[sigh] Onlooker, I see you're hoping to the very end for SLI in Mac Pros. Good luck. I voted no SLI Woodcrest board in Mac Pros. Apple is not out to make the ultimate 3D Workstation or Gaming machine. Their focus for Mac Pro is juicy profits and the pro video, film, and music market, with the target market within a certain envelope. Hardcore 3D Workstation is beyond the $2000-$5500 envelope, IMHO (Yeah you could custom the Mac Pro to over $10k but that is for very specific applications). They should offer top of the line Quadro, single card, if going with nVidia. Otherwise, FireGL, single card. 7600GS to 7600GT standard if nVidia, 7900GT and 7900GTX options. Well, those are my predictions. If SLI shows up on Monday, I'll personally PM you that I ate my shorts
if apple does put sli in it may just be x8 x8 not x16 x16
if apple does put sli in it may just be x8 x8 not x16 x16
We have that now, but it's pretty much a waste of money to have 2 cards do the same thing that one 16x card will do. The improvements from 2 8x cards to one 16x card isn't enough to justify the price of a second card. Why do you think none of us has done it. All the stats are online already from the original boards that had one 16x slot, and an 8x slot, and the difference once you took the 8x card out wasn't more than fractions of points because the 8x slot slows the 16x card to it's speed in both the 8x and the other 16x slot. Why do it?
We have that now, but it's pretty much a waste of money to have 2 cards do the same thing that one 16x card will do. The improvements from 2 8x cards to one 16x card isn't enough to justify the price of a second card. Why do you think none of us has done it. All the stats are online already from the original boards that had one 16x slot, and an 8x slot, and the difference once you took the 8x card out wasn't more than fractions of points because the 8x slot slows the 16x card to it's speed in both the 8x and the other 16x slot. Why do it?
Well, being on x8 doesn't cut a card to half speed, it can hurt performance a little. I'd rather be on x16 than x8, but I'd rather have SLI than one x16 card.
Evidently onlooker thinks that the x-speed of the card interface is the speed of the card... :loll:
No the second (x8) slot with the 16x card in that scenario reduces the speed of the first (16x) card to 8x so they can communicate properly. I don't know if I phrased that right but they equal out to 8x each because of the bottleneck. .
You are correct (both the x16 and x8 slot going to x8 in SLI mode), but if I understand correctly, that's only in SLI mode. In normal mode, the second card is ignored (or running a separate display).
And 7x00 series don't totally consume 16 lanes of bandwidth, but they would be slightly hurt by 8 lanes. And obviously there is overhead from splitting the rendering into two parts. So an SLI solution is going to be less than the sum of the parts.
It's probably better to buy a high end card than two SLI'd mid-range cards, but let's say I get a 7600GT now, and in 6-12 months, I spend $100 to get a second (once 8x00 series are out). I don't have to drop $300 or so for a 7900 or more for a 8600/8800, but I do get a decent speed boost.
Not only was (and is, AFAIK) Apple in the Hypertransport group, they used it in the Powermac G5 to provide a bus to the Airport, Firewire, and some other stuff.
No the second (x8) slot with the 16x card in that scenario reduces the speed of the first (16x) card to 8x so they can communicate properly. I don't know if I phrased that right but they equal out to 8x each because of the bottleneck...At least thats what I read.
Yes in PC boards originally and still with the lower end, only 16 PCIExpress lanes are dedicated to the GPU. Hence one card, get x16 lanes. Put 2 graphic cards in SLI, get x8 x8 lanes per GPU.
My understanding is, for a general GPU:
One GPU x16 , let's say is baseline 1x performance.
Two GPUS x8 x8, you get about 1.5x performance.
Two GPUS x16 x16, you get about 1.6-1.7x performance.
Very very rough figures. You'd have to trawl through TomsHardware and AnandTech but the figures I have above are kinda what's going on.
At the end of the day, its about hardcore gaming enthusiasts that go for SLI. Might as well ensure the board has two x16 lanes for each GPU if you're going SLI. Initial SLI solutions "split" it into x8 x8 because they wanted IMO to get it out to market fast. nVidia then rejigged their chipset once SLI caught on to give full x16 lanes to *each* GPU.
Not only was (and is, AFAIK) Apple in the Hypertransport group, they used it in the Powermac G5 to provide a bus to the Airport, Firewire, and some other stuff.
Yeah... Apple is a founding member of HyperTransport group. In the PowerMac G5, "Serial ATA, FireWire, USB, audio, and wireless technologies are integrated through two bidirectional 800MHz HyperTransport interconnects for a maximum throughput of 1.6 GBps, providing ample throughput for a host of peripheral devices." From no. 9 in http://www.apple.com/powermac/pciexpress.html
Unlike AMD though they don't use HyperTransport for the FrontSideBus (no mention of HyperTransport):
"Dual bidirectional frontside buses
Leveraging the dual frontside bus architecture pioneered in the original Power Mac G5, each processor in Power Mac G5 Quad systems has its own interface to the system controller — unlike traditional dual-processor systems, which constrain throughput by placing all processor resources on one bus. This high-performance frontside bus architecture also enables each core to discover and access data in the other cores’ caches, further increasing performance on dual-core and quad-core systems."
