If another company would buy AMD, AMD will lose X86 license from Intel.
Got a tip for us?
Let us know
Become a MacRumors Supporter for $50/year with no ads, ability to filter front page stories, and private forums.
MP 7,1 Waiting for Mac Pro 7,1 A1991 (no more)
- Thread starter Mago
- Start date
-
- Tags
- macpro 2016
- Sort by reaction score
If another company would buy AMD, AMD will lose X86 license from Intel.
Assuming Apple would buy AMD for Ryzen. I imagine if they did buy AMD, it would possibly be for other assets and they would continue to source Intel CPUs.
I beg you, please, PLEASE post this paragraph in a linux forum and send us the link. Also make sure that you get it cleared that you talk about apple and how big open standards benefactors they are. People there will take care of the rest.
The deal invokes possibility for AMD to use Intel fabs for next generation products.
P.S. Next generation graphics chips from Intel can be HSA compatible. To the degree that there will not be a problem with pairing AMD GPU with Intel CPU to get complete coherency.
P.S.2. This makes HSA software much more possible to be popularized...
Intel do support hUMA with their iGPU's or SoC's. But, what has been missing is opening hUMA for external dGPU's.
Last edited:
Why should Intel build something to be HSA compatible? The only reason for Intel to follow a standard that they did not invent (and thus do not control) is when they have lost a battle and are forced to adopt a standard. Like x86_64 from AMD, which effectively killed ia64.
But HSA is not very common at the moment. Intel could come up with its own flavour of HSA and push it to the market with its sheer sales volume. There would be so many Intel-HSA systems out there, that AMD HSA simply has no chance. And remember, it's not the best technology that wins the market, it's mostly the cheapest and widely available technology that wins the market. Just see how Intel more or less killed the low-budget GPU market with their internal GPU offerings.
Besides, HSA would move computing from the CPU to external computing units (i.e. external to the CPU). That's not in Intel's business interest. They make money by selling CPUs with a huge margin. If the CPU becomes less relevant, people would start to buy smaller CPUs and more external computing units from other vendors. Which means less revenue for Intel. Nothing Intel prefers to happen anytime soon … I would even say Intel tries to steer the industry away from this.
All I can say is this: observe server and enterprise market deals, and "shakeups".
errrr. HSA is implemented by more than just AMD. Trying AMD , ARM , Imagination Tech , Media Tek, Samsung, and Qualcomm.
http://www.hsafoundation.com/members/
In terms of deployed numbers of Intel PCs this is about an order magnitude larger. I'm not sure that doesn't meet 'common' litmus test. AMD stopped trying to solely 'hog' HSA years ago. Around the same time they didn't hog Mantle. The spread of Vulkan among the ARM vendors (major ones listed above) all got on board with HSA about the same time.
And smartphones don't outnumber classic PCs ? The vast majority of deployed processors these days have integrated GPUs.
Which is a major point. If going to have a SoC which consists of a CPU and a GPU on the same die, splitting up the work is something that comes in very handy.
Hooey. Intel needs to find enough valuable products to keep their fabs filled. They spent several billion on Altera. Farming out a narrow subset of computations to specialized FGPA and keeping mainstream "test and branch" high level control on the Intel CPU works. HSA ran a bit ahead of OpenCL's shared memory model but it is not necessarily in conflict with it. SPIR, Vulkan, and OpenCL are heading in same direction. Most of the folks in HSA haven't done a full HSA software stack, but to some extent they aligned enough of the hardware so that the future features the Khronos stack is heading toward.
There are factions in Intel who might want things to remain as balkanized. Nvidia pulling on the same strings. But cache coherent interconnects like Intel OmniPath demonstrates that the whole company is not off chasing Moby Dick and the past.
HSA may disappear and get subsumed into OpenCL/Vulkan, but "no chance because outnumbered"; not really matched up with who is and isn't signed up for HSA.
[doublepost=1494964184][/doublepost]
As a mandated standards requirement that all of these be present? It isn't technically impossible. However, it is extremely unlikely to ever pass as a standard. So pragmatically no.
