Hello, last year some of you mentioned that one had to use water cooling system due to heat generated by the i9-7900X. I am afraid to use water cooling as water might damage the expensive system. Has the situation been improved? Can I just use a large heatsink and fans instead? Over the past few months, is there any new CPU that has similar or higher performance than the i9-7900X but does not require water cooling system to keep it cool?
BFG-9000
Splendid
- Sep 17, 2016
- 4,252
- 810
- 29,590
I'm afraid you are misunderstanding the function of the PLX or IDT chips, which function as switches. Any single x16 device can get full bandwidth at a time, but all of them cannot operate at full speed simultaneously.
It's like expecting 15 devices attached to a 16-port gigabit switch to all be able to pull from a server at gigabit speeds simultaneously. There is a bottleneck, which is the single gigabit link to the server. Note the PCH also functions as a switch chip--its up to 24 PCIe lanes must go through DMI 3.0 to the processor which is equivalent to x4.
However, scientific applications tend to only send the instructions to the GPUs over the bus, which the GPUs then crunch using their local VRAM before sending the results back. So they probably aren't hampered much by restricted bus width.
While one of the GPUs could both run CUDA apps and a display at the same time, depending on your software you may run into some display issues, so it's really better to have a separate card for display. It can be connected to the PCH PCIe lanes to not take bandwidth from the main cards, as the GT710 came in x1 and the GT1030 is x4 only.
BTW, the actual i7-9700k next year will have two more cores than the present i7-8700k. Knowing Intel, it will probably require a new chipset.
It's like expecting 15 devices attached to a 16-port gigabit switch to all be able to pull from a server at gigabit speeds simultaneously. There is a bottleneck, which is the single gigabit link to the server. Note the PCH also functions as a switch chip--its up to 24 PCIe lanes must go through DMI 3.0 to the processor which is equivalent to x4.
However, scientific applications tend to only send the instructions to the GPUs over the bus, which the GPUs then crunch using their local VRAM before sending the results back. So they probably aren't hampered much by restricted bus width.
While one of the GPUs could both run CUDA apps and a display at the same time, depending on your software you may run into some display issues, so it's really better to have a separate card for display. It can be connected to the PCH PCIe lanes to not take bandwidth from the main cards, as the GT710 came in x1 and the GT1030 is x4 only.
BTW, the actual i7-9700k next year will have two more cores than the present i7-8700k. Knowing Intel, it will probably require a new chipset.
Thanks. There are not that many reviews on the ASUS WS X299 SAGE and most Amazon users seem to have returned it for a refund. Perhaps better to stay away from it.
I plan to order the components this week. I don't quite understand how the PCI count works. For the 8700K (16 lanes), 7820X (28 lanes) and 7900X (44 lanes), how many Nvidia 1080ti GPU can I run at the same time all at x16 speed? In addition to that, how many PCIe NVMe M.2 SSD can be run at full speed?
I plan to order the components this week. I don't quite understand how the PCI count works. For the 8700K (16 lanes), 7820X (28 lanes) and 7900X (44 lanes), how many Nvidia 1080ti GPU can I run at the same time all at x16 speed? In addition to that, how many PCIe NVMe M.2 SSD can be run at full speed?
BFG-9000
Splendid
- Sep 17, 2016
- 4,252
- 810
- 29,590
Only the 7900x can run two cards at x16 each, because 28 is less than 32.
The NVMe is a more difficult question, because it depends on how the motherboard is wired. On the consumer platforms they will be wired to the PCIe lanes of the PCH for sure, and note that a single x4 NVMe drive can fully saturate its DMI 3.0 link to the CPU by itself. The Skylake-X platform could easily wire three x4 NVMe to the spare CPU lanes (44-32=12 left over) but the board manufacturer may have designed the board to work with processors with fewer PCIe lanes so also wired the NVMe to the PCH. So you'd have to ask the manufacturer.
That's one of the problems caused by the Kaby Lake-X 7740x even existing. All Threadripper CPUs have 64 lanes so TR4 boards are much easier to design.
