Where does this obsession with GPUs attached to a Thunderbolt port come from? It is an extremely bad idea.
Sony is doing this with their ultra-portable. It basically means that you can buy an entry-level machine and get the highest-end graphics. I mocked up a price list here based on real-world MXM prices and models:
The Radeon 6970M in the highest iMac is on par with some of the fastest single desktop graphics cards you can buy and while that performance will reach on-die GPUs, it lets people get it now without breaking the bank.
I think people who buy a MBA for example, would love the idea of buying a $200-400 GPU to play high-end games at high quality.
Quote:
Originally Posted by wizard69
ECC RAM in a Mini? Interesting but I'm not sure that would be compelling in it's traditional markets.
There's no particular reason for going with ECC, it's just an extra that comes for free picking the Xeon chip. That low-powered Xeon seemed better than the i5/i7 alternatives for the Mini.
Quote:
Originally Posted by wizard69
Even if the Mini did come with two Thunderbolt ports I'm still not convinced it is a good idea. The speed at which you can share data between the CPU and the GPU is grossly compromised.
Intel designed TB for low-latency and essentially as an external PCI connection. It will certainly introduce some latency but the immediate impact is not clear. You're talking about nanosecond latencies so the real-world impact will depend on how you use it. You can buffer data to be operated on in the GPU memory.
Quote:
Originally Posted by solipsism
Even though the iMac has 2 Thunderbolt ports it's still a total of 20G/s as it's still a single Z68.
x4 PCIe 2.0 has 16Gbps of bandwidth so enough for two TB ports. Apple suggests you get two full speed ports, not shared 10Gbps (or 20Gbps total per port) but I suspect up to 8Gbps each:
Sx4 PCIe 2.0 has 16Gbps of bandwidth so enough for two TB ports. Apple suggests you get two full speed ports, not shared 10Gbps (or 20Gbps total per port) but I suspect up to 8Gbps each:
Only four lanes are used by Intel's Thunderbolt controller, the remaining lanes are used for things like Bluetooth and WiFi. Do the math and you'll realize that four PCIe 2.0 lanes are only good for 20Gbps of bandwidth, plus DMI between the Z68 chipset and Sandy Bridge is limited to 20Gbps itself. A single Thunderbolt port is capable of 20Gbps of bandwidth (10Gbps in each direction), so that works out well (if you don't use any of the other PCIe devices in the system at the same time). While the 21.5-inch iMac has a single Thunderbolt port, the 27-inch model has two. That's a total of up to 40Gbps of bandwidth to Thunderbolt devices, but only 20Gbps to the controller itself. Don't be fooled by the presence of two Thunderbolt ports on the 27-inch iMac, you don't get any more bandwidth than you would on the 21.5-inch model - you can just hook up more displays.
[INDENT][FONT="Arial"]Only four lanes are used by Intel's Thunderbolt controller, the remaining lanes are used for things like Bluetooth and WiFi. Do the math and you'll realize that four PCIe 2.0 lanes are only good for 20Gbps of bandwidth, plus DMI between the Z68 chipset and Sandy Bridge is limited to 20Gbps itself. A single Thunderbolt port is capable of 20Gbps of bandwidth (10Gbps in each direction), so that works out well (if you don't use any of the other PCIe devices in the system at the same time). While the 21.5-inch iMac has a single Thunderbolt port, the 27-inch model has two. That's a total of up to 40Gbps of bandwidth to Thunderbolt devices, but only 20Gbps to the controller itself. Don't be fooled by the presence of two Thunderbolt ports on the 27-inch iMac, you don't get any more bandwidth than you would on the 21.5-inch model - you can just hook up more displays.
DMI 2.0 and PCIe x4 are 20Gbps in both directions = 40Gbps of bandwidth. It could well be limited in some way but someone will have to test it out. Just wish manufacturers would hurry up with the devices.
Getting the PCIe 2.0 slot from Sonnet should at least allow people to hook up standard desktop GPUs or even fibre channel cards and max out the bandwidth on each port and see what the limits really are.