It isn't that you can't slap alternative modes on. Where the "one port for everything , everywhere" is in the implementation of those modes. In the "race to the bottom" ports there will be high pressure to cut optional stuff loose to trim costs. That means the port want do those things everywhere. Hence the fail.
This isn't really about dedicated ports. It is about the universal port being something more than simply just mechanically the same. The data that traverses the port would have to be universal also for it to be really universal. As long as have to ride on USB IF controlled port it is almost certain that there will be competition as to what is universal on the port ( USB 2.0 is mandated always) and the other stuff.
Related. AMD on their Analyst conference have just said that Datacenter is the main growth opportunity for them, and that they see that Datacenter can really benefit from HSA.
Koyoot implied that Intel might pick up HSA for their CPUs. I simply suggested that this is unlikely. I think the most obvious explanation - that Intel needs the patents because their licensing agreement with Nvidia did end - is the most plausible explanation, too. Besides, I commented on some of the financial aspects.
Since Intel has no smartphone chips (at least none of any relevance or market penetration), the HSA hardware and software in the smartphone segment are not Intel's concern - or you could say Intel has no saying in this sector. So anything you say in this regard may be right, but it does not matter to the question about HSA support in Intel's desktop and server CPUs.
Because Ryzen is enough big threat to Intel CPUs. Secondly support for HSA can kill Nvidia, and that is what both: Intel and AMD would want.
That is All I want to say at this moment about this.
It could start to make sense for Apple to consider Epyc for the mMP. It might be possible to do a modular system with an Infinity Fabric backbone, then you can slot in 1 or 2 CPUs (CPU modules) and maybe it would be possible to adapt the GPUs to Infinity Fabric, since it seems to be PCIe based anyway (but 64 lanes?).
This would make it really modular, and you could actually configure it as you wish, dual CPUs or dual GPUs, or both even.
Best of both worlds.
This would make it really modular, and you could actually configure it as you wish, dual CPUs or dual GPUs, or both even.
Best of both worlds.
Epyc are server CPUs not airmed at WorkStations, as a server CPU with 32 cores it need to throtle the cores to keep a safe temperature, I bet what ever you want its single thread execution beats nothing currently being produced, in WS the single thread speed its mandatory since not few process are single threaded (as program compiling, some CAD operations), in WS has no sense having more than 16 cores (unless you develop SMMP).
More sense should had the Ryzen Thread Ripper, but I consider it very unlikely to come to the mMP.
becoming a standard Intel PC was the most important beneficial thing that happened to the Mac. No matter how good the PowerPC macs were, the advantage of being a "no downside" option (aside from cost, and then only by virtue of a lack of low end) to an HP or Dell or Sony was the single most important thing for expanding the mac outside its existing base.
Even if Intel means progress is not as fast as some people would like - difference from what the rest of the market uses has to be explained and justified. better versions of what the rest of the market has are just "it's a better computer because..."
Mago
Epyc would be nice because of the 2S possibility, as many folks here crave for it. That's it.
TR is HEDT, I doubt Apple would use it, and no 2S there.
I don't need dual CPU, but some here do, that was my point.
Maybe lower core count SKUs.
Epyc would be nice because of the 2S possibility, as many folks here crave for it. That's it.
TR is HEDT, I doubt Apple would use it, and no 2S there.
I don't need dual CPU, but some here do, that was my point.
Maybe lower core count SKUs.
Vast majority of folks yelping about two socket really want > 12 cores and more than four DIMM slots. ThreadRipper would likely deliver that ( and ECC if like the Ryzen models have it turned on by default). A single Epyc even more so and blow away any PCI-e lane shortage. There number who need 40+ threads and are single user workload is small (and in may cases better served by a "x86 core farm" in some other room than sitting beside or on their desk. )
What AMD has done is add even more weight to the scale toward the one socket solution.
On the AMD Epyc marketing pages AMD has linked a study about one socket servers.
http://www.amd.com/system/files/2017-05/TIRIAS-AMD-Single-Socket-Server.pdf
The major take away is that 30-40% of the sub $10K , but in the Mac Pro's price range, two socket servers sold have only one CPU in them. ( the lower prices ones fell to 20%. Probably because filling with stuff like E5 2603 .. the very low clocked stuff that the Mac Pro would never use. )
15-30% of the 2 socket market can be taken by a very good 1 socket system if just don't let folks buy the empty socket they are unlikely to fill. There is always a fraction that will pay for "future proof" that they will probably never use.