The NVMe is a more difficult question, because it depends on how the motherboard is wired. On the consumer platforms they will be wired to the PCIe lanes of the PCH for sure, and note that a single x4 NVMe drive can fully saturate its DMI 3.0 link to the CPU by itself. The Skylake-X platform could easily wire three x4 NVMe to the spare CPU lanes (44-32=12 left over) but the board manufacturer may have designed the board to work with processors with fewer PCIe lanes so also wired the NVMe to the PCH. So you'd have to ask the manufacturer.
That's one of the problems caused by the Kaby Lake-X 7740x even existing. All Threadripper CPUs have 64 lanes so TR4 boards are much easier to design.
Thanks BFG-9000.
Perhaps I miss something or I don't know how to read the specs properly. For example, according to the specs, 7820X (28-lane CPU) can run two PCIe 3.0 concurrently at x16 speed each. As for the 7900X (44-lane CPU), it can run three PCIe 3.0 concurrently at x16 speed each. Am I wrong?
https://www.asus.com/ca-en/Motherboards/PRIME-X299-DELUXE/specifications/
I shall check with manufacturers about how they wire NVme.
Perhaps I miss something or I don't know how to read the specs properly. For example, according to the specs, 7820X (28-lane CPU) can run two PCIe 3.0 concurrently at x16 speed each. As for the 7900X (44-lane CPU), it can run three PCIe 3.0 concurrently at x16 speed each. Am I wrong?
https://www.asus.com/ca-en/Motherboards/PRIME-X299-DELUXE/specifications/
I shall check with manufacturers about how they wire NVme.
theonerm2
Splendid
- Dec 28, 2009
- 4,114
- 81
- 26,740
I don't believe you can run 2 cards at 16x on 28 lanes because 16x2 is 32. It works by adding the lanes together. 16x and 16x is 32 lanes. So you would need 32 lanes. Also you would need 48 for 3 cards running at 16x. Do you see how it works now?
BFG-9000
Splendid
- Sep 17, 2016
- 4,252
- 810
- 29,590
On that board the 2nd or 3rd x16 card is connected to the PCH. Which itself is connected to the CPU by x4, a bottleneck.
In case you were wondering why the number of PCIe lanes was cutoff exactly as they were, it's because that way anyone who needed more would have to upgrade past their Prosumer HEDT line all the way to the expensive Professional/Enterprise Skylake-SP Xeon line on LGA3647. On Intel, if you want more, you must pay more.
48 lanes per CPU means a board such as the ASUS WS-C621E SAGE can run four x16 if two CPUs are installed. Technically if you don't attach anything else PCIe to either CPU, a board could be made to run six x16.
Intel's grand plans were thrown into disarray when AMD released Ryzen, so they've been panicking and just tossing more cores at everything (even lowly i3) for now, because architectural changes take longer to implement. For what it's worth, AMD's Xeon equivalent line is Epyc, and all CPUs in that lineup have 128 lanes each.
In case you were wondering why the number of PCIe lanes was cutoff exactly as they were, it's because that way anyone who needed more would have to upgrade past their Prosumer HEDT line all the way to the expensive Professional/Enterprise Skylake-SP Xeon line on LGA3647. On Intel, if you want more, you must pay more.
48 lanes per CPU means a board such as the ASUS WS-C621E SAGE can run four x16 if two CPUs are installed. Technically if you don't attach anything else PCIe to either CPU, a board could be made to run six x16.
Intel's grand plans were thrown into disarray when AMD released Ryzen, so they've been panicking and just tossing more cores at everything (even lowly i3) for now, because architectural changes take longer to implement. For what it's worth, AMD's Xeon equivalent line is Epyc, and all CPUs in that lineup have 128 lanes each.
Thanks. I cannot decide among 8700K, 7900X, Threadripper or Epyc yet since I do not know which corresponding motherboards are compatible with Ubuntu Linux (the most important) and perhaps also with Hackintosh.
Am I right that PCIE4/PCIE5 components won't show up at least for another 1.5 years?
Am I right that PCIE4/PCIE5 components won't show up at least for another 1.5 years?
I have decided to go for either 8700K or 7900X. Within the next 6-8 months, will Intel release more powerful CPU that is similar in price as the current 7900X but with higher speed, have at least 6 cores and at least the same or more number of PCIe lanes?
