bit-tech.net

A Phenomenon in Warsaw

We, like many other hardware enthusiasts, have been waiting for a long long time for AMD to respond to Intel’s Core microarchitecture launch and there have been many die-hard AMD fans that have jumped ship after getting tired of waiting. Well, if you’ve been following the news column in the last couple of days, you've likely noticed that the wait is now over - AMD launched its first quad-core Phenom processors on Monday morning.

I spent the back end of last week with AMD and the new Phenom processors in Warsaw, Poland and, as a result of my time there, I came away with some ideas of what’s going on inside the company at the moment.

There’s no other way to phrase it than to say that the Phenom launch is a bit of a disappointment in my eyes, and with it I’ve lost quite a lot of faith in AMD. The processors have only launched at 2.2GHz and 2.3GHz, and the underlying execution core is almost exactly the same as the K8 execution core that they were designed to replace. The only saving grace is that Phenom’s pipeline is 33 percent wider and can handle more complex instructions, including some of the SSE4 instruction set (which is referred to as SSE4a by AMD) – these features, combined with the new shared L3 cache are where the chip gets its improved performance from.

I would have hoped for more than this from Phenom, given how long we’ve waited for the damn thing.

"The Phenom launch is a bit of a disappointment in my eyes, and with it I've lost quite a lot of faith in AMD."

Anyway, during the time that I was with AMD, I was able to run some tests on one of the new CPUs – albeit one that isn’t being released until the start of next year – to gauge how well the new architecture performs against Intel’s current processors. I can’t tell you exactly how well it performs, because the benchmarks we were allowed to run were a limited selection and the system was pre-configured by AMD.

Yes, that’s right – the testing environment was one controlled by AMD. AMD allowed us to tweak the system and mess with the BIOS, so I don’t think there was any malice in the way the company handled this limited benchmarking, but it could have been handled much better in my opinion.

The systems were curiously set up with 32-bit versions of Windows Vista, which meant we’d got nothing to really directly compare the Phenom to, as we at bit-tech moved our CPU test rigs to Windows Vista x64 last month. We’d hoped to move before that, but we’ve been working in the background to get a wide selection of stable application benchmarks working – this doesn’t just happen overnight.

I was surprised by AMD’s move to test under 32-bit given that Phenom is a 64-bit processor, and I didn’t really get a straight answer as to why this was the case. If you couple this with the fact that the limited selection of benchmarks was a mix of synthetic benchmarks which we don’t touch and application benchmarks that we do use, but with very different workloads, it’s quite hard to gauge how well Phenom really shaped up against the competition.

"I don’t think most consumers really care whether their quad-core processor is 'native' or not as long as it performs like a quad-core processor."

This was one of the main reasons why you haven't seen a full disclosure of Phenom’s performance on bit-tech thus far. The other reason is that we’d rather benchmark the chip on our own terms, where we’ve got full control of our own test environment and benches used. We felt that this would will give you, the loyal bit-tech readers, the information you’re looking for and not the information that any company wants you to see.

Despite this, I could at least get a general idea of how a 2.4GHz Phenom 9700 performed, thanks to a couple of the industry-standard benchmarks...and I have to say, I was underwhelmed. The long of the short was that it wasn’t much faster than a Core 2 Quad Q6600, nevermind the Core 2 Quad Q6700. The Q6700 has come down in price quite considerably since it was first launched last year as an Extreme Edition, making it an even worse overall picture.

The one conclusion I could draw from my time with Phenom was that the fastest chip that’s supposed to be available today, the X4 9600, isn't anywhere near as fast as Intel’s equivalently priced Core 2 Quad Q6600. However, the company was adamant that it had the better product based on the fact that Phenom is a more elegant solution since it’s a native quad-core chip.

AMD banged on, and on, and on about native quad-core during the briefings I had with representatives from the company. However, to be perfectly frank, I don’t think most consumers really care whether their quad-core processor is 'native' or not as long as it performs like a quad-core processor. Right now, at the very least, it looks like Intel’s decision to adopt a non-native quad-core processor was the right one to make, as there have been Intel quad-core processors on the market for over twelve months now.

By my reckoning, if you take performance and availability into account, AMD is around that same twelve months behind Intel and, given the fact that we’ve already received samples of the upcoming Core 2 Extreme QX9770 processor, Intel isn’t about to stop. With the next-generation Nehalem architecture scheduled to arrive next year, things don’t look particularly rosy for AMD at the moment. If you thought that was all though, you're in for a treat - there's still more.

