Slashdot Mirror
Arch-rivals Intel and AMD Team Up on PC Chips To Battle Nvidia (pcworld.com)
Intel and AMD, arch-rivals for decades, are teaming up to thwart a common competitor, Nvidia. On Monday, the two companies said they are co-designing an Intel Core microprocessor with a custom AMD Radeon graphics core inside the processor package. The chip is intended for laptops that are thin and lightweight but powerful enough to run high-end videogames, the companies said. From a report: Executives from both AMD and Intel told PCWorld that the combined AMD-Intel chip will be an "evolution" of Intel's 8th-generation, H-series Core chips, with the ability to power-manage the entire module to preserve battery life. It's scheduled to ship as early as the first quarter of 2018. Though both companies helped engineer the new chip, this is Intel's project -- Intel first approached AMD, both companies confirmed. AMD, for its part, is treating the Radeon core as a single, semi-custom design, in the same vein as the chips it supplies to consoles like the Microsoft Xbox One X and Sony Playstation 4. Some specifics, though, remain undisclosed: Intel refers to it as a single product, though it seems possible that it could eventually be offered at a range of clock speeds. [...] Shaking hands on this partnership represents a rare moment of harmony in an often bitter rivalry that began when AMD reverse-engineered the Intel 8080 microchip in 1975.
This is what Apple should be using in future Macs. Maybe they knew of Intel plans, that's why the MacBook Air and Mac mini haven't really been updated in such a long time. It's the two Macs that will have this new CPU first.
#DeleteFacebook
has frozen over.
What's next? Sonic on Nintendo consoles? Square-Enix games on computers?
#DeleteFacebook
About 5 years ago INTC tried to buy NVDA. They had enough money to do it, and the offer was going to be reasonable, but there was a sticking point about who would become CEO of the combined company. Paul Otellini of Intel was about to step down, and the assumption from NVIDIA's Jensen Huang was he would become the CEO of the combined Intel-NVIDIA. But Intel's board wasn't going to have it and promoted Brian Krzanich to CEO instead. And that's the story of how Intel managed to lose a ton of money and missed opportunities in 3D graphics and Compute.
The best of both worlds is not a big deal when neither of the worlds is really any better than an alternative in one of the categories.
NVidia trounces both on graphics. Not to mention far more capable Linux support than either has (even if NVidia's offering is not open source, it's still pretty damn sweet).
File under 'M' for 'Manic ranting'
The linchpin of the Intel-AMD agreement is a tiny piece of silicon that Intel began talking up over the past year: the Embedded Multi-die Interconnect Bridge, or EMIB. Numerous EMIBs can connect silicon dies, routing the electrical traces through the substrate itself. The result is what Intel calls a System-in-Package module. In this case, EMIBs allowed Intel to construct the three-die module, which will tie together Intelâ(TM)s Core chip, the Radeon core, and next-generation high-bandwidth memory, or HBM2.
AMD sell Intel bare dies that talk EMIB. Interesting thing is that Intel could do a deal with NVidia to supply GPU dies which use the same interface. Well except that Intel pays NVidia licence fees whereas the AMD Intel patent licensing agreement is completely one sided - AMD pays Intel but Intel gets IP rights to anything AMD invents for free.
It's not like AMD is selling Intel a synthesizable core or even a hard macro. And Intel being Intel they probably pay people to do competitor analysis on AMD stuff anyway. So getting bare dies doesn't tell them anything that they don't already know.
And as a lot of people have noted Apple use Intel and AMD GPUs but not NVidia ones. Post Itanium I think Intel regards Apple as its non commodity low volume/high margin market.
So it all makes sense. It'll be interesting to see what the chip costs and if EMIB graphics has performance and/or power advantages over PCI Express run chip to chip. With USB you can strip out the analog transceivers for HSIC
https://www.synopsys.com/dw/dw...
Could you do something similar for PCIe? Turns out you can
http://eecatalog.com/pcie/2012...
Meanwhile, modem makers were looking for a suitable interconnect for next-generation LTE networks. These networks will have air interfaces capable of throughputs beyond the 40 MB/s typically possible with HSIC USB. Further, there was a desire to deploy other SuperSpeed applications such as mass storage in a chip-to-chip environment. The SuperSpeed Inter-Chip USB (SSIC USB) group selected M-PHY as the physical layer, and developed a reference model that bridges from the PIPE 3.0 reference model to the M-PHY physical layer. This allows existing USB 3.0 IP to be quickly adapted for SSIC USB use by deleting (or disabling) the legacy USB 2.0 support, replacing the USB PIPE 3.0 implementation with a shim plus an M-PHY implementation, and making minor changes to the link layer of the USB 3.0 IP.
In September 2012, PCI-SIG and the MIPI Alliance announced an initiative to similarly adapt PCIe to run over M-PHY. Because of the work already done by the USB-IF SSIC USB group, the adaptation will likely include a similar reference model based on PIPE 3.0, simplifying early prototyping and architectural verification.
PCIe over M-PHY is likely to be quickly accepted in the Ultrabook and x86-based tablet PC market because it will allow reuse of hardware and software IP while lowering system power requirements. Adoption may be slower in smartphones and ARM-based tablets, because thereâ(TM)s less experience in using PCIe in those systems.
