It's time for the PCPer Mailbag, our weekly show where Ryan and the team answer your questions about the tech industry, the latest and greatest GPUs, the process of running a tech review website, and more!
This week, our very own Josh Walrath tackles your questions about process tech, racing wheels, and AMD:
00:31 – PCIe 4.0 for Volta?
02:10 – Intel beats AMD to HBM APU?
02:59 – What process tech gets Josh all excited?
04:25 – Significance of discrete GPU sales to AMD's bottom line?
06:09 – New products still using old process nodes?
07:50 – CPU redundancy?
10:08 – Racing wheel recommendations under $200?
12:07 – Direct drive wheels?
13:31 – Frequency and IPC improvements for Zen+?
16:06 – Josh takes over PCPer?
Be sure to subscribe to our YouTube Channel to make sure you never miss our weekly reviews and podcasts, and please consider supporting PC Perspective via Patreon to help us keep videos like our weekly mailbag coming!
Intel beats AMD to HBM
Intel beats AMD to HBM APU?
That’s no APU that’s an Intel SOC with it’s not so performant integrated graohics wired up to a discrete semi-custom AMD Radeon GPU die via PCIe traces on the EMIB/MCM. And the only part of the EMIB getting any silicon interposer etched traces is the semi-custom Radeon die interface to its HBM2 stack. So AMD and Intel’s relationship as far as graphics is concerned is no different that any of AMD’s relationships to any other AIB discrete GPU partners. Intel is just purchasing semi-custom Radeon dies and creating a wee little nano-motherboard arrangement for some space savings and some better graphics via that semi-custom Radeon graphics that Intel is purchasing directly from AMD.
The only interesting IP is the EMIB’s embedded smaller silicon interposer that is placed in the EMIB’s organic substrate and etched with those 1024 bit memory traces to the HBM2 with only the interface parts of the Radeon Die and the HBM2 die stack interfaced via a smaller sliver of silicon interposer materal(Bridge Chip). The ramainder of the interfacing Between the Intel SOC(And its Intel integrated graphics) and the discret Radeon Die is some via regular PCIe traces in the EMIB’s organic(PCB Like) substrate that is done the old fashon way. Intel’s SOC is still interfacing to the Discrete Radeon die via PCIe traces on that EMIB.
So no APU just a daughter card MCM stacked/interfaced mezzanine like on a larger Motherboard in that Intel NUC image posted online. Intel gets no Radeon IP license in the deal Intel only sources the Discrete Radeon Die like any other AIB partner.
I guess if Apple wants to save space and Intel will have this EMIB and discrete Radeon die offering available to keep Apple’s business and if AMD ever brings to market any Raven Ridge APU SKUs with a single stack of HBM2 then Intel is SOL because AMD’s Raven Ridge APU graphics interfaces CPU core to GPU cores via the Infinity Fabric and not PCIe. And All AMD needs to do is get a interposer that’s big enough to host the APU’s monolithic die and a single stack of HBM2 and that interposer will be smaller than AMD uses for its Discrete GPU offerings so it’s probably the next step for AMD and some high end Ryzen APU offerings for the Desktop/Mini-Desktop APU market.
The question was phrased with
The question was phrased with “APU” in quotes so they did know which product was being referred to.
The main purpose likely is for Intel to get familiar with a good GPU product as they realize that GPU’s are the future and CPU’s are going to become increasingly less relevant.
Intel hired Raja from AMD so it will be interesting to see where Intel is concentrating their research efforts. Probably minimal to no effort in the higher-end desktop market but rather the AI, server etc markets.
The Intel + AMD EMIB did surprise me though, because I thought AMD would not be interested and instead just use their own APU in place of the Intel CPU (so APU + AMD dGPU). Obviously there were good business reasons to create this product that currently benefit AMD and Intel both, so it will be interesting to see what happens in the future.
