HNHacker News
TopNewBestAskShowJobs

wtallis

22,096 karma · joined September 20, 2009

submissionscomments
wtallis··on Apple Silicon and macOS VMs: Faster LLM Inference with llama.cpp
So those generation numbers aren't really anchored to Apple's hardware designs. It's just counting from when Apple introduced the Metal API, and the first several generations were when the GPU cores Apple was using were still nominally PowerVR designs.
wtallis··on Exploiting System Management Mode with a very long interrupt
> failing after arbitrary timeouts introduces hard-to-debug failures under load

What needs to fail here is the instruction doing insanely slow MMIO. That's not going to be too hard to debug; none of the examples of suitably slow instructions are anywhere close to reasonable, and a fault on a vmovdqu in MMIO address space is a big red flag.

And this attack requires enough ridiculous behavior from coordinating software beyond just the single super-slow instruction that it's hard to imagine any reasonable workload being affected if this case starts causing a fault.

wtallis··on Illinois just passed a law that puts Linux on the hook for age verification
This kind of age signalling with no verification whatsoever does not help the adtech companies at all.
wtallis··on Illinois just passed a law that puts Linux on the hook for age verification
If anyone is distributing it within Illinois jurisdiction (or maybe even distributing it elsewhere while maintaining a business presence in Illinois), they could be at risk of qualifying as an operating system provider under Illinois law.
wtallis··on Illinois just passed a law that puts Linux on the hook for age verification
I think it's pretty obvious that your explanation cannot be complete or accurate. Algorithmic feeds usually have a strong influence of what they want to show you, commonly pushing stuff you aren't interested in, or anything that will keep you on their app/site longer so they have more opportunities to show you ads. There are strong conflicts of interest here.
wtallis··on Exploiting System Management Mode with a very long interrupt
I wonder whether a Thunderbolt accessory can cause arbitrarily long IO operations. What's the upper bound before some other layer gives up and causes the instruction to fault?
wtallis··on 2027 memory capacity is reportedly sold out
None of what you said actually explains how HBM could require 3x the wafers.

The base interface die has eight or 12 memory dies stacked on top of it. The wafer cost of that interface die is therefore small.

If the process of die thinning and TSV stacking reduced yields by a factor of three, nobody would consider HBM mature enough to put into production, and especially not mature enough to be increasing stack height from one generation to the next.

Trace length has approximately nothing to do with die size. If anything, designing for shorter traces means you can get away with smaller PHYs at either end.

What might go some ways toward explaining such a huge difference in die size is that the TSVs themselves take up significant die area and must be fairly numerous to carry both a large number of signal wires and all the power and ground required by the stack. But it's wildly implausible that the die area consumed by the TSVs would be significantly larger than the die area consumed by the memory arrays themselves, or that anyone would build a memory die where the memory array was not a large majority of the total die area.

If there's any truth to that ~3x higher wafer requirement for the same number of bits as compared to DDR5, it must be a combination of several factors and probably includes something non-obvious and dubious, like counting all the area of the passive interposers that go between HBM stacks and GPUs (those interposers aren't competing for the same fab space).

wtallis··on AMD acquires Taalas to boost inference performance by etching models in silicon
> (each HBM3 die is >1000mm² of silicon)

Did you mean that each HBM3 stack is that large? Because it only takes one glance to see that the memory chips are much smaller than reticle-sized GPUs they sit next to.

wtallis··on Nvidia’s Vera Whitepaper Has a Thread Loose
> who knows?

You could try reading the footnotes, which include a link to https://www.amd.com/content/dam/amd/en/documents/solutions/a...