Sony is doing this with their ultra-portable. It basically means that you can buy an entry-level machine and get the highest-end graphics. I mocked up a price list here based on real-world MXM prices and models:
What can I say to me it is a rip off! Maybe that is American economics speaking but I just don't see any value in connecting a GPU in this manner.
Quote:
The Radeon 6970M in the highest iMac is on par with some of the fastest single desktop graphics cards you can buy and while that performance will reach on-die GPUs, it lets people get it now without breaking the bank.
The problem as I see it is that you are missing key points here. The nature of GPU usage in modern OS'es is best leveraged with tightly coupled hardware. Moving the GPU off the device so that it is a considerable distance form the CPU, its caches and I/O buses is not the way of the future.
Quote:
I think people who buy a MBA for example, would love the idea of buying a $200-400 GPU to play high-end games at high quality.
I'm not convinced myself. First off gaming doesn't appear to be a big issue with the AIRs. Second that means good GPU performance is never with you in a portable sense. The whole point of the AIR platform is portability so I'm not seeing a lot of people jumping at the chance to use this hardware.
Quote:
There's no particular reason for going with ECC, it's just an extra that comes for free picking the Xeon chip. That low-powered Xeon seemed better than the i5/i7 alternatives for the Mini.
In some ways it would make for a nicer Mac Mini platform. Considering the frequency of RAM failures having hardware detection of complete failures may be useful to many. Not to mention the auto correction of bit errors.
Quote:
Intel designed TB for low-latency and essentially as an external PCI connection. It will certainly introduce some latency but the immediate impact is not clear. You're talking about nanosecond latencies so the real-world impact will depend on how you use it. You can buffer data to be operated on in the GPU memory.
There is no question that TB is orders of magnitude slower than access to you processors caches and on chip buses.
Quote:
x4 PCIe 2.0 has 16Gbps of bandwidth so enough for two TB ports. Apple suggests you get two full speed ports, not shared 10Gbps (or 20Gbps total per port) but I suspect up to 8Gbps each:
I think you need some perspective here. Look up the bandwidth of GPU memory systems or dedicated GPU cards and then look up memory bandwidths for Sandy Bridge processors. 10Gbps is chump change when you look at what is happening on a modern CPU these days.
Only four lanes are used by Intel's Thunderbolt controller, the remaining lanes are used for things like Bluetooth and WiFi. Do the math and you'll realize that four PCIe 2.0 lanes are only good for 20Gbps of bandwidth, plus DMI between the Z68 chipset and Sandy Bridge is limited to 20Gbps itself. A single Thunderbolt port is capable of 20Gbps of bandwidth (10Gbps in each direction), so that works out well (if you don't use any of the other PCIe devices in the system at the same time). While the 21.5-inch iMac has a single Thunderbolt port, the 27-inch model has two. That's a total of up to 40Gbps of bandwidth to Thunderbolt devices, but only 20Gbps to the controller itself. Don't be fooled by the presence of two Thunderbolt ports on the 27-inch iMac, you don't get any more bandwidth than you would on the 21.5-inch model - you can just hook up more displays.
That and that there is apparently already two different TB chips in the wild. In one case Apple bypassed the DMI bottle neck and made use of the on Processor PCI Express channels. While this gets around some of the bandwidth limitations it highlights that we already are seeing Apple implement TB in different ways. Beyond that the TB controller is apparently a cross bar switch, so adding more PCI-Express ports to the switch should be easy and thus allow for more bandwidth.
However that bandwidth really means nothing when you compare it to tightly coupled GPU/CPU systems. When looking at PCI-Express connected hardware, GPU's on that bus will quickly go the way of the Dodo bird. Especially as OS code and app code takes greater and greater advantage of heterogeneous computing.
The nature of GPU usage in modern OS'es is best leveraged with tightly coupled hardware. Moving the GPU off the device so that it is a considerable distance form the CPU, its caches and I/O buses is not the way of the future.
Look up the bandwidth of GPU memory systems or dedicated GPU cards and then look up memory bandwidths for Sandy Bridge processors. 10Gbps is chump change when you look at what is happening on a modern CPU these days.
