LinkBack URL
About LinkBacksRead more.According to the source this processor runs at 3.6/4.1GHz and gets a Physics Score of 12,8XX.
Read more.According to the source this processor runs at 3.6/4.1GHz and gets a Physics Score of 12,8XX.
Hmm, I was getting quite happy with AMD not creating a product stack by turning off features on fully functional chips ala Intel. I hope this isn't a sign. There are few things I hate more than buying something that has had extra effort put in to cripple it just to create a product range. Disabling dodgy cores / iGPUs to make use of otherwise scrap parts is different and I don't mind that as long as customers know what they're getting into.
IIRC SMT requires extra hardware so technically they could be using dies that didn't make the cut, although the extra circuitry is pretty minimal so i imagine defects would be pretty low.
It does require extra hardware for the thread management and storage to shift between the two as far as i understand.Originally Posted by Corky34
This cpu seems a bit weird in all honesty.
I thought the design AMD were using were a 'chiplet' design eaning they didn't have as much waste or damaged parts.
It doesn't seem like there's much point to the 3500 when the rest isn't exactly overly expensive to buy (in relative terms)
I don't know that much about AMD's current fab, not having visited one, but that was what immediately went through my mind too. It's certainly plausible that they have some way, designed in, to modify chips that fail at full-spec to be used as lower spec, or even those thar pass at high spec if the demand for lower spec exists.
A cousin, if you like, to the old cutting of bridges to determine chip speed. I certainly suspect it'd be too expensive to design a whole second design, just for this, so I strongly suspect it's something of the above type.
And, phileidiot, while I kinda agree about chip-crippling, the other way to look at it is that by doing so, they can offer a range of chips, at a range of prices, to suit a range of user budgets and needs. There's inevitably going to be an element of design for the high end, then .... not sure cripple is the word I'd use, but rather "limit" .... for those with lesser needs and tighter budgets
I think it makes perfect sense. After all, you have to do the designing for high end to meet that market, and this way you can meet lower markets with virtually no marginal cost implications either at design or fab stages.
Pleiades (23-08-2019)
Crippling implies intention whereas this is just recycling defective dies.
The extra hardware for threading is heavily embedded into the core. If there is a fault there, then the core won't work *at all* and will have to be disabled.
There are two reasons for disabling threading:
1/ Less work per clock helps on peak thermals, which might help either boost clock speeds on marginal parts or lower the rated TDP.
2/ They felt it could justify a cheaper part.
I'm going for option 2 for this one, and evidence that AMD have plenty of silicon kicking around now. Like the 2200G/3200G parts. If this comes in cheap enough it could be an awesome budget gaming part.
Speculative execution mitigation?
I was about to post something along the same lines myself. Much of the hardware used for implementing SMT is integral to the core and hence SMT cannot be broken on these parts in isolation. There are IIRC a small amount of parts which are statically partitioned and in theory could be defective but they're pretty microscopic and the chances of a defect lying within those parts would be vanishingly small, and I'd guess almost certainly not worth binning separately for - they'd be better just treating the whole core as bad and using as a lower core-count part. In reality, as far as I know, SMT is just a market segmentation tickbox, not a yield concern. Though binning based on power as quoted is a possibility.
Speculative execution is a crucial part of modern, high-performance CPUs. The SMT exploits of which I assume you're referring to are core specific, so I doubt that's relevant here.
Salvaging partial dies as you describe is entirely normal and used extensively throughout the semiconductor industry to vastly increase overall yields, especially in the sort of large, high-performance dies we see with PC hardware. Just for some current examples, all of AMD's Zen2 chiplets are fabbed with 8 cores, IIRC Intel's 8C16T 9900k, 8C8T 9700k and 6C6T 9600k share a die, Nvidia's 2070 Super, 2080 and 2080 Super share a die, and AMD's 5700 and 5700XT share a die. Memory devices e.g. NAND/DRAM are built with generally lots of redundancy.
I would suspect a cheaper lower TDP part . @35w . After all who *needs* anything more than 6 cores . or maybe they are getting ready to incorporate a gfx slice and get the 3500G ?
I can't be bothered quoting all the people I'm replying to. So you can ALL read this. Because I am so important and what I think is just so valuable you can't afford to miss it. Allow me to educate you all.
Now I'm safe in the knowledge that no one is reading this anymore, I can say what I like. Arse parsnips.
I understand the business reasons for creating a range but doing so by purposefully making something and then breaking part of it is just antithetical to how my mind works. I'm a Yorkshireman. I believe in extracting every bit of value from everything I can. Taking something and making it better is good. Taking something and making it worse is bad. I an a simple minded idiot.
Whilst I don't know about AMD, I know Intel chips can simply not function if the bits required for hyperthreading are broken. I suspect the same is true for AMD. Therefore this wouldn't be recycling otherwise defective dies (which I can definitely get behind) but intentionally removing a fully functional feature to justify a price point.