- May 27, 2010
- 28,410
- 5,784
- 144,340
I believe Cascade Lake is scheduled for the end of this year, so there will be an incremental improvement to the current Skylake-X CPUs. Basically going to be "Coffee Lake-X" equivalent. I doubt Intel is going to switch up the lane scheme too much. They might add 44 lanes to the lower half of the line-up to try and compete with AMD.
The potential 9700k on Z390 will probably still only have x16 lanes total from the CPU. Just what you should expect from a consumer grade platform.
Going to be a long wait. The usual advice is to buy when you need it unless a launch is imminent.
The potential 9700k on Z390 will probably still only have x16 lanes total from the CPU. Just what you should expect from a consumer grade platform.
Going to be a long wait. The usual advice is to buy when you need it unless a launch is imminent.
- Dec 26, 2012
- 30,759
- 4,553
- 142,790
The largest single difference between big air and big water is top end ability. For anything below @250w, either is just as good as the other and air has the distinct advantage of being far cheaper mostly. But, just like any smaller air cooler, even big air has its limits, that being @250-300w. A full custom loop is only limited by the liquid, flow, radiators, all of which can be increased as needed. Drop a couple of D5 pumps into a bathtub full of water and there isn't a cpu on the planet that'll heat it up.
The X cpus are 140w ish TDP. That's thermal design power. Happens to be a bunch of bs as it's a number set by Intel after running a specified series of applications at very moderate usage. Peak power is what you'd expect at a Prime95 full load 100% torture test, it's the maximum wattage a cpu at stock settings will supply (temp wattage and power wattage are different, but close enough not to worry about exact differences) and usually runs at @1.5-2x TDP. With the X cpus count on 2x TDP +. Before OC. That's roughly @280w at stock settings, so good luck doing anything challenging with even the best air coolers. You are pretty much maxed out before any OC considerations.
The only time you'll have any luck on air cooling on the X cpus is with gaming, as most games don't use the core count and even then are hard pressed to use 70% of that. For production or any of its relative apps that will use most if not all cores to decent levels, you need liquid cooling, or an alternative such as peltier. No ifs ands or butts.
The X cpus are 140w ish TDP. That's thermal design power. Happens to be a bunch of bs as it's a number set by Intel after running a specified series of applications at very moderate usage. Peak power is what you'd expect at a Prime95 full load 100% torture test, it's the maximum wattage a cpu at stock settings will supply (temp wattage and power wattage are different, but close enough not to worry about exact differences) and usually runs at @1.5-2x TDP. With the X cpus count on 2x TDP +. Before OC. That's roughly @280w at stock settings, so good luck doing anything challenging with even the best air coolers. You are pretty much maxed out before any OC considerations.
The only time you'll have any luck on air cooling on the X cpus is with gaming, as most games don't use the core count and even then are hard pressed to use 70% of that. For production or any of its relative apps that will use most if not all cores to decent levels, you need liquid cooling, or an alternative such as peltier. No ifs ands or butts.
Thanks. I am interested in INTEL CPUs that support AVX512 VNNI. Anybody knows when they will be available?
https://en.wikipedia.org/wiki/AVX-512
The price difference between a 8700K system and a 9700X system is about 1K. I need to buy now but if a better CPU that supports AVX512 VNNI will be available in a few months, perhaps I just buy the 8700K now and upgrade later.
Am I right that when the PCIe 4 or 5 systems come out 1.5-2 years later, even the best performed 4-GPU 1080Ti/Titan system will become incomparable?
https://en.wikipedia.org/wiki/AVX-512
The price difference between a 8700K system and a 9700X system is about 1K. I need to buy now but if a better CPU that supports AVX512 VNNI will be available in a few months, perhaps I just buy the 8700K now and upgrade later.
Am I right that when the PCIe 4 or 5 systems come out 1.5-2 years later, even the best performed 4-GPU 1080Ti/Titan system will become incomparable?
- Dec 26, 2012
- 30,759
- 4,553
- 142,790
Pcie 3 has been out for years, and up until recently with the high end pascal gpus, even using up the bandwidth offered at x8 pcie 2 wasn't easy. And even dual 1080ti can't fill the bandwidth of pcie 3 x8. It would take a couple of epically massive gpus to fill out the bandwidth of x16.