wtallis··on Nvidia’s Vera Whitepaper Has a Thread Loose
In a hypothetical world where AMD wasn't allowed to do AMD64, and Intel stayed committed to Itanium: Itanium would still have sucked, and both PowerPC and SPARC would have out-sold Itanium by an even wider margin than they did in this reality. Itanium could only have succeeded if AMD64 wasn't possible and literally all of the competing 64-bit architectures were killed off by their owners so they could jump on the Itanium bandwagon. Itanium managed to kill off PA-RISC and Alpha and (high-end) MIPS roadmaps, but it still had competitors that were not just viable but actually more successful.
wtallis··on Drivers Are Taking Extreme Measures to Avoid New Car Technology
I've seen enough complaints about features like automatic emergency braking to be quite scared—not of the tech features, but of the drivers who are so overconfident that they are unwilling to accept any false positives from a safety feature and would rather endanger others by eschewing those features entirely.
wtallis··on Show HN: Open-source engine running Gemma 4 26B in 2 GB RAM on any M-series Mac
Reads don't wear out flash memory to any meaningful extent.
wtallis··on Microsoft is using TPM chips to crack down on pirated Windows activations
OEM versions of Windows are just copies with license terms dictating that they are to be installed on and tied to a specific new PC, as opposed to a full retail license that is not tied to a specific PC and may be transferred to a different machine. What you have linked to is quite clearly Windows 11 Pro, not Windows 11 Enterprise and certainly not Windows 11 Enterprise LTSC.
wtallis··on Apple Is the King of AI and Nobody Knows It
AMD's Strix Halo has a 256-bit memory bus. Apple's Max class chips have a 512-bit memory bus, Ultra class chips have a 1024-bit memory bus. Unless you want to go for datacenter hardware, AMD's offerings simply aren't in the same league in terms of memory bandwidth.
wtallis··on AMD and Cerebras Launch AI Inference Solution
LLM inference is a two-phase process. The first phase is prompt processing aka prefill. It's compute-heavy but requires relatively low memory bandwidth. The second phase is token generation aka decode, which doesn't require much in the way of FLOPs but wants as much memory bandwidth as possible.

This announcement is for a system to do the first phase on Helios and the second phase on WSE.

wtallis··on Geekbench 7
> A very frequent real-world workload is the compilation of a software project, which scales almost perfectly with the number of cores.

Geekbench has that as a subtest.

> For the other multi-threaded tasks, like running certain CAD/EDA applications, I use well-optimized programs, either by others or by myself.

Those seem pretty clearly out of scope. You want a workstation benchmark, not a consumer benchmark.

> This would be fine, if it would be more clearly specified when the Geekbench benchmark is distributed.

The descriptions of what Geekbench is are adequate. You're asking for Geekbench to exhaustively list what it isn't to stop you from imagining it to be a one-size-fits-all benchmark, which couldn't possibly exist.

> But others appear to not be aware of this and I have seen frequently erroneous claims on the Internet that some laptop or even some smartphone would match or exceed the performance of a desktop or of a bigger laptop, even one with a CPU with more cores, based on the published Geekbench multi-threaded results, while in reality that desktop or big laptop has a much greater throughput.

Again you seem to be unable to stop asserting that workstation or server use cases are the only ones that are real.

wtallis··on Geekbench 7
> While it would be nice to have a multi-threaded benchmark that demonstrates a realistic scaling towards many threads, I have no reason whatsoever to trust that the Geekbench implementers are capable to optimize the inter-thread communication well enough for reaching the limits imposed by the hardware.

You're contradicting yourself. Do you want realistic scaling, or the kind of highly-tuned hand-optimized scaling that almost no real-world applications (especially in the consumer space) ever get?

In both of your comments here, you make very narrow assumptions about what constitutes a "real" workload, and you're excluding ordinary everyday consumer PC software from what you consider "real". Highly optimized software running on large servers aren't the only "real" workloads, and they're not the kind of "real" workload that Geekbench is trying to represent. For the kind of hardware and use cases that Geekbench actually targets, using the platform/compiler default synchronization primitives and doing work in short bursts interspersed with significant idle time is reasonable and realistic.

wtallis··on Geekbench 7
I don't think you've put forth a coherent argument here.

SPEC CPU is distributed in source form, to be compiled by the end user. There are restrictions on what kind of compiler options and other games you can play when making an official submission to SPEC, which Intel (and others) get around by publishing "estimated" SPEC scores and using those for marketing claims. In practice, the amount of bullshit in Intel's claimed SPEC scores is constrained only by what Intel's legal department is willing to risk getting sued over.

Geekbench is distributed in binary-only form with all binaries for all platforms originating from the author of Geekbench. All the major chip designers pay to see the source code and participate in the development, but the final decisions about what code goes in and how it's compiled are made by the author of Geekbench, not by the chip vendors. They just get to lobby, and AMD has every opportunity to call out Intel and Geekbench should the compiler options for the x86 builds be unfair.

Intel has a long history of cheating on SPEC, but cheating on Geekbench is more difficult because of the necessity of runtime binary patching and they didn't publicly try until this year, with limited success and lots of negative press: https://www.geekbench.com/blog/2026/03/geekbench-6-and-intel...

wtallis··on Geekbench 7
Geekbench is not just intended to be used to benchmark consumer systems, it is intended to be used by consumers to benchmark their systems and to have its scores interpreted by consumers. For that audience, the most important thing that a multi-core score needs to do is counteract the naive assumption that a 20-core CPU will be twice as fast as a 10-core CPU. Having one of the top-line scores be a perfectly-scaling multi-core score that you know will be widely misinterpreted by your audience and used by vendors to mislead consumers would be serious benchmarking malpractice.
wtallis··on Geekbench 7
Yes, they have a valid reason for the change. But I think it's still making a tradeoff, and has a clearly identifiable downside because the content of the multi-core test suite is no longer as closely aligned with the content of the single-core test suite.

