Now that two days have passed since Bulldozer’s release and I’ve had some time to think about what AMD’s trying to achieve, I think I’m ready to put my opinion in paragraph.
Bulldozer is a disappointment for various reasons, of which only a couple relate to its actual results from reviews. Most of them seem to indicate some kind of management issue over at AMD (not to say Intel doesn’t have any, but they seem to do a better job of hiding it). To list:
Performance not meeting resourcesThe undeniable conclusion is that while Bulldozer may be great for server workloads (I await Interlagos final-silicon benchmarks before concluding this), it clearly is not as great for most enthusiast applications. If we look at multithreaded application performance (because looking at lightly threading makes me feel sorry for Bulldozer), going from Thuban to Zambezi, we see an approximately 15% increase in load power consumption and double the transistor count, for performance increases that vary
from 36.6% for par2 recovery to 14% for 7zip MIPS. Accompanying this is a 9% reduction in die size. Contrast this with the 27% decrease in die size and 29% increase in transistor count, 8% increase in load consumption, and
corresponding change in performance over the same benchmarks* that we observe going from Lynnfield to Sandy Bridge. Both are observed over a single node shrink from 45nm to 32nm. In other words, while Bulldozer shows varying degrees of improvement over Thuban, the improvements are nowhere near matching what was accomplished with Sandy Bridge, a microarchitecture that is clearly a better fit for enthusiast application performance. To put it bluntly,
it is not a good fit for the enthusiast market.
* Yes, this is but a subset of the benchmarks available. I pick them not to bias Sandy Bridge, but to avoid distraction from synthetic benchmarks, focusing instead on more pertinent real-world multithreaded benchmarks. One might even note that I picked benchmarks where Bulldozer performance was not downright painful to look at. 7zip is the only anomaly to the suite of real-world performance tests, since no real-world compression test was done for FX-8150.Some may argue that, regardless of the number of transistors used, the die size, the TDP of the chip and its power consumption, as long as it performs better with a corresponding scaling in price, it should be considered a better product. Unfortunately, this argument also falls flat. On Newegg, the 1100T is $189; the FX-8150, $279 (though $249 MSRP). Ignoring the pricier 9-series chipset you’ll need for AM3+, that’s a 30% increase in price, for performance increases that don’t scale as much. The Sandy Bridge flagship, i7-2600K, at $319, was 7% pricier than its i7-860 predecessor ($299) at launch, but the performance increase justifies the price increase.
Another popular argument is that the FX-8150 compares favourably to an i7-2600K in x264 and similar heavily threaded applications for a lower price and is thus obviously the superior choice. This is putting the cart before the horse. The FX-8150 is not a better product because it is cheaper; it is cheaper because it is
not a better product. You don’t think AMD would have loved to price the FX-8150 at $299? Then you quite sorely misunderstand AMD’s position. They did not pick the value-for-money position willingly, but were forced into it. A larger chip with more transistors would be costlier to manufacture, and pricing it lower than a competitor that has fewer transistors and a smaller die size is simply not desirable. Low-margin pricing is a risky game that chip designers try to avoid with good reason.
Market misfitSince the first Bulldozer preview we’ve seen and understood that AMD is aiming for a different architecture, one that is geared for greater overall throughput.
RWT surmised it is geared for heavily multithreaded server workloads, and none of the hardware previews revealed enough numbers for a conclusive guesstimate of single-thread performance. The unspoken thought was that AMD would be able to make a comeback,
somehow, and without knowing enough about Bulldozer that was all we could hope for.
I’m not about to accuse AMD of withholding information. Intel has been much more forthcoming with microarchitecture information since Conroe; that is a privilege they enjoy from being in the lead and having picked a successful path in Nehalem. AMD is in no such position. Rather, what I’m questioning is this: AMD execs could not have been unaware of Bulldozer’s real-world performance in enthusiast benchmarks. Surely they must have known how it would perform against the X6 and Sandy Bridge at least a year before release.
Naturally, they could not back out of the enthusiast market at that point; I’m not sure if the enthusiast market is really that profitable, but they seem to need revenue from it quite badly. Instead of playing down Bulldozer as an enthusiast chip, they drummed up marketing (“World’s first 8-core desktop processor!” etc), and convinced OEMs to come up with
44 31 AM3+ SKUs for Bulldozer alone. Rebranding the 8-series chipsets to include AM3+ support is inevitable, but wherefore the need to overprice them? Think about it:
44 31 SKUs over 3 chipsets, on micro-ATX and ATX alone! Even FM1 only got
42 33 SKUs over 2 chipsets on mini-ITX, micro-ATX and ATX. I know the inevitable comparison here is to Intel’s
>200 >160 SKUs for H/P/Z 6-series, but let’s not get too distracted: The point I’m trying to make is that AMD sold [the idea of] Bulldozer convincingly enough that they managed to drum up almost as many AM3+ SKUs as FM1 SKUs. Did they intend to sell Bulldozer at the same volume as their mainstream Llano offerings, despite knowing its performance in enthusiast workloads?
