Array
(
    [content] => 
    [params] => Array
        (
            [0] => /forum/index.php?threads/apple-introduces-m2-ultra.18108/
        )

    [addOns] => Array
        (
            [DL6/MLTP] => 13
            [Hampel/TimeZoneDebug] => 1000070
            [SV/ChangePostDate] => 2010200
            [SemiWiki/Newsletter] => 1000010
            [SemiWiki/WPMenu] => 1000010
            [SemiWiki/XPressExtend] => 1000010
            [ThemeHouse/XLink] => 1000970
            [ThemeHouse/XPress] => 1010570
            [XF] => 2021370
            [XFI] => 1050270
        )

    [wordpress] => /var/www/html
)

Apple introduces M2 Ultra

Daniel Nenni

Admin
Staff member

With a faster CPU and GPU, plus support for even more unified memory, it takes Mac performance further than ever
M2 Ultra is the largest and most capable chip Apple has ever created. (Photo: Business Wire)
M2 Ultra is the largest and most capable chip Apple has ever created. (Photo: Business Wire)

June 05, 2023 01:17 PM Eastern Daylight Time
CUPERTINO, Calif.--(BUSINESS WIRE)--Apple® today announced M2 Ultra, a new system on a chip (SoC) that delivers huge performance increases to the Mac® and completes the M2 family. M2 Ultra is the largest and most capable chip Apple has ever created, and it makes the new Mac Studio™ and Mac Pro® the most powerful Mac desktops ever made. M2 Ultra is built using a second-generation 5-nanometer process and uses Apple’s groundbreaking UltraFusion® technology to connect the die of two M2 Max chips, doubling the performance. M2 Ultra consists of 134 billion transistors — 20 billion more than M1 Ultra.

Its unified memory architecture supports up to a breakthrough 192GB of memory capacity, which is 50 percent more than M1 Ultra, and features 800GB/s of memory bandwidth — twice that of M2 Max. M2 Ultra features a more powerful CPU that’s 20 percent faster than M1 Ultra, a larger GPU that’s up to 30 percent faster, and a Neural Engine that’s up to 40 percent faster.1 It also features a media engine with twice the capabilities of M2 Max for blazing ProRes® acceleration. With all these advancements, M2 Ultra takes Mac performance to a whole new level yet again.

“M2 Ultra delivers astonishing performance and capabilities for our pro users’ most demanding workflows, while maintaining Apple silicon’s industry-leading power efficiency,” said Johny Srouji, Apple’s senior vice president of Hardware Technologies. “With huge performance gains in the CPU, GPU, and Neural Engine, combined with massive memory bandwidth in a single SoC, M2 Ultra is the world’s most powerful chip ever created for a personal computer.”

Industry-Leading UltraFusion Technology
M2 Ultra is built from two M2 Max dies connected through UltraFusion, Apple’s industry-leading, custom-built packaging technology. UltraFusion uses a silicon interposer that connects the dies with more than 10,000 signals, providing over 2.5TB/s of low-latency interprocessor bandwidth.

UltraFusion’s architecture enables M2 Ultra to appear as a single chip to software. This means code doesn’t need to be rewritten to utilize the extreme performance of M2 Ultra and makes UltraFusion unlike anything else in the industry.

Unparalleled Performance and Power Efficiency
The 24-core CPU of M2 Ultra consists of 16 next-generation high-performance cores and eight next-generation high-efficiency cores, delivering up to 20 percent faster performance than M1 Ultra. With Mac Studio powered by M2 Ultra, colorists using DaVinci Resolve will experience up to 50 percent faster video processing compared to Mac Studio with M1 Ultra.

The GPU can be configured with 60 or 76 next-generation cores. This is up to 12 more cores and up to a 30 percent improvement compared to the incredibly powerful GPU of M1 Ultra. Rendering 3D effects using Octane on Mac Studio with M2 Ultra is up to 3x faster than Mac Studio with M1 Ultra.

Game-Changing Unified Memory Architecture
Apple’s unified memory architecture, a hallmark of Apple silicon, delivers incredible bandwidth, low latency, and unmatched power efficiency. M2 Ultra features 800GB/s of system memory bandwidth, far greater than anything found in a PC. And it can be configured with a massive 192GB of unified memory, which enables workflows not possible on a PC. For example, M2 Ultra can train massive machine learning workloads in a single system that the most powerful discrete GPU can’t even process.

Advanced Custom Technologies Supercharge Machine Learning, Video, and More
M2 Ultra integrates Apple’s latest custom technologies right on the chip, maximizing performance and efficiency:
M2 Ultra features a 32-core Neural Engine, delivering 31.6 trillion operations per second, which is 40 percent faster performance than M1 Ultra.
  • The powerful media engine has twice the capabilities of M2 Max, further accelerating video processing. It has dedicated, hardware-enabled H.264, HEVC, and ProRes encode and decode, allowing M2 Ultra to play back up to 22 streams of 8K ProRes 422 video — far more than any PC chip can do.

