increase in integrated GPU performance (even moved on die) over current integrated solutions.
I'll be quite pleasantly surprised if Llano manages a 2x increase. And absolutely amazed if it approaches 3x. But, I'm honestly looking at a 50-100% increase as best case...
[Edit] Although in compute limited (virtually no memory access) situations it may be possible to approach 5x. But certainly not in games.
Regards,
SB
Thinking about that in conjunction with the state of computing these days, there are no sucky onboard graphics solutions anymore.
It kind of makes me appreciate how good (at least with regards to OS and application display and acceleration) on board is now. Even Intel's worst is still very good at the things it needs to be good at. Something that just couldn't be said even 10 years ago.
Oh, and I just salvaged a state of the art 10 MB 1.5 inch high HDD. I think I may just keep the sucker.
End of minor sidetrack.
On topic, I still highly doubt even a 5x increase in integrated GPU performance (even moved on die) over current integrated solutions.
I'll be quite pleasantly surprised if Llano manages a 2x increase. And absolutely amazed if it approaches 3x. But, I'm honestly looking at a 50-100% increase as best case...
[Edit] Although in compute limited (virtually no memory access) situations it may be possible to approach 5x. But certainly not in games.
Regards,
SB
End of minor sidetrack.
On topic, I still highly doubt even a 5x increase in integrated GPU performance (even moved on die) over current integrated solutions.
I'll be quite pleasantly surprised if Llano manages a 2x increase. And absolutely amazed if it approaches 3x. But, I'm honestly looking at a 50-100% increase as best case...
[Edit] Although in compute limited (virtually no memory access) situations it may be possible to approach 5x. But certainly not in games.
Regards,
SB
It depends on the architecture of the chip right? For instance the Cypress class GPUs are more bandwidth efficient than the RV770 class of chips. If this trend is continued then surely whatever architecture is installed in Fusion, and if its based on the updated uncore of Southern? Northern? islands we may very well see a precipitous increase in bandwidth efficiency on top of any other architectural changes like quad channel ram and on die cache which makes sense in effect due to the huge savings in stripping out an extra chip + memory architecture and putting that onto the CPU die. I don't think Fusion is your low end Athlon II solution.
OT: How many people work in Daves office? Im trying to figure out how big the cake was that they had to celebrate.... Of course the cake is a lie as Dave ate most of it before it even made it in the door.
Sorry, it is. Note that Bulldozer won't be in notebooks that fast so Llano will be holding the entire notebook line with exceptions to Bobcat.
Which might in fact work well for them - if Llano is a core component in every AMD notebook, switchable graphics will either i) start being extremely common (Optimus wins are here and there but not as much as the sea change some tech sites proclaimed) ii) there'd be no need for switchable graphics if Llano performs up to 9600M GT / GT330M level (which it does have potential to, just that bandwidth is not there yet as of now)
And then you can cut off execution units on the GPU/CPU to "enhance value" for the lower end.
I guess the questions are based around how good Llano is on the 32nm node and how they can mitigate the issue of bandwidth. Obviously their desktop GPU line is getting progressively better and I suspect that the earliest Fusion CPUs will be 2 core + GPU models so that ought to make the cache model a little more friendly.