With sli and cf all but dead, in 2 years (hopefully) DX12 will actually start using its mgpu ability. If that happens, you might just see quad gpu mobo's start coming back, but until then it's pointless really. So will pcie 4 be a big boost? I'm not thinking so, not unless in the next 2 years gpu output basically doubles and the gtx 3080ti treats a 1080ti the way a 1080ti treats a 660ti.
Many ppl gripe about Intel dragging out the tick-tock cpu upgrades. They only do that so they can stay in business. There's only so much that can be done with a chunk of silicon. Intel and amd have been lookin to the 5nm process for years, but that's basically the theoretical limit of silicon based chips. They are already working on 7nm processes. That goes for the gpu too. Same basic architecture. So without a breakthrough of epic proportions, pcie 5 will be a moot point as the chip just won't have the capability to use it to any extent. Gpu or cpu. Which is why Intel finally got off its duff and started adding cores vrs pushing performance on quad core cpus. It's reaching silicon limits.
With sli and cf all but dead, in 2 years (hopefully) DX12 will actually start using its mgpu ability. If that happens, you might just see quad gpu mobo's start coming back, but until then it's pointless really. So will pcie 4 be a big boost? I'm not thinking so, not unless in the next 2 years gpu output basically doubles and the gtx 3080ti treats a 1080ti the way a 1080ti treats a 660ti.
Many ppl gripe about Intel dragging out the tick-tock cpu upgrades. They only do that so they can stay in business. There's only so much that can be done with a chunk of silicon. Intel and amd have been lookin to the 5nm process for years, but that's basically the theoretical limit of silicon based chips. They are already working on 7nm processes. That goes for the gpu too. Same basic architecture. So without a breakthrough of epic proportions, pcie 5 will be a moot point as the chip just won't have the capability to use it to any extent. Gpu or cpu. Which is why Intel finally got off its duff and started adding cores vrs pushing performance on quad core cpus. It's reaching silicon limits.
Thanks. INTEL should give more lanes to higher-end CPUs. In the case of the 8700K, it would be perfect if it had over 40 lanes. Yet, INTEL made it 16. It is like INTEL is playing games and making people pay a lot and change motherboard all the time. If I choose the INTEL path now, the worse thing that could happen is to buy a new motherboard and a new CPU because DDR4-3600 these days can be used in both INTEL and TR platforms. Am I correct?
- May 27, 2010
- 28,410
- 5,784
- 144,340
As I said you would expect the 8700k to have only 16x lanes. The vast majority of people on that chipset don't need multi-GPU or high end expansion cards. That is exactly what HEDT is for. That is clear market segmentation and justifiable. Lanes cost silicon and board space, having the capability would increase the cost of consumer boards to the levels of the HEDT platform. Not going to put a $100 CPU into a $250 motherboard...not to mention the increased cost of the CPUs themselves. (Larger chips mean less chips per wafer)
DDR4-3600 is probably a little extreme for threadripper, would be a serious overclock given the way they set up the memory bus. 3600 on Intel is the start of the very high end, so overclocking would again be very necessary, but likely doable. The more ram you add though the harder that gets. If you go by the official numbers it would be 2666 for Intel and AMD.
I am in a dilemma. 7900x has AVX-512 support but it can only handle 2 GPUs all at x16 even it is expensive. As for the TR, it has no support for AVX-512 and performs relatively poorly in AVX-2 applications. However, it supports 3 GPUs at at all x16. The ideal system is to have AVX-512 support and 4 GPUs all running at x16.
For games, x16 or x8 does not matter but for CUDA computations, the performance does decrease if using x8 rather than x16.
In terms of performance in moving large amount of data to and from the GPUs, overall system-wide performance and memory/PCI bandwidth, which is better? X299 or X399? I think they both use quad channels.
For games, x16 or x8 does not matter but for CUDA computations, the performance does decrease if using x8 rather than x16.
In terms of performance in moving large amount of data to and from the GPUs, overall system-wide performance and memory/PCI bandwidth, which is better? X299 or X399? I think they both use quad channels.
