5 ms·

This is not a great article, if you've followed the news and around the M1 in the last few weeks. Sure, the Bloomberg audience are not computer geeks, but still

by em500 6y ago

This is not a great article, if you've followed the news and around the M1 in the last few weeks. Sure, the Bloomberg audience are not computer geeks, but still. "If they live up to expectations, they will significantly outpace the performance of the latest machines running Intel chips, according to people familiar with the matter who asked not to be named because the plans aren’t yet public." It's not entirely obvious that "the latest machines running Intel chips" here probably refers only to the latest Macs running Intel chips. We already know that the M1/Firestorm beats Intel handily in single thread, so this just means they'll use enough cores to beat the 28 core Xeon-W in their current Mac Pro. This is of course completely uninteresting, since 1) they've long announced they'll be Intel-free by 2022 and 2) they're not going to release new Macs any slower than the current Intels. Going by Geekbench they need probably a 16 to 20 Firestorm CPU to beat the current Xeon W-3275M Mac Pro. A 32 Firestorm CPU would completely obliterate it. To make a statement, they'll probably do just that. The more interesting question to me is if they're going to go for the absolute performance crown, currently the AMD EPYC 7H12, a 64 core Zen2 CPU which benches around 60k in Geekbench5 (vs 19k for the 28-core Mac Pro). By mid-2021, that that will be probably a 64 core Zen3 EPYC at around 72k (+20% IPC vs Zen2). Apple can probably match that with a 48 Firestorm CPU. A 32 core CPU won't do it, even with a generous IPC increase next year.

admiralspoo 6y ago
Apple will likely have HBM on the 32 core packages, which should provide a significant IPC uplift over the LPDDR4 used in the M1 Macs. I'm pretty certain a 32-core Apple Silicon machine will handily beat a 64 core Zen3, likely at half the power, especially since the AMD's chiplet designs runs into significant memory bandwidth and communication issues.
- dragontamer 6y ago
  > Apple will likely have HBM on the 32 core packages First of all, that's ludicrously expensive. You need an active interposer that's larger than the chip to do something like that. 4x HBM stacks means 4096 wires to run between your compute-device and the RAM, that's a lot of wires in a tiny area, and only possible with advanced packaging (expensive) methodologies. Second: HBM2 based designs have poor memory-capacity: 8GB per stack means 4x HBM2 caps out at 32GBs under most reasonable configurations. Going into 2021 with a ludicrously expensive 32GB machine seems a bit weak. I don't think anyone in the high-end market is targeting that level of memory capacity. Even if we're generous and HBM capacity doubles between now and then, I'm not sure if 64GBs is enough for workstation tasks. HBM works for supercomputers, because when 32GB isn't enough, you just buy another node. A64FX, the current CPU-based supercomputer that uses HBM, only has 32GB/node. Then Fugaku bought 150,000 nodes.
  - imtringued 6y ago
    I have a "low budget" PC with 32GB RAM and 8 cores. I don't know why anyone would want a workstation with less than 128GB RAM. Of course, the Apple world is extremely weird and you get massively overcharged for commodities like RAM.
    - dragontamer 6y ago
      Well... my point isn't so much "Apple is stupid", as much as "Apple would be stupid to use HBM2... and I don't think Apple is stupid". I think Apple would want to release a 64GB or 128GB Mac Pro... but only to upsell its customers to a 256GB or higher Mac Pro. The ability to configure more RAM into a system is key for marketing (even if this RAM is soldered onto the motherboard: like LPDDR4x, it wouldn't be too expensive to make different motherboards especially in a niche $10,000 computer like the Mac Pro). But if Apple did go with HBM2 itself... well... then its physically unable to go above 32GB (today), or maybe ~64GB (if bigger HBM2 modules come out next year). -------- Sticking with LPDDR4x just seems like the most obvious, and best, decision for Apple. So that's my expectation. But then again: making a 8-way decoder + fat-as-heck execution pipeline wasn't part of my expectation either. So maybe Apple will surprise us again.
  - AnthonyMouse 6y ago
    They could use HBM as L4 cache. Though it still wouldn't be cheap.
    - dragontamer 6y ago
      HBM's latency characteristics are about the same as any other DRAM however (including DDR4 or LPDDR4x). So you have a high-bandwidth, but no better latency, L4 chip. And that's kind of hard to deal with and optimized against. Xeon Phi had a DDR4 + HMC (where HMC was a early competitor to HBM), and getting good performance out of that was unusual. DDR4 ended up having slightly better latency, so the HMC cache would hurt some workloads.
eloff 6y ago
Unless Apple intends to get into the server business, I think the market at the top is very small. They really only have to be competitive with the lower core consumer chips to accomplish their goal. I think they should try for the performance crown though and the server business. There's a lot of money in that. In servers power consumption is an important consideration and they have an advantage there as well. ARM based architectures are also a lot more relaxed in that they don't have to present a nearly sequentially consistent view of cache and memory between multiple cores like x86. That may give them an edge scaling the core count. It's horrible to program for if you are coding lock free algorithms, but almost nobody does that and fewer should.
- em500 6y ago
  Their current 28-core Mac Pro (which even has a rack version) is probably not a volume business either, that doesn't stop them from making them. It's probably much cheaper to go for the performance crown with their own CPU now than with Intel/AMD in the past. See it as Apples Bugatti Veyron/Chiron :)
  - eloff 6y ago
    They didn't have to make a special chip for it though, it can be economical to offer a small volume product made of off the shelf components. It's a losing venture to do it with custom designed chips - that's extremely expensive and I doubt the market is large enough to justify it. Unless they intend to get into servers. I'm very curious to see what they do.
    - _ph_ 6y ago
      They have to make a server-class chip if they want to have something like the Mac Pro on ARM too. To actually go into the server business they might have to add a bit more I/O capability like Epyc vs. Threadripper, but the basic cpu should be more than up to it, if it has 32 performance cores or more. Even just a rack-mounted M1 would be interesting for many server applications and anything above more so.
- 6y ago
formerly_proven 6y ago
Note that the W-3275M is part of Cascade Lake, which is a really small iteration on Skylake-X, the most notable change is probably that it includes hardware mitigations. Otherwise fairly little changed about the cores themselves since 2015.
tromp 6y ago
> they're going to go for the absolute performance crown Or, in Elon Musk speak, a "hardcore smackdown", with x86 playing the role of the gas engine.