* Numbers adjusted to discount open-box productsYes, certainly, Intel might have done the same thing in those shoes. But we’d have hated them for it. Even cutting AMD some slack, only a fanboy could believe such marketing again. They’ve disappointed with Phenom I (“World’s first true quad-core!” ... sound familiar? Remember how it did against C2Q?), and now they’ve disappointed with Bulldozer. It’s fine if Bulldozer, being geared for server workloads, wouldn’t do well in the enthusiast space, but they should have marketed it likewise then. Drumming up so much support for it knowing it would let enthusiasts down is a bad, bad PR move, and no doubt the bad taste is going to stay.
Mismanagement of the FX brandI don’t mean to say the FX brand is some sort of godly name that has been sullied, but it would not be an overstatement to say that the FX brand name has had a really good reputation owing to Athlon 64’s success in its heyday. It has good PR vibes associated with it, and is a valuable brand name that could be used to sell new processor lines.
No company would willingly throw such a brand name away on a product that isn’t up to expectations. Dell reserves its Ultrasharp line of monitors only for non-TNs, HP Elitebooks are enterprise-level notebooks with build quality that matches the name, and I don’t think I need to say anything about Lenovo Thinkpads. FX processors enjoy the same reputation, and one would definitely expect it to at least best the X6 conclusively, even if not quite matching Sandy Bridge across the board.
They could certainly have used a different model name to market Bulldozer, but again they seemed to be desperate to drum up support for a product that was ill-fitted to enthusiast needs (overclocking, encoding, rendering, etc). The result is that they have wasted a lot of consumer goodwill, and the FX brand is no longer as reputed or profitable as it was before Bulldozer’s release.
Mistrust of AMD managementI’ll say it: trust in AMD has been eroding. Without knowing what’s going on in Sunnyvale I won’t make any guesses about what AMD management is thinking, but looking at AMD-related news lately I don’t think it’s unreasonable to say that things look less than sunny. A number of high-level execs have left, execs who’ve brought about the success of earlier products. But this is just one aspect of it.
The other thing is that AMD simply needs a better PR/marketing team. Intel, the success of their products notwithstanding, has a more downplayed PR team that doesn’t overpromise and doesn’t set themselves up for blame as easily. AMD’s PR team, on the other hand, puts forth the argument that their architecture is “forward-looking” and “aimed at tomorrow’s workloads” (Intel said that about the Pentium 4 too). It’s fine to say that in the server space, where workloads are more readily characterisable (virtualisation, HPC, database, hosting, etc) and hardware more readily tunable for specific needs, but it won’t fly in the enthusiast market; there, it simply looks like blame-pushing (
if only devs would compile their apps for optimum performance on
our obviously superior platform, we would get better benchmark numbers!)*. If one is going to sell a “new”, “special” microarchitecture in the consumer/enthusiast space, one had better do one’s best to ensure software support is present, if not at least promised in the near future, before launching one’s product. It’s great if it works for tomorrow’s workloads, but if it doesn’t fit today’s workloads well, people are only going to buy it tomorrow. [See market misfit above].
* No doubt x86 is partially to blame for this, but again let’s not get distracted: AMD’s disadvantage in Intel-influenced instruction sets aside, the architecture nonetheless demonstrates its unsuitability to lightly-threaded and enthusiast workloads.I’ll avoid pushing too much blame to them; Bulldozer is not an easy product to sell to enthusiasts, not with that kind of performance. (It was a mistake to push Bulldozer in the enthusiast space in the first place IMO.) I am of the opinion that the amount and level of marketing for a product should match its performance, and if it doesn’t perform then don’t make it out to be the best thing ever. AMD’s selling of Bulldozer leaves me with the impression that the Bulldozer project hasn’t been very well-managed from the start. This is a different accusation from saying that Bulldozer is a poor product; don’t mix the two up.
So what now?AMD has promised 50% better performance-per-watt for the Bulldozer architecture by 2014. That is a pretty vague promise, as no mention is made of which workload this promise applies to, nor is it a firm promise of better performance since they could just lower power consumption by 33% keeping performance equivalent. All we know for now is that “AMD is committed to improving Bulldozer”. That’s hardly newsworthy.
On x86 monopolyI have no reason to hate AMD, since I haven’t preordered Bulldozer. But we’ve all been eagerly awaiting updates/upgrades from AMD that will “keep Intel on its toes”. I’m not naive enough to think that we have to keep AMD alive at all costs because their exit from the x86 market is going to bring back $500 mainstream Intel processors. Intel does have competitors, and they’re not in the x86 market. The end of PCs isn’t going to come anytime soon, but Intel can easily bring it about with suicidal pricing policies. Let’s not even mention the antitrust suits that various groups are waiting to spring on Intel should AMD make its exit ... Rather, what I worry is that Intel will have less motivation to improve on things it has to catch up on, namely IGP performance. Product stagnation is not a desirable outcome for consumers either.