BFG-9000
Splendid
- Sep 17, 2016
- 4,252
- 810
- 29,590
Perhaps there is a 7900x board that lets you run 4 GPUs all at x8. Unfortunately that Prime X299-Deluxe only allows two cards at x16, plus two additional cards that have to share the DMI 3.0 with all of the SATA and USB and networking.
And the 8700k is hardly a "higher end CPU." It fits into their very cheapest Consumer platform for their most budget-conscious customers and sells for just over a few hundred bucks. To put things into historical perspective:
And the 8700k is hardly a "higher end CPU." It fits into their very cheapest Consumer platform for their most budget-conscious customers and sells for just over a few hundred bucks. To put things into historical perspective:
Thanks. I have been checking various reviews but they have conflicting information. It seems that there is a confusion on the number of "true PCIe 3.0x16" connections "directly from the CPU". Is this depending on the CPU or motherboard? Since the ability to move large amount of data between the GPU and the motherboard/cpu is important, I want to know which has more "true PCIe 3.0x16" connection "directly from the CPU" in the case of i9-7900x and the Threadripper (I guess all three versions are the same).
For example, the TR should support three PCIe 3.0 all running at x16 but one site mentioned that the MSI X399 Gaming Pro has only two slots "wired for true" x16 connections. The other site mentions 4x PCIe 3.0x16 as PCIe slots for graphics (from CPU)".
http://www.tomshardware.com/reviews/msi-x399-gaming-pro-carbon-ac-threadripper-motherboard,5307.html
https://www.anandtech.com/show/11685/amd-threadripper-x399-motherboards/9
These seem to be conflicting info or I just miss something to understand it.
If the TR and 7900X both support "true PCIe 3.0x16 connections directly from the CPU", does that mean even the TR has more lanes, there is no advantage over the 7900X in terms of performance to move large amount of data between the GPUs and the CPU?
For example, the TR should support three PCIe 3.0 all running at x16 but one site mentioned that the MSI X399 Gaming Pro has only two slots "wired for true" x16 connections. The other site mentions 4x PCIe 3.0x16 as PCIe slots for graphics (from CPU)".
http://www.tomshardware.com/reviews/msi-x399-gaming-pro-carbon-ac-threadripper-motherboard,5307.html
https://www.anandtech.com/show/11685/amd-threadripper-x399-motherboards/9
These seem to be conflicting info or I just miss something to understand it.
If the TR and 7900X both support "true PCIe 3.0x16 connections directly from the CPU", does that mean even the TR has more lanes, there is no advantage over the 7900X in terms of performance to move large amount of data between the GPUs and the CPU?
BFG-9000
Splendid
- Sep 17, 2016
- 4,252
- 810
- 29,590
It definitely depends on the motherboard too. With Threadripper usually 48 lanes are dedicated to x16 slots and 12 to x4 NVMe. MSI chose to provide the usual three NVMe and 4 x16 PCIe slots so had to split the last 16 lanes in two to x8/x8. It is possible for a motherboard manufacturer to instead wire only the 4th x16 slot to x8 but supply just one NVMe x4.
For 7900x, both ASUS ROG Rampage VI Extreme and MSI X299 XPower Gaming AC claim x16/x8/x8/x8 with a 44-lane CPU. WS X299 Sage must have a switch chip because it can run seven GPUs.
If you buy 8700k now, then in a few months you can upgrade to 9700k Cannon Lake with 2 more cores by throwing away your motherboard and CPU. Which still won't get you AVX512 VNNI because that's not scheduled to arrive until Ice Lake, which will probably require yet another motherboard. But at least you can keep that DDR4... unless Ice Lake requires DDR5!
For 7900x, both ASUS ROG Rampage VI Extreme and MSI X299 XPower Gaming AC claim x16/x8/x8/x8 with a 44-lane CPU. WS X299 Sage must have a switch chip because it can run seven GPUs.
If you buy 8700k now, then in a few months you can upgrade to 9700k Cannon Lake with 2 more cores by throwing away your motherboard and CPU. Which still won't get you AVX512 VNNI because that's not scheduled to arrive until Ice Lake, which will probably require yet another motherboard. But at least you can keep that DDR4... unless Ice Lake requires DDR5!
Thanks. What advantage do I get in regard to more PCIe3.0x16 lans by having two more cores in Cannon Lake?