  • The display engine supports up to six Apple Pro Display XDR® displays, driving more than 100 million pixels.

  • The latest Secure Enclave, along with hardware-verified secure boot and runtime anti-exploitation technologies, provides best-in-class security.
Better for the Environment
The power-efficient performance of M2 Ultra unlocks new possibilities for even the most demanding pro users without sacrificing environmental responsibility. Today, Apple is carbon neutral for global corporate operations, and by 2030, plans to have net-zero climate impact across the entire business, which includes manufacturing supply chains and all product life cycles. This also means that every chip Apple creates, from design to manufacturing, will be 100 percent carbon neutral.

Mac Transition to Apple Silicon Now Complete
With M2 Ultra powering the new Mac Pro, the Mac transition to Apple silicon is now complete, revolutionizing the laptop and desktop experience. Fueled by continued innovation in Apple silicon, it’s the beginning of a new era for the Mac.

Apple revolutionized personal technology with the introduction of the Macintosh in 1984. Today, Apple leads the world in innovation with iPhone, iPad, Mac, Apple Watch, and Apple TV. Apple’s five software platforms — iOS, iPadOS, macOS, watchOS, and tvOS — provide seamless experiences across all Apple devices and empower people with breakthrough services including the App Store, Apple Music, Apple Pay, and iCloud. Apple’s more than 100,000 employees are dedicated to making the best products on earth, and to leaving the world better than we found it.

1 Results are compared to previous-generation Mac Studio systems with Apple M1 Ultra, 20-core CPU, 64-core GPU, and 64GB of RAM.
2 Results are compared to previous-generation Mac Studio systems with Apple M1 Ultra, 20-core CPU, 64-core GPU, 128GB of RAM, and 8TB SSD.
3 Testing was conducted by Apple in April 2023 using preproduction Mac Studio systems with Apple M2 Ultra, 76-core GPU, and 192GB of RAM, as well as a PC system with NVIDIA RTX A6000 graphics with 48GB GDDR6. Performance measured using select publicly available transformer models. Performance tests are conducted using specific computer systems and reflect the approximate performance of Mac Studio.

NOTE TO EDITORS: For additional information visit Apple Newsroom (www.apple.com/newsroom), or email Apple’s Media Helpline at media.help@apple.com.

© 2023 Apple Inc. All rights reserved. Apple, the Apple logo, Mac, Mac Studio, Mac Pro, UltraFusion, and ProRes are trademarks of Apple. Other company and product names may be trademarks of their respective owners.

Contacts​

Todd Wilder
Apple
wilder@apple.com
 
i noticed in Apple's yesterday event when it come to M2 computer. Apple compare M2 to Intel, not M1.
 
  • Like
Reactions: VCT
They do this because a lot of mac users are still on Intel based macs. This is simply a marketing pitch to push them to upgrade

Intel based Macs are all phased out and not available on Apple online store anymore. Apple is determined to clear out any reference to Intel processors and start a fresh new life.
 
One thing that does confuse me about the M2 Ultra... is all of the DRAM in-package as with the other M-series CPUs, or does it use external DRAM? I'm having a trouble believing 192GB of DRAM and all of that Apple silicon fits in a single package. Perhaps I'm just lacking imagination.
 
One thing that does confuse me about the M2 Ultra... is all of the DRAM in-package as with the other M-series CPUs, or does it use external DRAM? I'm having a trouble believing 192GB of DRAM and all of that Apple silicon fits in a single package. Perhaps I'm just lacking imagination.
It's all in the package. That's how they get high memory bandwidth and Unified Memory architecture (CPU and GPU sharing the memory).
 
It's all in the package. That's how they get high memory bandwidth and Unified Memory architecture (CPU and GPU sharing the memory).
I don't think they're getting better bandwidth, LPDDR5 memory only goes so fast, but if it is in-package they sure do save a lot of power and board area. (And probably cost versus a DIMM solution, just guessing.) I'm actually agreeing you that all 192GB is probably in-package, but it still seems like a lot of DRAM chips to stuff in a package with those co-joined twin big CPU dies. Hmmm.
 
I don't see how comparing a client chip to an HPC server chip is interesting.
@blueone. I upvoted. Still I think it is interesting for high-end video production at least according some bloggers I've read. I'm way out of my depth. Also Apple sells systems, not chips; And prefer high volume uses for their technology. I've often wondered if Apple's M-series could be used for high-end in-house servers. Use for cloud seems remote. I'd love to hear other thoughts on these topics.
Do you mean the R1? If so, there's very little design detail on it that I've found.
Yes I do mean R1. Thanks for catching that. Maybe Johny Srouji held a session on just that for WWDC.
 
I don't see how comparing a client chip to an HPC server chip is interesting.
Why do you think Ultra is not a server chip? Why do you think GH could not be used by the same folks who buy a Mac Pro? The only thing stopping NVidia from going after the desktop-server market is likely that every GH made is far more valuable in other SKUs. But in many ways these are comparable silicon and package sizes. Arguably the GH can do everything the M2 Ultra can do. Maybe the opposite is true, too.
 