In the end, it comes down to real-world performance. IGPs are fine and are the way forward for graphics but for the moment, high-end dedicated cards still offer clear advantages, especially when we are talking about Intel's IGPs. It doesn't matter if an on-die GPU has all manner of theoretical advantages, if you can't run a game at the highest quality or can't have a CAD/3D viewport with a certain amount of polys or graphics software like Motion with as many different layers in HD, it makes no difference. Results are everything.
I'm with you on the AMD options but I don't see Apple going this route and with Intel IGPs, the Thunderbolt GPUs are the next best thing.
Quote:
Originally Posted by wizard69
First off gaming doesn't appear to be a big issue with the AIRs. Second that means good GPU performance is never with you in a portable sense.
There will be people who want to do the whole Starcraft/Warcraft thing but not have to buy a high-end MBP to run them at a decent quality. The external GPUs are used in the same way you use docking solutions for external drives. You don't need to take them with you but it's nice to have the performance when you are at a desk with a big screen.
Comments
Where does this obsession with GPUs attached to a Thunderbolt port come from? It is an extremely bad idea.
Sony is doing this with their ultra-portable. It basically means that you can buy an entry-level machine and get the highest-end graphics. I mocked up a price list here based on real-world MXM prices and models:
The Radeon 6970M in the highest iMac is on par with some of the fastest single desktop graphics cards you can buy and while that performance will reach on-die GPUs, it lets people get it now without breaking the bank.
I think people who buy a MBA for example, would love the idea of buying a $200-400 GPU to play high-end games at high quality.
ECC RAM in a Mini? Interesting but I'm not sure that would be compelling in it's traditional markets.
There's no particular reason for going with ECC, it's just an extra that comes for free picking the Xeon chip. That low-powered Xeon seemed better than the i5/i7 alternatives for the Mini.
Even if the Mini did come with two Thunderbolt ports I'm still not convinced it is a good idea. The speed at which you can share data between the CPU and the GPU is grossly compromised.
Intel designed TB for low-latency and essentially as an external PCI connection. It will certainly introduce some latency but the immediate impact is not clear. You're talking about nanosecond latencies so the real-world impact will depend on how you use it. You can buffer data to be operated on in the GPU memory.
Even though the iMac has 2 Thunderbolt ports it's still a total of 20G/s as it's still a single Z68.
x4 PCIe 2.0 has 16Gbps of bandwidth so enough for two TB ports. Apple suggests you get two full speed ports, not shared 10Gbps (or 20Gbps total per port) but I suspect up to 8Gbps each:
http://www.apple.com/uk/imac/performance.html
Sx4 PCIe 2.0 has 16Gbps of bandwidth so enough for two TB ports. Apple suggests you get two full speed ports, not shared 10Gbps (or 20Gbps total per port) but I suspect up to 8Gbps each:
http://www.apple.com/uk/imac/performance.html
[INDENT][FONT="Arial"]Only four lanes are used by Intel's Thunderbolt controller, the remaining lanes are used for things like Bluetooth and WiFi. Do the math and you'll realize that four PCIe 2.0 lanes are only good for 20Gbps of bandwidth, plus DMI between the Z68 chipset and Sandy Bridge is limited to 20Gbps itself. A single Thunderbolt port is capable of 20Gbps of bandwidth (10Gbps in each direction), so that works out well (if you don't use any of the other PCIe devices in the system at the same time). While the 21.5-inch iMac has a single Thunderbolt port, the 27-inch model has two. That's a total of up to 40Gbps of bandwidth to Thunderbolt devices, but only 20Gbps to the controller itself. Don't be fooled by the presence of two Thunderbolt ports on the 27-inch iMac, you don't get any more bandwidth than you would on the 21.5-inch model - you can just hook up more displays.
DMI 2.0 and PCIe x4 are 20Gbps in both directions = 40Gbps of bandwidth. It could well be limited in some way but someone will have to test it out. Just wish manufacturers would hurry up with the devices.
Getting the PCIe 2.0 slot from Sonnet should at least allow people to hook up standard desktop GPUs or even fibre channel cards and max out the bandwidth on each port and see what the limits really are.