Do you know how many true PCIe3.0x16 are connected directly to the CPU in the case of Gigabyte X299 Aorus Gaming 9 and Gigabyte X399 Aorus Gaming 7?
Do you know how many true PCIe3.0x16 are connected directly to the CPU in the case of Gigabyte X299 Aorus Gaming 9 and Gigabyte X399 Aorus Gaming 7?
Hello, I found the manual for the WS X299 Sage. I need some expert opinions. Please refer to:
http://dlcdnet.asus.com/pub/ASUS/mb/Socket2066/WS_X299_SAGE/Manual/E14028_WS_X299_SAGE_UM_V2_WEB.pdf?_ga=2.259423439.435600446.1523725088-932830197.1519525979
On p.119, there is a block diagram. I noticed the following:
- None of the PCIe X16 slot is connected directly to the CPU.
- They seem to operate at 100MHz PCIE 16x/8x
1. Does that mean by adding two PLX chips, none of the PCIe x16 card/slot can operate as fast as if they were directly connected to the CPU?
2. How efficient is that PEX8747 PLX chip? Will the latency affect performance?
3. If I use only 1 or 2 GPUs operating at PCIe 3.0x16x16, do I gain better performance using this MB or standard MB likes Gigabyte X299 Aorus Gaming 9, Asus X299 Prime-A or MSI X299 SLI Plus MS-7A93 in the case of using 7900X and Gigabyte X399 AORUS Gaming 7, MSI X399 Gaming PRO Carbon AC or Asus Prime X399-A?
Am I right that I can only plug in one NVMe SSD M.2 to this motherboard? I may need two SSD.
Thanks.
http://dlcdnet.asus.com/pub/ASUS/mb/Socket2066/WS_X299_SAGE/Manual/E14028_WS_X299_SAGE_UM_V2_WEB.pdf?_ga=2.259423439.435600446.1523725088-932830197.1519525979
On p.119, there is a block diagram. I noticed the following:
- None of the PCIe X16 slot is connected directly to the CPU.
- They seem to operate at 100MHz PCIE 16x/8x
1. Does that mean by adding two PLX chips, none of the PCIe x16 card/slot can operate as fast as if they were directly connected to the CPU?
2. How efficient is that PEX8747 PLX chip? Will the latency affect performance?
3. If I use only 1 or 2 GPUs operating at PCIe 3.0x16x16, do I gain better performance using this MB or standard MB likes Gigabyte X299 Aorus Gaming 9, Asus X299 Prime-A or MSI X299 SLI Plus MS-7A93 in the case of using 7900X and Gigabyte X399 AORUS Gaming 7, MSI X399 Gaming PRO Carbon AC or Asus Prime X399-A?
Am I right that I can only plug in one NVMe SSD M.2 to this motherboard? I may need two SSD.
Thanks.
- Dec 26, 2012
- 30,759
- 4,553
- 142,790
You are obsessing over stuff that makes less than a fraction of a milli-second difference. The Asus WS is as good as it gets for high end, single cpu workstation type mobo's, so everything you are researching is pretty much guaranteed to have been researched by the bright-sparks over in the Asus Think-tank. If it was any kind of relevant performance hit, it wouldn't be on that board at all.
As is, you'll not be running all 4 of those gpus to 100%, even mining extremists can't as they are limited mostly to x1 or x4 at absolute best.
I see it as nothing more than wasting gas running all over town looking for the gas station that's 1¢ cheaper per gallon...
Why need 2x NVMe? A 1-2Tb 960 Pro is a massive amount of space to do anything with, within the boundaries of multiple programs. You could even use a regular Sata3 256Gb ssd as a scratch disk, with an external Nas for mass storage and backup.
As is, you'll not be running all 4 of those gpus to 100%, even mining extremists can't as they are limited mostly to x1 or x4 at absolute best.
I see it as nothing more than wasting gas running all over town looking for the gas station that's 1¢ cheaper per gallon...
Why need 2x NVMe? A 1-2Tb 960 Pro is a massive amount of space to do anything with, within the boundaries of multiple programs. You could even use a regular Sata3 256Gb ssd as a scratch disk, with an external Nas for mass storage and backup.
