Users silicon interposer. Twice the density of anything in the industry.
10000 links, 2.5 TB/s
4x bandwdith of leading edge
114 billion transistors
800 GB/s DRAM bandwidth, 128 GB of unified memory
16 performance cores, 4 efficient cores
64-core GPU
Fun comparisons
GPU comparisons
Overall
@dylan522p any idea what silicon interposer is being used?
One of the big things with M1 Ultra is going to be that GPU with 800 GB/s of bandwidth. Assuming the GPUs on each die can work collectively on a single task.
Actually, that's really important - can the GPUs work on a single task together? or is it an Xe Link scenario?
800GB/s is also not the real story. A single core can barely access 200 GB/s, so it's a collective effort to use that bandwidth
Announcing Mac Studio
The small one is Mac Studio. There's also Mac Studio display.
Guessing on price, Mac Studio probably starts at $4k ? More?
Single aluminium extrusion -
3.7-inch by 7.7-inch by 7.7-inch
Double sided blower
Custom circular power supply
Low impedance heatsink
'Barely hear it' for most workloads
4 x TB ports
10 GbE RJ45
2 x Type-A
HDMI
2 x TB4 / Type-C 10 Gbps on front
SDXC port
Supports high impedance
Supports 5 x 4K displays
CPU Performance
GPU Performance
Exact benchmarks likely to be on the website later
More CPU
More GPU
128 GB unified memory for GPU.
This'll be good for ML
8 TB of 7.4 GB/s SSD performance
No way to add extra though except through TB4 ports
The specifications seem great.
The problem is that to take these pictures and do this tweet stream, I'm using custom scripts I've written using software that has no Mac counterpart.
and FWIW, this looks like a system crashed during Johny's section
The annual @ieee_isscc#ISSCC22 conference is coming up next month and the presentation list is now live. Here are some of the talks I'm really looking forward to.
Session 2 is all CPUs, hoping to see if Intel says more about PVC and SPR 1/x
2/ This one is a bit out of left field. Intel is going to talk about ultra-low-voltage Bitcoin ASICs. The DS1 in this talk means there's going to be a demo of it (perhaps more than simulation work?)
3/ @tenstorrent is going to talk more about Wormhole, it's 3rd generation big 700mm2+ chip. Uses GDDR6 and 16 x 100 GbE for scale out - you can connect as many chips together in a 2D array to create the AI training chip you need with predicable on-chip/off-chip latency
Looking at the two new Sunway papers up for the Gordon Bell. None of them are Exaflop on FP64, for clarification.
The Quantum paper showcases 1.2EF using FP32, 4.4EF using mixed, on 41.9M cores. No FP64.
The nuclear paper showcases 298PF using FP64 on 40.4M cores.
It's worth noting that the definition of 'core' is being stretched here. Each chip is listed as having 390 cores - that's 6 groups of (8x8 compute elements + 1 management element). The management element has 512-bit SIMD, unclear what the compute elements can do with vectors
I suspect the management element also acts as the front-end for the compute elements in compute element-only mode.
So what we're really counting here is just execution ports that aren't AGUs or L/S
The Intel Advanced Architecture Development Group (AADG) is responsible for creating the next leap in Microprocessor Design and the future of the x86 environment at Intel by helping to solve the ‘Innovator’s Dilemma’ when it comes to a new core design for Intel.
This dilemma can be simplified into finding the answers to questions like:
Now I've had some time for the @Intel announcements yesterday to sink in, here's my brain dump.
A number of analysts were quite reserved when @PGelsinger joined Intel, stating that no matter what he did, we wouldn't see it for years. I said straight away that Pat can steer the intent of the company as soon as he sat in the seat. Today is a clear message
That being said, Intel last week is still the same as Intel today. Saying stuff doesn't mean much unless Intel does pivot, and it will take a few years to enable that pivot, but there will be a strong undercurrent of things to come throughout, with a continuous focus on 2023