On integrated graphicsI have similar expectations of AMD; smaller companies don’t get any more slack in the innovation or product improvement department. The lack of Eyefinity/triple-display support in APUs, lower hardware post-processing quality compared to the higher-end chips, and other little bugs (
broken chroma upsampling in non-BD playback, for one) was a big disappointment for me, even if its raw performance is way ahead of HD Graphics. What this tells me is that AMD’s IGPs play second fiddle to its dedicated GPUs. I don’t mind if raw performance falls behind the PCIe cards; that is only to be expected. But they have made no mention of getting these issues fixed in future iterations, and I have no idea if future developments in GPU
features will make their way into the APU.
On the other hand, Intel only has one focus for graphics: IGP. Whatever they are working on in graphics, you know it’s going to work its way into their IGP. They’ve promised fixes for the 23.976-fps-without-UAC issue, better hardware AA, as well as HDMI 1.4, DX11, OpenCL 1.1, OpenGL 3.1 in Ivy Bridge; that pretty much brings Ivy Bridge IGP’s feature set up to date and completely matching AMD’s and Nvidia’s (if they fulfill their promises). Go ahead and poke fun at their IGP performance; it is entirely justified. But one cannot deny that for non-gamers, HD Graphics is becoming more and more viable as a fully featured GPU. We could not even say this of their IGP two generations ago (in Clarkdale), but they have improved greatly with each generation. Once Ivy Bridge rolls around, dedicated HTPC cards may no longer be needed on Intel-based platforms.
AnxietyIt may sound like I’m selling Intel here, but what I mean to say is that AMD and Nvidia have a lot to be worried about even in the graphics arena, their stronghold: the market does not consist of gamers alone. Right now videophiles/consumers are still buying dedicated graphics cards for HTPCs (some with valid reasons, others based on outdated impressions of IGPs), but if Intel catches up in features, that is one market segment decimated for AMD and Nvidia.
Meanwhile, all we hear from AMD and Nvidia regarding new products is gaming, gaming, gaming. There is some mention of compute, but hardly any fire following the smoke; we haven’t seen much use of it apart from CUDA encoding, Folding@Home and MD5 brute-forcing. CUDA is popular, but largely in the HPC space, not so much in consumer applications. DirectCompute seems to be taking off, but very slowly. They should push for it much more aggressively in applications if they’re going to keep any advantage for a dedicated graphics card in anything other than gaming.
Looking forwardWhile I still can’t say Intel is my favourite company, I can say right now that it is the company that excites me the most. They’re promising things other than raw gaming performance, and delivering them. To date: triple-display on IGP, greater hardware video encode/decode capability, Thunderbolt support (on limited Panther Point motherboards, presumably more widespread in Lynx Point), new low-profile board options (
thin ITX). AMD and Nvidia do have things on the roadmap, but they all seem to be geared for gaming and rendering needs. No doubt they would excite gamers and 3D modellers, but they promise nothing to non-gamers, typical consumers and non-gamer enthusiasts (yes, they do exist; look on AVForums and SPCR forums) other than better raw performance. Case in point: Nvidia doesn’t seem at all interested in OOTB single-card support for displayport or triple-display.
There is one exception, however:
AMD Graphics Core Next (GCN; yeah, I did say they need a better PR/marketing team). This is supposedly what AMD’s future APUs will be, and it promises orders-of-magnitude increase in multithreaded integer performance without any developer intervention (or minimal intervention, at most).
This, not Bulldozer, is what could turn AMD’s fortunes. And this is what I’d love to see more of.
Unfortunately,
it’s not slated to be released in Trinity (which will be using VLIW4), so
the earliest we’ll be seeing this is 2013. What we’ll be seeing on Radeon 7000-series instead is
Fusion System architecture (FSA), basically AMD catching up with compute-oriented architectural developments that Nvidia had put in place in the Fermi architecture, released last year.
Mini-update insert: VR-Zone with a bit more IDF coverage on Haswell overclocking. More fine-grained clocking options sounds yummy, even if I’m not an overclocker (if anything I’m more likely to undervolt instead). To quote the article:
“A year later, in early 2013, the pinnacle of Intel’s 22 nm process show off, the initial Haswell processor, is expected to go another step further, where CPU core, GPU, memory, PCI and DMI ratios are all set independently here, on top of fine grain BCLK base clock available within the Lynx Point chipset.”Sounds like Haswell could be even easier to overclock than Sandy/Ivy Bridge.
