Slashdot Mirror
Qualcomm Eyes Intel With Centriq 2400 Arm Server Chip (eweek.com)
Qualcomm is now challenging rival Intel in the rapidly changing data center market. From a report: The company is now selling its long-awaited Centriq 2400 Arm-based server processor that is aimed at the fast-growing cloud market and that Qualcomm officials say beats Intel in such crucial areas as power efficiency and cost. Officials from Arm and its manufacturing partners have for several years talked about pushing the Arm architecture into the data center as an alternative to Intel, and some manufacturers like Cavium and Applied Micro in recent years have rolled out systems-on-a-chip (SoCs) based on the 64-bit Armv8-A design. However, Qualcomm represents the most significant Arm chip maker in terms of scale and resources to challenge Intel, which holds more than 90 percent of the global server chip market. Qualcomm's Centriq chips offer up to 48 single-threaded cores running up to 2.6GHz and are manufactured on Samsung's 10-nanometer FinFET process. The processors sport a bidirectional segmented ring bus with as much as 250G bps of aggregate bandwidth to avoid performance bottlenecks, 512KB of shared L2 cache for every two cores and 60MB of unified L3 cache. There also are six channels of DDR4 memory and support for up to 768GB of total DRAM with 32 PCIe Gen 3 lanes and six PCIe controllers. They also support Arm's TrustZone security technology and hypervisors for virtualization.
A few issues...
1. I'm assuming no VMWare? How well does Xen run on ARM?
2. Can GCC/CLANG optimize for a server profile? I'm assuming that, until now, all the work for the ARM target has been on code compactness and efficiency over performance.
3. Looks like the chip has plenty of available bandwidth, does it have the transactional horsepower to fill it?
We've seen chip makers trying to push re-purposed, low-powered chips into servers before, and the results have been underwhelming. If the raw CPU throughput is there, and a compiler/OS/server stack can be created that works well with it on existing server workloads, it may have a shot, but that's a lot of if's.
My Other Computer Is A Data General Nova III.
There has been a lot of talk about Qualcomm ARM chips taking over from Intel. The problem is when you look at the benchmarks they're rather underwhelming. Eg.
http://weborus.com/snapdragon-...
The Snapdragon 835 is a great device if you're running Android. If you're running something like Photoshop I predict performance is going to be disappointing. Microsoft's 'Windows on a Snapdragon' video shows Photoshop running. It doesn't mention performance
https://www.youtube.com/watch?...
It's the same with server stuff. And of course Intel have threatened people with a patent lawsuit on SIMD
https://newsroom.intel.com/edi...
Protecting x86 ISA Innovation
Intel invests enormous resources to advance its dynamic x86 ISA, and therefore Intel must protect these investments with a strong patent portfolio and other intellectual property rights. The following graph shows that relentless instruction set innovation translates into a deep and dynamic patent portfolio with over 1,600 patents worldwide relating to instruction set implementations.
https://imgur.com/a/x0K2V
New x86 Instructions and Related Patents
Intel carefully protects its x86 innovations, and we do not widely license others to use them. Over the past 30 years, Intel has vigilantly enforced its intellectual property rights against infringement by third-party microprocessors. One of the earliest examples, was Intelâ(TM)s enforcement of its seminal âoeCrawford â(TM)338 Patent.â In the early days of our microprocessor business, Intel needed to enforce its patent rights against various companies including United Microelectronics Corporation, Advanced Micro Devices, Cyrix Corporation, Chips and Technologies, Via Technologies, and, most recently, Transmeta Corporation. Enforcement actions have been unnecessary in recent years because other companies have respected Intelâ(TM)s intellectual property rights.
However, there have been reports that some companies may try to emulate Intelâ(TM)s proprietary x86 ISA without Intelâ(TM)s authorization. Emulation is not a new technology, and Transmeta was notably the last company to claim to have produced a compatible x86 processor using emulation (âoecode morphingâ) techniques. Intel enforced patents relating to SIMD instruction set enhancements against Transmetaâ(TM)s x86 implementation even though it used emulation. In any event, Transmeta was not commercially successful, and it exited the microprocessor business 10 years ago.
Only time will tell if new attempts to emulate Intelâ(TM)s x86 ISA will meet a different fate. Intel welcomes lawful competition, and we are confident that Intelâ(TM)s microprocessors, which have been specifically optimized to implement Intelâ(TM)s x86 ISA for almost four decades, will deliver amazing experiences, consistency across applications, and a full breadth of consumer offerings, full manageability and IT integration for the enterprise. However, we do not welcome unlawful infringement of our patents, and we fully expect other companies to continue to respect Intelâ(TM)s intellectual property rights. Strong intellectual property protections make it possible for Intel to continue to invest the enormous resources required to advance Intelâ(TM)s dynamic x86 ISA, and Intel will maintain its vigilance to protect its innovations and investments.
If Microsoft can't transform SSE instructions into an ARM SIMD instruction set due to patents on SSE, Photoshop will suck if it's run through Microsoft's x86 to ARM64 JIT engine. And the odds are something like Photoshop is using bits of SSE which are still patented and will be for some time.
Even if you don't emulate and run code nati
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;
The challenge has been performance per watt and performance per dollar is crap. Single threaded performance is complete and utter crap, and even though it's a 48 core socket, even 48 of them suck next to a comparable Intel Xeon.
If you want a whole bunch of cores and you don't care about terrible per core performance Intel can sell you a Xeon Phi. It's got a whole load of Atom cores
https://en.wikipedia.org/wiki/...
Knights Landing will be built using up to 72 Airmont (Atom) cores with four threads per core,[66][67] using LGA 3647 socket[68] supporting for up to 384 GB of "far" DDR4 2133 RAM and 8â"16 GB of stacked "near" 3D MCDRAM, a version of the Hybrid Memory Cube. Each core will have two 512-bit vector units and will support AVX-512 SIMD instructions, specifically the Intel AVX-512 Foundational Instructions (AVX-512F) with Intel AVX-512 Conflict Detection Instructions (AVX-512CD), Intel AVX-512 Exponential and Reciprocal Instructions (AVX-512ER), and Intel AVX-512 Prefetch Instructions (AVX-512PF).[69]
Is it selling well? Probably not. I.e. there's some evidence that people don't want to move from 4-8 Core i7 style large cores to 64-72 Atom cores. I.e. people buying server CPUs care about single thread performance.
Which makes me think ARM based servers are not going to kill x64 ones.
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;