Thanks. I do think that I am getting obsessive. Too bad the TR cannot do even x16x16x16 despite its relatively larger number of lanes compared with the i9. The advantage seems to be just the cost.
Looks like not many people are using the Asus WS so getting support might be an issue. How is support from Asus?
Based on all these discussions, what is your personal view on choosing between 7900X and 1900X? Each platform has pros and cons.
I plan to use the workstation to do AI research and process big data. A colleague recommended getting two NVMe SSDs one for storing the Linux OS and other programs. The other for data.
Looks like not many people are using the Asus WS so getting support might be an issue. How is support from Asus?
Based on all these discussions, what is your personal view on choosing between 7900X and 1900X? Each platform has pros and cons.
I plan to use the workstation to do AI research and process big data. A colleague recommended getting two NVMe SSDs one for storing the Linux OS and other programs. The other for data.
BFG-9000
Splendid
- Sep 17, 2016
- 4,252
- 810
- 29,590
Sigh, it's obvious you aren't reading the answers in this very thread.
Any switch chip will provide full performance if only one device is attached at each end. Just as if your gigabit switch had only two things plugged in, everything runs at full line speed. More than that and they will need to share bandwidth if used simultaneously.
The point is not if dropping to x8 or sharing bandwidth reduces performance, but whether adding more cards anyway still improves things.
X299 can do x16/x8/x8/x8 in most boards but you may find one that can do x16/x16/x8
Threadripper can do x16/x16/x16 just fine, you just need to either find a X399 board that does this if nothing is plugged into the 4th slot, or else use a board with only 3 slots like the ASRock X399M Taichi (that one's mATX so you'd need a x16 extender to run a 3rd double-width card).
Intel can do x16/x16/x16 too if you use a 48-lane CPU like Skylake-W (Xeon W-2135, 2145 or 2155) on a C422 chipset board such as the Gigabyte MW51-HP0.
If you want to use more cards than that all at x16, then like I said you'd need multiple Skylake-SP Xeon processors or a single Epyc.
All of these Intel chips can do some form of AVX512 but honestly if your application can run in CUDA it would probably be faster running in the GPUs. FWIW, if your application can run in GPUs then CPU speed and cores should hardly matter, and the 4-core Xeon W-2125 only has a MSRP of $444 which is about the same as the current price of the 8-core TR 1900x. So for x16/x16/x16 you can choose between AVX512 and higher single-thread performance, or twice as many cores with three times as many NVMe. You could always add money to either platform to get more cores too.
Any switch chip will provide full performance if only one device is attached at each end. Just as if your gigabit switch had only two things plugged in, everything runs at full line speed. More than that and they will need to share bandwidth if used simultaneously.
The point is not if dropping to x8 or sharing bandwidth reduces performance, but whether adding more cards anyway still improves things.
X299 can do x16/x8/x8/x8 in most boards but you may find one that can do x16/x16/x8
Threadripper can do x16/x16/x16 just fine, you just need to either find a X399 board that does this if nothing is plugged into the 4th slot, or else use a board with only 3 slots like the ASRock X399M Taichi (that one's mATX so you'd need a x16 extender to run a 3rd double-width card).
Intel can do x16/x16/x16 too if you use a 48-lane CPU like Skylake-W (Xeon W-2135, 2145 or 2155) on a C422 chipset board such as the Gigabyte MW51-HP0.
If you want to use more cards than that all at x16, then like I said you'd need multiple Skylake-SP Xeon processors or a single Epyc.
All of these Intel chips can do some form of AVX512 but honestly if your application can run in CUDA it would probably be faster running in the GPUs. FWIW, if your application can run in GPUs then CPU speed and cores should hardly matter, and the 4-core Xeon W-2125 only has a MSRP of $444 which is about the same as the current price of the 8-core TR 1900x. So for x16/x16/x16 you can choose between AVX512 and higher single-thread performance, or twice as many cores with three times as many NVMe. You could always add money to either platform to get more cores too.
TRENDING THREADS
-
-
-
Discussion What's your favourite video game you've been playing?- Started by amdfangirl
- Replies: 4K
-
-
Question Converting a laptop hard drive into an external hard drive
- Started by hack3rcon
- Replies: 15
Facebook
Twitter
Instagram