(One could just as easily argue that the single-core test suite ought to exclude any task typically done with multiple threads in real-world usage.)

wtallis··on Geekbench 7
> But, of all of the 2-pick-only options, choosing a custom mix and calling it multi-core requires guessing your actual workload on both the content of the mixed test suite as well as the scaling profile of the CPU to reason with.

Yes, that's pretty much the whole point of Geekbench: to offer non-experts a way to measure CPU performance and produce scores that are relevant to a specific class of devices and users and use cases, rather than artificially inflated.

Benchmarking experts can already use something like SPEC, which does offer single-thread and both types of multi-thread test modes (SPECrate and SPECspeed).

wtallis··on Geekbench 7
Geekbench is not intended for benchmarking large workstations. Look at how the web site actually describes the benchmark. It's for phones and laptops and similar devices. A test mode that uses a machine in a way that is only relevant to high-end workstation or server use would be actively harmful to the suitability of Geekbench for its intended audience of consumer hardware and consumer users.
wtallis··on Geekbench 7
Please stop framing the multi core scaling issue as a matter of the benchmark being good or bad. Geekbench 5 scores scale better than Geekbench 6 scores because Geekbench 5's multicore test runs N independent copies of the same workload while Geekbench 6 runs one workload that has to be split across the available cores, with non-zero coordination between threads.

The Geekbench 5 approach of pretending Amdahl's Law doesn't exist is sometimes a valid benchmarking strategy, but generally is the wrong choice for benchmarking consumer workloads and devices, and that's what Geekbench is ostensibly targeting.

The fact that Geekbench 6 scores don't increase linearly with the addition of more CPU cores is not a weakness of the benchmark, it's the benchmark demonstrating an important real-world effect.

The change that Geekbench 7 makes to exclude some subtests from the multicore suite entirely will definitely have the effect of making the overall multicore score scale better with the addition of more cores, but most of the audience for those scores is going to miss out on the fact that the multicore test now measures a narrower range of tasks than the single-core test suite.

wtallis··on ECC and DDR5
The client/consumer desktop market hasn't actually disappeared, and soldered memory is almost unheard-of in that market segment.
wtallis··on Firefox 153 available with support for Vulkan video decoding, JPEG-XL
The iGPU's video decoder could even be drastically less efficient than the dGPU's decoder, and it would still be better to use the iGPU, because just powering up the PCIe link and VRAM for the dGPU costs far more power than the video decoding itself.
wtallis··on Mac gaming is finally getting the overpowered upgrade it deserves
You might want to not only re-evaluate what you consider a real GPU, but also take a look at what the latest Intel integrated graphics is capable of (Panther Lake laptop chips with 12-core GPU chiplet fabbed by TSMC). Intel is still way behind NVIDIA and AMD in the discrete graphics card market, and their drivers don't have the decades of accumulated hacks to work around badly-written games, but their good iGPUs aren't a nightmare of incompatibility and missing features like they were 15 years ago.
wtallis··on I'm a USB-C Maximalist
I was including USB history before USB-C, because HDMI and DisplayPort have been around much longer than USB-C. And limiting the discussion to "fully-featured USB-C cables" is ignoring most of the real problems.
wtallis··on I'm a USB-C Maximalist
The critical difference is that USB versions keep adding new wires and active components, while the display cables largely stick to tightening the analog signal integrity requirements. Whether an old HDMI cable works for a new display is largely a question of how high the data rate is for the new resolution and refresh rate, vs how much the manufacturer of the cable overshot their analog bandwidth target.
wtallis··on Apple's rumored M7 Ultra targets 1.5TB and Blackwell-class AI performance
> I guess I just don't really buy your argument, because if their CPUs had turned out poorly, you could have applied the same argument to their CPUs instead of the GPUs.

Doesn't the argument work fine on the CPU side? Apple doesn't seem to be hurting for lack of a Threadripper competitor.

wtallis··on UI Degrades over Time
You say "that's exactly why they do it" then go on to describe something entirely different: an organization with plenty of options to spend their development resources on, and they make their decision not out of a lack of options but due to misaligned incentives. Which is exactly the point I already made.
← PreviousPage 4 of 34Next →