Sony is doing this with their ultra-portable. It basically means that you can buy an entry-level machine and get the highest-end graphics. I mocked up a price list here based on real-world MXM prices and models:
What can I say to me it is a rip off! Maybe that is American economics speaking but I just don't see any value in connecting a GPU in this manner.
The Radeon 6970M in the highest iMac is on par with some of the fastest single desktop graphics cards you can buy and while that performance will reach on-die GPUs, it lets people get it now without breaking the bank.
The problem as I see it is that you are missing key points here. The nature of GPU usage in modern OS'es is best leveraged with tightly coupled hardware. Moving the GPU off the device so that it is a considerable distance form the CPU, its caches and I/O buses is not the way of the future.
I think people who buy a MBA for example, would love the idea of buying a $200-400 GPU to play high-end games at high quality.
I'm not convinced myself. First off gaming doesn't appear to be a big issue with the AIRs. Second that means good GPU performance is never with you in a portable sense. The whole point of the AIR platform is portability so I'm not seeing a lot of people jumping at the chance to use this hardware.
There's no particular reason for going with ECC, it's just an extra that comes for free picking the Xeon chip. That low-powered Xeon seemed better than the i5/i7 alternatives for the Mini.
In some ways it would make for a nicer Mac Mini platform. Considering the frequency of RAM failures having hardware detection of complete failures may be useful to many. Not to mention the auto correction of bit errors.
Intel designed TB for low-latency and essentially as an external PCI connection. It will certainly introduce some latency but the immediate impact is not clear. You're talking about nanosecond latencies so the real-world impact will depend on how you use it. You can buffer data to be operated on in the GPU memory.
There is no question that TB is orders of magnitude slower than access to you processors caches and on chip buses.
x4 PCIe 2.0 has 16Gbps of bandwidth so enough for two TB ports. Apple suggests you get two full speed ports, not shared 10Gbps (or 20Gbps total per port) but I suspect up to 8Gbps each:
I think you need some perspective here. Look up the bandwidth of GPU memory systems or dedicated GPU cards and then look up memory bandwidths for Sandy Bridge processors. 10Gbps is chump change when you look at what is happening on a modern CPU these days.
That and that there is apparently already two different TB chips in the wild. In one case Apple bypassed the DMI bottle neck and made use of the on Processor PCI Express channels. While this gets around some of the bandwidth limitations it highlights that we already are seeing Apple implement TB in different ways. Beyond that the TB controller is apparently a cross bar switch, so adding more PCI-Express ports to the switch should be easy and thus allow for more bandwidth.
However that bandwidth really means nothing when you compare it to tightly coupled GPU/CPU systems. When looking at PCI-Express connected hardware, GPU's on that bus will quickly go the way of the Dodo bird. Especially as OS code and app code takes greater and greater advantage of heterogeneous computing.
The nature of GPU usage in modern OS'es is best leveraged with tightly coupled hardware. Moving the GPU off the device so that it is a considerable distance form the CPU, its caches and I/O buses is not the way of the future.
Look up the bandwidth of GPU memory systems or dedicated GPU cards and then look up memory bandwidths for Sandy Bridge processors. 10Gbps is chump change when you look at what is happening on a modern CPU these days.
In the end, it comes down to real-world performance. IGPs are fine and are the way forward for graphics but for the moment, high-end dedicated cards still offer clear advantages, especially when we are talking about Intel's IGPs. It doesn't matter if an on-die GPU has all manner of theoretical advantages, if you can't run a game at the highest quality or can't have a CAD/3D viewport with a certain amount of polys or graphics software like Motion with as many different layers in HD, it makes no difference. Results are everything.
I'm with you on the AMD options but I don't see Apple going this route and with Intel IGPs, the Thunderbolt GPUs are the next best thing.
First off gaming doesn't appear to be a big issue with the AIRs. Second that means good GPU performance is never with you in a portable sense.
There will be people who want to do the whole Starcraft/Warcraft thing but not have to buy a high-end MBP to run them at a decent quality. The external GPUs are used in the same way you use docking solutions for external drives. You don't need to take them with you but it's nice to have the performance when you are at a desk with a big screen.