The above posts about TDP is also what came to my mind and it really does make a lot of sense. It would explain the slightly reduced clock speeds as well - they're so slightly reduced they might as well not have bothered but it would certainly make sense if they're trying to keep within a lower TDP. Also, the very blunted benchmark they gave may be as a result of having to limit the boost to keep within the lower TDP. I'd frankly expect the loss of hyperthreading to make less difference to the score than is shown here.
To clarify, I'm more than happy for defective dies to be used and if they have a way to use chips which are marginal by disabling hyperthreading then I'm all for it. If they're just doing it to artificially create a range then, logically I get it. I just really hate the practice. I want to see an engineer working to squeeze the most out of their budget and be creative to provide the customer with something great. The idea of being asked to build something by strategically breaking something else is just awful. You want to motivate your engineers to build something great within limitations. If I was asked to make something worse for the sake of marketing, I'd be a demoralised drone within weeks.
I think part of that comes from seeing pacemakers which have the hardware to provide certain features which can really help a patient but they're not enabled by the software so as to create a product range. The end result being if someone gets a pacemaker put in by a hospital / country that buys cheaper devices, the patient misses out.... and it's pointless as the hardware is all there, it's intentionally crippled. I suspect that's one reason I dislike the practice as it has resulted in some very difficult situations for me personally and my patients which is going to bias my view.
That's pretty bad so I can understand your viewpoint.
Having been that engineer many times over I have to say it generally seems better vs the alternatives. If you engineer a specific product for a lower price point then you end up with more designs to debug, and even if there are common components and software items you can end up complicating those which can break the original high end product. It can work out better to just disable bits of one product in terms of time to market and reliability. What the marketing guys do with the price points is often hidden to me, but I'm actually happy with the engineering as not having to work on a genuine low end product frees up time to add new features and performance that benefit the entire range.
As part of that, I have never skimped on safety features. That might be tricky on something like a pacemaker though where I imagine every aspect could be considered a safety feature.
I'd have been surprised if it wasn't still normal. It seems like an obvioys thing to do. Examples much appreciated. Thanks, watercooled.
That does make sense. They did try a pacemaker business model whereby a pacemaker's firmware could be updated with various add ons which seems like a fantastic idea. The company only has to get one set of hardware approved (which I think is the main reason for doing what they do) and everyone gets a basic pacemaker at lower cost to implant and then as we identify various needs in clinic, we update the firmware with the required software and the bill is footed every month.
The problem being that the bean counters don't trust us and so when they trialled this model they found that hospitals rather liked the idea of the "one off cost" of a pacemaker with very little ongoing cost. They didn't want to hand us control of spending more money and so it was likely made so hard to get an upgrade or add-on authorised that no one got them. And the end result is that the company makes less money (even less with a payment infrastructure to maintain) and patients don't get the features.
I can see why they won't trust us. People like me will throw a load of stuff at a patient in order to extend the life of the device and reduce the chance of a patient needing surgery to replace it. And that would cost extra money in this scenario and so the end result would be me being forced to do an analysis of how much extra life I get out of a device, the cost of a new pacemaker and how many patients actually live long enough post implant (they're often in their 70s or 80s at implant) in order for it to even be an issue.... an analysis I'm not equipped to do and don't have the time to do and so it'll not get done. The bean counters win.
There is absolutely a place for really basic, cheap pacemakers with a low current drain that don't do anything fancy and last a really long time or are really small (they're normally designed from the ground up to be like that rather than top of the line units cut down). But the problem is that the bean counters don't see what we do (unless they watch Casualty) and we don't see what they do (much like your marketing guys blinding you to their plan). As such, their KPIs are often vastly at odds with mine.... mine being "patient not dead, ideally not crippled".
Ford are a pain for safety features being missed for money. They removed the rear seat belt tensioners from their top line Mustang in certain markets becuase apparently, they didn't think it'd get tested in those markets. Which turned it into something of a death trap for rear passengers in a crash as the rest of it was designed to work with seatbelt tensioners. Their cars also have a propensity to catch fire rather more frequently than other brands...
This was kind of expected, since they probably have ton of dies that can't be used for 3600, so 3500 would make sense and it likely will game fairly similar to 2600, they likely are just waiting for old supplies to run out. Plus whole 4 cores 8 thread vs 6 cores 6 threads is pretty muct on "it doesn't matter" basis anyway and this will be better than 2400G or 3400G, since neither of two is Zen2, APU are always one node behind. But it really will work nice as something for RX570/580 build. Not to mention that price will likely drop a bit quite quickly.
There are currently 1 users browsing this thread. (0 members and 1 guests)
Posting Permissions
Reply With Quote