by em500 6y ago

admiralspoo 6y ago

Apple will likely have HBM on the 32 core packages, which should provide a significant IPC uplift over the LPDDR4 used in the M1 Macs. I'm pretty certain a 32-core Apple Silicon machine will handily beat a 64 core Zen3, likely at half the power, especially since the AMD's chiplet designs runs into significant memory bandwidth and communication issues.

dragontamer 6y ago
> Apple will likely have HBM on the 32 core packages First of all, that's ludicrously expensive. You need an active interposer that's larger than the chip to do something like that. 4x HBM stacks means 4096 wires to run between your compute-device and the RAM, that's a lot of wires in a tiny area, and only possible with advanced packaging (expensive) methodologies. Second: HBM2 based designs have poor memory-capacity: 8GB per stack means 4x HBM2 caps out at 32GBs under most reasonable configurations. Going into 2021 with a ludicrously expensive 32GB machine seems a bit weak. I don't think anyone in the high-end market is targeting that level of memory capacity. Even if we're generous and HBM capacity doubles between now and then, I'm not sure if 64GBs is enough for workstation tasks. HBM works for supercomputers, because when 32GB isn't enough, you just buy another node. A64FX, the current CPU-based supercomputer that uses HBM, only has 32GB/node. Then Fugaku bought 150,000 nodes.
- imtringued 6y ago
  I have a "low budget" PC with 32GB RAM and 8 cores. I don't know why anyone would want a workstation with less than 128GB RAM. Of course, the Apple world is extremely weird and you get massively overcharged for commodities like RAM.
  - dragontamer 6y ago
    Well... my point isn't so much "Apple is stupid", as much as "Apple would be stupid to use HBM2... and I don't think Apple is stupid". I think Apple would want to release a 64GB or 128GB Mac Pro... but only to upsell its customers to a 256GB or higher Mac Pro. The ability to configure more RAM into a system is key for marketing (even if this RAM is soldered onto the motherboard: like LPDDR4x, it wouldn't be too expensive to make different motherboards especially in a niche $10,000 computer like the Mac Pro). But if Apple did go with HBM2 itself... well... then its physically unable to go above 32GB (today), or maybe ~64GB (if bigger HBM2 modules come out next year). -------- Sticking with LPDDR4x just seems like the most obvious, and best, decision for Apple. So that's my expectation. But then again: making a 8-way decoder + fat-as-heck execution pipeline wasn't part of my expectation either. So maybe Apple will surprise us again.
- AnthonyMouse 6y ago
  They could use HBM as L4 cache. Though it still wouldn't be cheap.
  - dragontamer 6y ago
    HBM's latency characteristics are about the same as any other DRAM however (including DDR4 or LPDDR4x). So you have a high-bandwidth, but no better latency, L4 chip. And that's kind of hard to deal with and optimized against. Xeon Phi had a DDR4 + HMC (where HMC was a early competitor to HBM), and getting good performance out of that was unusual. DDR4 ended up having slightly better latency, so the HMC cache would hurt some workloads.

eloff 6y ago

Unless Apple intends to get into the server business, I think the market at the top is very small. They really only have to be competitive with the lower core consumer chips to accomplish their goal. I think they should try for the performance crown though and the server business. There's a lot of money in that. In servers power consumption is an important consideration and they have an advantage there as well. ARM based architectures are also a lot more relaxed in that they don't have to present a nearly sequentially consistent view of cache and memory between multiple cores like x86. That may give them an edge scaling the core count. It's horrible to program for if you are coding lock free algorithms, but almost nobody does that and fewer should.

em500 6y ago
Their current 28-core Mac Pro (which even has a rack version) is probably not a volume business either, that doesn't stop them from making them. It's probably much cheaper to go for the performance crown with their own CPU now than with Intel/AMD in the past. See it as Apples Bugatti Veyron/Chiron :)
- eloff 6y ago
  They didn't have to make a special chip for it though, it can be economical to offer a small volume product made of off the shelf components. It's a losing venture to do it with custom designed chips - that's extremely expensive and I doubt the market is large enough to justify it. Unless they intend to get into servers. I'm very curious to see what they do.
  - _ph_ 6y ago
    They have to make a server-class chip if they want to have something like the Mac Pro on ARM too. To actually go into the server business they might have to add a bit more I/O capability like Epyc vs. Threadripper, but the basic cpu should be more than up to it, if it has 32 performance cores or more. Even just a rack-mounted M1 would be interesting for many server applications and anything above more so.
6y ago

formerly_proven 6y ago

Note that the W-3275M is part of Cascade Lake, which is a really small iteration on Skylake-X, the most notable change is probably that it includes hardware mitigations. Otherwise fairly little changed about the cores themselves since 2015.

tromp 6y ago

> they're going to go for the absolute performance crown Or, in Elon Musk speak, a "hardcore smackdown", with x86 playing the role of the gas engine.