An Intel 12 core or AMD 16 core with 8 DIMMs slots would skim off a sizable chunk of the 2 socket crowd.
Yeah, but how many of us here really have or need dual socket systems? And I'm not saying want to have or would like to, but really need and use it to the fullest? Really interested, not just picking.
[doublepost=1495130355][/doublepost]And Epyc might come in 16 and 24 cores flavor as well, it's rumored.
[doublepost=1495130451][/doublepost]Xeon SP refresh for 2018 will be named Cascade Lake.
[doublepost=1495130355][/doublepost]And Epyc might come in 16 and 24 cores flavor as well, it's rumored.
[doublepost=1495130451][/doublepost]Xeon SP refresh for 2018 will be named Cascade Lake.
IMO, the mMP is only relevant to the folks for whom the top trim iMac is still weak sauce for their needs. Within that cohort, is a 2S topology the only way to satisfy their demands? I think a lot of us would be fine with a 1S build as long as there were enough PCIe lanes to support at least 2 bomber GPUs and plenty of data I/O - especially if it meant a higher clock rate.
Especially in 2017, there is a case to be made for forking heavy duty processing tasks into two categories:
1) can be completed in a few hours or overnight on your local mMP
2) needs a render farm or dedicated box to get done by deadline
I would argue for a WS that can hold it's own on tasks that would melt an iMac, but not so bomber that it costs a fortune and needs it's own 20 amp circuit.
What would really rock, is a slick peer to peer networking (resource sharing) protocol that could efficiently exploit multiple mMPs as an ad hoc render farm. Let's say the mid level mMP is a $5,000 1S with a credible GPU, top level $10,000 fully loaded with at least 2 beast GPUs. For some folks the loaded machine is just what the doctor ordered, rock on. That said, for me, and I'd guess many others, I'd rather have more mid-level (this is mid-level based on next gen components, not current h/w) machines that could service more users when needed and also help crunch big jobs faster the rest of the time.
Bottom line - I want a mMP that takes advantage of the latest tech/parts in their cost/benefit "sweet spot" and an easy to implement ad hoc render farm protocol so one can hook several mMPs together to tackle monster renders/etc.
Especially in 2017, there is a case to be made for forking heavy duty processing tasks into two categories:
1) can be completed in a few hours or overnight on your local mMP
2) needs a render farm or dedicated box to get done by deadline
I would argue for a WS that can hold it's own on tasks that would melt an iMac, but not so bomber that it costs a fortune and needs it's own 20 amp circuit.
What would really rock, is a slick peer to peer networking (resource sharing) protocol that could efficiently exploit multiple mMPs as an ad hoc render farm. Let's say the mid level mMP is a $5,000 1S with a credible GPU, top level $10,000 fully loaded with at least 2 beast GPUs. For some folks the loaded machine is just what the doctor ordered, rock on. That said, for me, and I'd guess many others, I'd rather have more mid-level (this is mid-level based on next gen components, not current h/w) machines that could service more users when needed and also help crunch big jobs faster the rest of the time.
Bottom line - I want a mMP that takes advantage of the latest tech/parts in their cost/benefit "sweet spot" and an easy to implement ad hoc render farm protocol so one can hook several mMPs together to tackle monster renders/etc.
Not ridiculous at all:
- There is a lawsuit filed that is seeking class-action status
- Apple themselves are aware that they have defective GPUs
- The VFX team for Deadpool said that Mac Pros burned out regularly and had to be routinely replaced
- There is a problem GPU thread that has now reached over 1000 posts
- This guy has had four fail in a row
- I can't find it now, but one of the guys here bought 12 Mac Pros and experienced GPU failures on 11 of them
Last edited:
I've lost a ton of time due to nMP GPU issues - we ended up rendering everything out to an image sequence so that we could at least get rid of the glitched frames before encoding to ProRes, but what a garbage way to have to work.