Check the previous pages for real-world tests with different programs.
1.5x improvement is possible (look at the 6600GT vs 6600GT SLI). On x16 x16 SLI, okay, maybe it goes up, say 1.6x or 1.7x, maybe 1.7x is pushing it a bit. All depends on the benchmarks you run and at what settings, etc. I'll concede I haven't been able to find specific benchmarks about x8 x8 vs x16 x16. But again SLI x8 x8 or x16 x16, clearly, 1.4x to 1.5x improvement is possible...!!
At the very high end SLI looks to be just silly eg. SLI 7900GTX because you get a huge friggin frame rate anyway and then the game becomes CPU limited at that stage. I like the 6600GT results because the 6600GT, 7600GT, 7900GT are cards that would give you bang for buck by popping in another for SLI. Well, SLI is mostly for bragging rights anyway, but I like the idea, of keeping the card in there, just adding another card, and not throwing the old card out for a new card. I just checked the prices of a 7900GT though, it's about the $200 mark Long story short, SLI as upgrade strategy, somewhat doubtful, for hard enthusiasts, go for it. One caveat: look at the 7950GX2 vs the 7900GTX, performance and price, the 7950GX2 kicks ass... http://www.tomshardware.com/2006/06/...gx2/page6.html
Yeah... Apple is a founding member of HyperTransport group. In the PowerMac G5, "Serial ATA, FireWire, USB, audio, and wireless technologies are integrated through two bidirectional 800MHz HyperTransport interconnects for a maximum throughput of 1.6 GBps, providing ample throughput for a host of peripheral devices." From no. 9 in http://www.apple.com/powermac/pciexpress.html
Unlike AMD though they don't use HyperTransport for the FrontSideBus (no mention of HyperTransport):
"Dual bidirectional frontside buses
Leveraging the dual frontside bus architecture pioneered in the original Power Mac G5, each processor in Power Mac G5 Quad systems has its own interface to the system controller — unlike traditional dual-processor systems, which constrain throughput by placing all processor resources on one bus. This high-performance frontside bus architecture also enables each core to discover and access data in the other cores’ caches, further increasing performance on dual-core and quad-core systems."
AMD doesn't use a FSB. They use two HyperTransport links, one to the RAM, and one to the Southbridge.
Comments
[sigh] Onlooker, I see you're hoping to the very end for SLI in Mac Pros. Good luck. I voted no SLI Woodcrest board in Mac Pros. Apple is not out to make the ultimate 3D Workstation or Gaming machine. Their focus for Mac Pro is juicy profits and the pro video, film, and music market, with the target market within a certain envelope. Hardcore 3D Workstation is beyond the $2000-$5500 envelope, IMHO (Yeah you could custom the Mac Pro to over $10k but that is for very specific applications). They should offer top of the line Quadro, single card, if going with nVidia. Otherwise, FireGL, single card. 7600GS to 7600GT standard if nVidia, 7900GT and 7900GTX options. Well, those are my predictions. If SLI shows up on Monday, I'll personally PM you that I ate my shorts
if apple does put sli in it may just be x8 x8 not x16 x16
if apple does put sli in it may just be x8 x8 not x16 x16
We have that now, but it's pretty much a waste of money to have 2 cards do the same thing that one 16x card will do. The improvements from 2 8x cards to one 16x card isn't enough to justify the price of a second card. Why do you think none of us has done it. All the stats are online already from the original boards that had one 16x slot, and an 8x slot, and the difference once you took the 8x card out wasn't more than fractions of points because the 8x slot slows the 16x card to it's speed in both the 8x and the other 16x slot. Why do it?
We have that now, but it's pretty much a waste of money to have 2 cards do the same thing that one 16x card will do. The improvements from 2 8x cards to one 16x card isn't enough to justify the price of a second card. Why do you think none of us has done it. All the stats are online already from the original boards that had one 16x slot, and an 8x slot, and the difference once you took the 8x card out wasn't more than fractions of points because the 8x slot slows the 16x card to it's speed in both the 8x and the other 16x slot. Why do it?
Could you mabye rephrase that? I'm not following,
Evidently onlooker thinks that the x-speed of the card interface is the speed of the card... :loll:
No the second (x8) slot with the 16x card in that scenario reduces the speed of the first (16x) card to 8x so they can communicate properly. I don't know if I phrased that right but they equal out to 8x each because of the bottleneck. .
At least thats what I read.
And 7x00 series don't totally consume 16 lanes of bandwidth, but they would be slightly hurt by 8 lanes. And obviously there is overhead from splitting the rendering into two parts. So an SLI solution is going to be less than the sum of the parts.
It's probably better to buy a high end card than two SLI'd mid-range cards, but let's say I get a 7600GT now, and in 6-12 months, I spend $100 to get a second (once 8x00 series are out). I don't have to drop $300 or so for a 7900 or more for a 8600/8800, but I do get a decent speed boost.