I don't think they're getting better bandwidth, LPDDR5 memory only goes so fast, but if it is in-package they sure do save a lot of power and board area.
LPDDR has been superior perf to the sibling DDR for a while now. That 192GB is delivering 4Bps bandwidth per byte of storage. It is doing that at higher clock rate than DDR (currently 8500 vs 5200) with similar static latency but much lower dynamic latency due to much less queueing. Energy per bit in the Apple packaging will be somewhere between 4 and 10x less than DDR DIMMs. The 96 chips support 16 banks so there can be about 1,000 operations in flight, allowing the GPU and cores to run very parallel.

The only weak point of LPDDR5x will be RAS, and capacity. It only fixes single bit errors and it does not synchronously indicate uncorrectable errors. This can to some extent be offset by testing at manufacture under extremes, as is done for automobile grade, although I have never seen Apple talk about the RAS.

The only reason to use DDR is you need terabytes of memory and bulletproof RAS. 192GB of LPDDR is using 24GB stacks, which is currently the packaging limit.
 
LPDDR has been superior perf to the sibling DDR for a while now. That 192GB is delivering 4Bps bandwidth per byte of storage. It is doing that at higher clock rate than DDR (currently 8500 vs 5200) with similar static latency but much lower dynamic latency due to much less queueing. Energy per bit in the Apple packaging will be somewhere between 4 and 10x less than DDR DIMMs. The 96 chips support 16 banks so there can be about 1,000 operations in flight, allowing the GPU and cores to run very parallel.
I agree with all of these points. I was just pointing out that the memory bandwidth will be the same with DIMMs or chips in package. Bandwidth is determined by the speed of the memory channels, how many of them there are, and the speed of the memory chips.
The only weak point of LPDDR5x will be RAS, and capacity. It only fixes single bit errors and it does not synchronously indicate uncorrectable errors. This can to some extent be offset by testing at manufacture under extremes, as is done for automobile grade, although I have never seen Apple talk about the RAS.
Me either,
The only reason to use DDR is you need terabytes of memory and bulletproof RAS. 192GB of LPDDR is using 24GB stacks, which is currently the packaging limit.
Like servers do.
 
Why do you think Ultra is not a server chip?
Limited memory, networking, and connectivity. M1/M2 also has a lot of video support you don't want to waste die area or power on in a server chip. The M2 Ultra would make a decent workstation CPU, but it might not support enough memory for some use models, like video content creation. Actually, I'm still wondering if the M2 Ultra can really replace a Xeon-W for that reason.
Why do you think GH could not be used by the same folks who buy a Mac Pro? The only thing stopping NVidia from going after the desktop-server market is likely that every GH made is far more valuable in other SKUs. But in many ways these are comparable silicon and package sizes. Arguably the GH can do everything the M2 Ultra can do. Maybe the opposite is true, too.
GH uses too much power and lacks client capabilities. Also, the use of NVLINK and/or IB are useless in clients.

I'm not sure why you're beating this drum so loudly. Two great products for two completely different markets.
 
Irony: Back in 2019 Apple was rumored to be working on an R1 processor for sensor integration as a co-processor for A-series chips. https://www.macrumors.com/2019/09/09/rose-r1-coprocessor-2019-iphones/
The new R1 is somewhat similar:
  • A second R1 chip is responsible for all of the information that's coming in from the cameras, sensors, and microphones. It is able to stream images to the displays within 12 milliseconds, and Apple has said this chip is able to provide a "virtually lag-free" view of the world.
 
I agree with all of these points. I was just pointing out that the memory bandwidth will be the same with DIMMs or chips in package. Bandwidth is determined by the speed of the memory channels, how many of them there are, and the speed of the memory chips.
DDR5 DIMMs have 1 channel per 8 chips. LPDDR5 has a channel for each chip. DRAM core arrays are closely similar in each generation from each manufacturer, but the whole point of DDR, LPDDR, GDDR, HBM, is the different interfaces wrapped around the arrays for different tradeoffs when interfacing to the host. Apple has doubled down on extracting the performance potential of LPDDR, which explains a lot of their power/perf advantage.
Like servers do.
Servers typically have around 8GB per dual-thread core. The M2 Ultra has 8GB per single-thread, though it is sharing with a GPU (like future servers will need to do, as inferencing becomes ubiquitous). The remaining difference is RAS. I expect the next generation of LPDDR will go beyond single bit correction and provide better probity, at which DDR may lose a lot of its market share.
 
I'm not sure why you're beating this drum so loudly. Two great products for two completely different markets.
Only temporarily. It is a good bet that Nvidia wants to grow into server markets including workstations, and the heavily video oriented Mac Pro market is a logical place to start. I'm always looking ahead.
 
Back
Top