Of course you could do that for PCIe run chip to chip too. Still maybe you could use lower voltages over EMIB.
echo -e 'global _start\n _start:\n mov eax, 2\n int 80h\n jmp _start' > a.asm; nasm a.asm -f elf; ld a.o -o a;
Putting aside AMD's very newest chip for a moment, there are basically three different kinds of use cases:
A) I want the best performance I can get within my $X budget.
B) it's a server serving many clients (lots of threads)
C) It's a single thread and I don't care how much it costs because I'm spending taxpayer money, I want the very fastest single-thread performance, cost be damned
Intel specializes in case C. Raw single-thread performance, cost be damned.
AMD will give you more cores for the dollar, so it competes well in case B, servers running many threads. AMD also traditionally costs significantly less, so it fits case A, getting the best CPU you can within a certain price range.
That's a generalization, though. It's best to compare one CPU model to another, evaluating based on the needs of your specific application and budget.
Makes sense. Intel graphics are still a failure.
Remember when the industry panicked when Intel bought Chips & Technologies and the Real3D patents?
That didn't go so well. Who else had a shoebox full of Intel i740 cards bought at fire-sale prices?
Kriston
Ryzen still competes on price more than performance....No Amd chip can touch the intel 8700k.
Good-bye
And many libraries I had to use were pre-compiled to generic specs rather than optimized.
Or they were compiled with icc, which adds check for whether the emitted code is running on an Intel chip, and jumps to generic unoptimized code if not. If it IS an Intel chip, the code does the usual flag checks to see which modern instructions are available and runs the appropriate code, but on non-Intel it doesn't get as far as checking flags.
Finally! A year of moderation! Ready for 2019?
> I regretted that because I found that at that time in history while some code did run equally well on these that in general the software libraries for AMD just weren't tuned as well for these chips. Many optimizations not taken.
Part of that was do to Intel's shenanigans.
Intel's "cripple AMD" function in their compiler
The BB 10 OS is derived from QNX, which BlackBerry bought when they realized their own OS was a POS that couldn't scale into the future. (Very much like Apple declaring Copland a lost cause, and started deciding between BeOS and NeXTSTEP in 1996).
QNX is one of the premier embedded RTOS's, and is used in the majority of the automotive industry, including most automobile infotainment systems.
Licensing fees from QNX is probably one of BlackBerry's largest remaining revenue sources at this point.
-- Sometimes you have to turn the lights off in order to see.
Intel gets a lot of bad press for this, but it's worth remembering why they did it: a bunch of x86 chips advertised the relevant CPUID feature flags and then either didn't implemented the instructions correctly (IDT, I'm looking at you), or (in AMD's case) implemented them entirely in slow microcode so that the fast paths ended up being slower than the versions that used the older instructions. AMD complained when they emitted code that used the newer instructions that were much faster on Intel chips than AMD, then they complained when Intel whitelisted CPUs where they'd tested that there actually was a speedup. There was basically no way for Intel to win here.
I am TheRaven on Soylent News
It's a brilliant move. If it's successful, they will control the GPU on CPU market for the 64 bit 86 platform globally. Intel still holds the majority of sales and it'll be painful to compete with them directly as Intel has enough cash to match AMD at any price point as well as enough volume and customers to out-produce and out-market AMD.... but... if AMD gets a portion of every Intel sale because of the GPU, they get a lot of cash for just a bit of support work and help to raise brand awareness for Radeon while raising their own market share of GPUs vs their more direct rival, Nvidia. Right now, Intel makes more GPUs than Nvidia and AMD combined for the consumer market and many of their on-die GPUs are good enough for most users... but, this will make them good enough for most gamers!
Long term, AMD can kill Nvidia by providing decent gaming graphics for both AMD and Intel CPUs and offering discrete graphics cards for those that they aren't good enough for.... and likely they'll come up with a method for the embedded GPU to work with the discrete GPU as a sort of crossfire-like connection. They already do this with separate AMD cards of different types. Nvidia won't be able to take advantage of that capability most likely either.
Not quite as good is still a fair assessment, though a bit complicated.
Intel has insisted on less scalable designs to avoid some dramatic latency penalties.
AMD has gotten bigger core count by going to a more scalable design, albeit with cross-complex latency penalties, which make things complicated.
Ignoring that, the AMD core isn't *quite* as good single threaded performance per watt or performance per clock, but has more cores to make up for it in the price point.
Things get more complex as you go to Epyc. Intel's offering starts to have offerings that have multiple memory controllers on a package, but still a very fast communication channel to use it. Epyc has more memory channels and PCIe lanes per socket, however they get divided up in somewhat limiting ways for memory performance and fast IO. This makes it of course very appealing for internet traffic (where the latency penalties of the design are a rounding error) and for bulk PCIe storage (where you are still *way* better than SATA or SAS). It is appealing for running many small VMs that don't mind single channel performance, but less so for big ones. Also Epyc is going up against Skylake-X, which has AVX-512 and as such has overwhelming floating point performance for workloads that can use it.
Basically, back in the Netburst days, it was a slam dunk for AMD technology. Intel did netburst and IA-64 when AMD stuck with a sane design, got hyper-transport and AMD64, basically advantaged in every single way except software ecosystem. Intel got smarter and then AMD went down Bulldozer path.
AMD versus Intel is now about the same as it was in the K6 generation, AMD good enough and as such much more cost effective, but Intel still edges them out per core and the pressure has thankfully made them move and deliver Coffee Lake.
XML is like violence. If it doesn't solve the problem, use more.