SLI or CrossFire-depends on the chipset-Intel's own is likely, so no SLI.
In the wake of AMD's acquisition of ATI, I'd not be surprised about any future collaboration between nVidia and intel.
Now, whether this extends to Apple at this time, who can say.
BTW, wasn't Apple in the HyperTransport group a couple of years ago?
T
No the second (x8) slot with the 16x card in that scenario reduces the speed of the first (16x) card to 8x so they can communicate properly. I don't know if I phrased that right but they equal out to 8x each because of the bottleneck...At least thats what I read.
Yes in PC boards originally and still with the lower end, only 16 PCIExpress lanes are dedicated to the GPU. Hence one card, get x16 lanes. Put 2 graphic cards in SLI, get x8 x8 lanes per GPU.
My understanding is, for a general GPU:
One GPU x16 , let's say is baseline 1x performance.
Two GPUS x8 x8, you get about 1.5x performance.
Two GPUS x16 x16, you get about 1.6-1.7x performance.
Very very rough figures. You'd have to trawl through TomsHardware and AnandTech but the figures I have above are kinda what's going on.
At the end of the day, its about hardcore gaming enthusiasts that go for SLI. Might as well ensure the board has two x16 lanes for each GPU if you're going SLI. Initial SLI solutions "split" it into x8 x8 because they wanted IMO to get it out to market fast. nVidia then rejigged their chipset once SLI caught on to give full x16 lanes to *each* GPU.
Not only was (and is, AFAIK) Apple in the Hypertransport group, they used it in the Powermac G5 to provide a bus to the Airport, Firewire, and some other stuff.
Yeah... Apple is a founding member of HyperTransport group. In the PowerMac G5, "Serial ATA, FireWire, USB, audio, and wireless technologies are integrated through two bidirectional 800MHz HyperTransport interconnects for a maximum throughput of 1.6 GBps, providing ample throughput for a host of peripheral devices." From no. 9 in http://www.apple.com/powermac/pciexpress.html
Unlike AMD though they don't use HyperTransport for the FrontSideBus (no mention of HyperTransport):
"Dual bidirectional frontside buses
Leveraging the dual frontside bus architecture pioneered in the original Power Mac G5, each processor in Power Mac G5 Quad systems has its own interface to the system controller — unlike traditional dual-processor systems, which constrain throughput by placing all processor resources on one bus. This high-performance frontside bus architecture also enables each core to discover and access data in the other cores’ caches, further increasing performance on dual-core and quad-core systems."
My understanding is, for a general GPU:
Two GPUS x8 x8, you get about 1.5x performance.
That looks like wishful thinking to me. I don't remember any of the tests I read coming even close to that.
I remember it being like 1.2x or 1.3x which didn't impress me.
Check the previous pages for real-world tests with different programs.
1.5x improvement is possible (look at the 6600GT vs 6600GT SLI). On x16 x16 SLI, okay, maybe it goes up, say 1.6x or 1.7x, maybe 1.7x is pushing it a bit. All depends on the benchmarks you run and at what settings, etc. I'll concede I haven't been able to find specific benchmarks about x8 x8 vs x16 x16. But again SLI x8 x8 or x16 x16, clearly, 1.4x to 1.5x improvement is possible...!!
At the very high end SLI looks to be just silly eg. SLI 7900GTX because you get a huge friggin frame rate anyway and then the game becomes CPU limited at that stage. I like the 6600GT results because the 6600GT, 7600GT, 7900GT are cards that would give you bang for buck by popping in another for SLI. Well, SLI is mostly for bragging rights anyway, but I like the idea, of keeping the card in there, just adding another card, and not throwing the old card out for a new card. I just checked the prices of a 7900GT though, it's about the $200 mark
Yeah... Apple is a founding member of HyperTransport group. In the PowerMac G5, "Serial ATA, FireWire, USB, audio, and wireless technologies are integrated through two bidirectional 800MHz HyperTransport interconnects for a maximum throughput of 1.6 GBps, providing ample throughput for a host of peripheral devices." From no. 9 in http://www.apple.com/powermac/pciexpress.html
Unlike AMD though they don't use HyperTransport for the FrontSideBus (no mention of HyperTransport):
"Dual bidirectional frontside buses
Leveraging the dual frontside bus architecture pioneered in the original Power Mac G5, each processor in Power Mac G5 Quad systems has its own interface to the system controller — unlike traditional dual-processor systems, which constrain throughput by placing all processor resources on one bus. This high-performance frontside bus architecture also enables each core to discover and access data in the other cores’ caches, further increasing performance on dual-core and quad-core systems."
AMD doesn't use a FSB. They use two HyperTransport links, one to the RAM, and one to the Southbridge.
AMD doesn't use a FSB. They use two HyperTransport links, one to the RAM, and one to the Southbridge.
RAM is part of the cpu it does use HyperTransport for cpu to cpu ram access