Intel’s had a strange and expensive path trying to break into mobile. They spent billions trying to get x86 more efficient to combat ARM, but just like their past missteps I wonder where the EMIB approach would actually make them much money… how much demand is there for mobile devices with x86 CPU’s and good dGPU’s really?
AMD has some EXCELLENT GPU and CPU products now that suffer mostly from lack of great driver, compiler and other software support.
This was exactly what I
This was exactly what I meant. Though I was also wondering about the business case of helping out a competitor (even if you’re getting paid) versus being first to market yourself.
So Josh
the Reverse of
So Josh
the Reverse of Hyperthreading/SMT, will we ever see it exist? 2/ cores -> 1 thread for improved IPC? Suppose you’d have to be able to predict the future to do it. Or like have the CPU cores somehow know what the other was working on in memory on alternating cycles?
Intel bought this company a while back that I think was working on something like it.
https://www.theregister.co.uk/2016/09/09/intel_soft_machines/
And here’s a fake Ryzen Slide for a potential name
https://i.imgur.com/ZWmbvoB.png
Yes they did but was Soft
Yes they did but was Soft Machines IP really that great and if so why did Soft Machines have so much trouble attracting investment. Only time will tell and Intel may just sit on the IP until market forces(Competition) really forces Intel’s hand. Intel is not in the technology business Intel is in the Technology drizzled out slowly for more profits business! So unless Intel’s back is against the wall then Intel will keep new technology introductions to a minimum and Milk, Milk, Milk that IP for all its worth!
That’s one way to look at it
That’s one way to look at it – another way is they’re maximizing profits, a practice every board member and stockholder really insists upon. Also, getting too far ahead is not good, you would essentially be competing with yourself, and causing your recent products to become too quickly obsolete. And too much success causes antitrust problems as well, which Intel has been burned by more than once ($1.25 billion to AMD, $1.45 billion to the EU). If Intel suddenly invented a new process that could create CPUs with 10 GHz clock speeds, their only safe bet would be to sit on it until some other company was getting close, then magically come up with it. And sell it at a really high price (like they already do with flagship parts). Like the whole Ryzen situation that still has fools declaring how scared Intel must be to react so “radically”. Yes, business is all about making money, not pleasing some random bitchy nerd on the internet. Hard to believe that anyone is foolish enough to think otherwise…
And don’t forget that each
And don’t forget that each time Intel gets some competition that it can’t deal with it just pays the OEMs to ignore AMD’s better product. So those antitrust problems where not the result of any real succsess on Intel’s part. And you Peckerwoods think that profits above all else matters even if laws are broken to get the money made.
You are a classic Crony Capitalist and Intel has held back innovation long enough. AMD’s real mistake is focusing too much on the x86 ISA and not enough on other ISAs like the RISC ISAs that are better for the low power markets. So AMD looks to be becoming a smaller Intel with that crappy x86 lagacy bloat of an ISA when AMD should be focusing on better solutions for mobile that colud include that K12 custom ARM core paired with AMD’s graphics IP.
Intel has been locked out of the Phone and mainstream tablet market and AMD needs to focus more on becoming a company that is not tied exclusively to the x86 ISA as there will be Power9 competition in that server/HPC market also. Intel is not going to remain relevent much longer as will the x86 ISA and the mobile market. It looks like AMD is resting to much on that x86 ISA when it should also be focusing on some of its ARM based IP. AMD needs to get into new markets with a different RISC based Mobile Friendly ISA and pair that ISA with its graphics IP.
Sure Ryzen is out there but the x86 market does not have the majority of devices market that folks make use of on an everyday on an all the time basis like the smartphone market and the mainstream tablet market. And Google will be making use of its own Power9 based internal server product so that’s just more x86 ISA shrinkage in the larger world market of devices not tied to the x86 ISA. AMD really needs to stop focusing on markets that are just shrinking and any market share gained is from a fixed or shrinking market like the PC market or discrete GPU market where AMD has to battle it out over a shrinking x86 or discrete GPU pie with Intel and Nvidia and start thinking more about using s RISC ISA for the lowest power using phone/tablet market.
AMD needs to focus on the GPU AI/Blockchain(Not necessarly coin) markets and the professional GPU compute markets and also some of the RISC ISA based CPU markets where AMD’s Graphics IP may become useful paired with a custom ARM CPU core/s APU design. The x86 server HPC market is also going to see competition from the OpenPower market licensees and Google is just one of those Licensees and that means that AMD better stop trying to be a smaller version of Intel and only be tied to that mess of an ISA the x86 ISA because the HPC server markets run on the Linux OS and Linux already has that RISC market from ARM phones to Government supercomputers running via the ARM RISC ISA and the Power9/Power9 RISC ISA. And there are some exaflop supercomputers planned using the ARMv8A/newer RISC ISA.
If AMD dropps its K12 project that will just be conformation that AMD is not serious about the entire devices market as the x86 ISA based market is dwarfed by the RISA ISA based market on a total devices basis. And AMD can not afford to tie itself to a any shrinking or stagnent markets in the long run. I’m sure that Apple is planing to continue to beef up its ARM designs and the at some point in time port MacOS fully onto some of its ARM ISA based products and Apple can really move away from any dependency on the x86 ISA market.
Vega is fine for non gaming graphics workloads and compute but AMD needs to take those Vega cUCs and the rest of the Vega IP and tape out a new Vega gaming focused base die design with a little more ROPs and a little less shaders in the mix. Vega’s TMU/Texture resources are already enough to compete with Volta, and Nvidia beats AMD everytime on that ROP count metric that allows Nvidia to win the FPS contest.
I can see why AMD lost Raja because too much of that AMD as an x86 CPU only company remains at AMD and Intel sure will make us of Raja for some compute/AI focused product to compete with Nvidia’s GV100 compute/AI products and Nvidia already has an inside deal with IBM and some power9 products so Nvidia sure is not thinking about x86 in the long run either.
Josh seemed drunker than a
Josh seemed drunker than a station wagon full of Indians in that segment. Sorry..Native Americans!
God knows we love him though!
God knows we love him though!
I… wasn’t actually drunk.
I… wasn't actually drunk.
question for future
question for future mailbag:
If I want to build a system with an Intel i7-7700K or 8700K and 64GB of RAM, which configuration will provide the best performance:
1 stick of 64GB
2 sticks of 32GB
4 sticks of 16GB
Intel mainstream platforms
Intel mainstream platforms are dual channel, so performance scales up to two sticks. No performance benefit of using 4.
True, though in some
True, though in some situations you may benefit above 3200MHz CAS15.
If you need 64GB then it’s likely you are doing video editing or something similar, so it’s hard to say how much benefit you’d get above a certain bandwidth as it’s very, very application specific.
You can also put stress on the CPU memory controller if you start going above 1.35V on the DDR4 memory so I’d avoid higher bandwidth/higher voltage kits in general unless benchmarks justify the small gain to you.
As said, the i7-8700K motherboards are Dual Channel so even if it makes sense to use all FOUR slots (i.e. 4x16GB) that’s still Dual Channel as the data is split between PAIR A, or PAIR B and is NOT split between four sticks.
So again, for an i7-8700K you should probably stick with Dual Channel 3200MHz.
Other:
QUAD CHANNEL setups would be for 10+ fast cores typically where the amount of data is sufficiently high to require a high-speed bandwidth between the CPU and system memory.
You don’t tend to see Quad Channel support on a motherboard unless it supports CPU’s that could benefit from this (and most probably have EIGHT slots in total I’d guess to support at least 128GB system memory).
Question for the mailbag:
I
Question for the mailbag:
I noticed nVidia’s V100 and Titan V use 3-4 HBM2 chips while AMD’s products use 2 memory chips, effectively giving AMD lower bandwidth despite the same total memory size. As memory bandwidth is the major selling point of HBM, any guesses why AMD’s products are apparently limited in this way?