At the eleventh hour of the Phenom production, AMD found an issue in the L3 cache translation lookaside buffer at very heavy loads which prevented the company from launching it at 2.4GHz. When AMD announced this to the members of press present at the event, there was that horrible silence where opinion in the room just sank.

The company did state that there is a BIOS fix available which users would be able to turn on or off in AMD Overdrive – the company’s new tweaking utility, which actually looks pretty swanky – as the situations when end users are likely to encounter this problem are very small in number. The penalty for enabling the L3 cache TLB fix all the time is around a 10 percent performance decrease, which I guess is why AMD is leaving users with the option to enable or disable the fix as they see fit. The company’s representatives said that the issue would be fixed soon with a new stepping, which will probably be what the higher-clocked chips will use when they’re launched early next year.

Back to top

Tim Smalley

However, I think there’s more to it than this one problem in and of itself. I got this inkling after asking one question during the event about how mature the 65nm process was. See, we’ve still not seen Athlon 64 X2 chips based on the new process that are shipping in volume at speeds higher than 2.6GHz. The answer we got was that the process was taking longer than expected to mature – part of the reason for this was that the company had focused so hard on getting Phenom out of the door in a respectable amount of time.

Obviously, AMD’s approach to the event was that the company was there to talk about its platform as a whole, as it is the only company that can flaunt a true enthusiast platform. Never before has a manufacturer provided every major microprocessor in the enthusiast’s system infrastructure – the CPU, chipset and discrete GPU. Many would argue that Intel has had a platform for a long time, but the truth is that Intel’s graphics processors are far and away from discrete, and they’re certainly not targeted at enthusiasts.

As a result of this, it wasn’t much of a surprise when AMD spent a long time talking about its 'Spider' platform. One key point that I picked up on was the point referring to cost of ownership – AMD was making a lot of noise about the fact that the new Phenom processors will work in current AM2 motherboards after a simple BIOS update, while also working in new motherboards with an AM2+ socket. Additionally, future AM3 CPUs, which will come with both DDR2 and DDR3 support, will work in all of the aforementioned sockets.

The reason for this is that the pin counts are exactly the same – the later sockets just introduce more features into the fray, like HT3.0 and dual power planes with socket AM2+, and DDR3 support with socket AM3.

"AMD's 65nm process is taking longer than expected to mature."

I couldn’t help but ask why AMD opted to put so much effort into consolidating socket AM2 through 3, but pushed socket 939 customers out into the cold. In order for 939 users to upgrade, they’d need to not only buy a new CPU, but also a new motherboard and new memory - pushing the price out of reach for next to no performance benefit. With Intel’s Core microarchitecture following so close after the relatively lacklustre socket AM2 launch, I’m pretty sure most enthusiasts either didn’t upgrade, or upgraded to a system based on Intel’s new CPU architecture.

I was greeted by some rather puzzled looks from AMD’s execs, and I got a two part answer to my question. The first came from Dave Everitt, EMEA Product and Platforms Manager at AMD. He said that “in order to get the bandwidth required for HyperTransport 3.0, we needed a few more pins. The way we’d pinned it out [in the past] meant that was how it turned out to be.”

On its own, that was a pretty feasible answer. I don’t quite know how far out AMD was working on DDR3, but I’ll assume that it wasn’t before the socket 939 pin layout had already been finalised.

The second part of the answer I got came from Jochen Polster, Sales and Marketing Vice President for Worldwide Partner Infrastructure Products at AMD. “We had to make the switch somewhere, right?” expressed Polster. “We decided to concentrate on AM2 because we couldn’t consolidate into one of the older [sockets].”

As Dave highlighted, there were quite possibly technical reasons behind this need to move to a newer socket - but I found Jochen’s answer to be a bit strange in some ways. To me, it felt like AMD had realised the mistake it made with the transition from socket 939 to socket AM2. The biggest reason for me saying this is that socket AM3 processors will not only work in socket AM3 motherboards, but also older socket AM2 and socket AM2+ boards. That’s a good thing – as AMD isn’t suddenly going to alienate the AMD-faithful this time when they go out to upgrade their current AM2 systems in the next year and a half.

"...it felt like AMD had realised the mistake it made with the transition from socket 939 to socket AM2."

On the whole then, I think you should be able to gather that I’ve not been particularly impressed with the launch of Phenom thus far, but I think that there’s definitely still time for AMD to redeem itself in the enthusiast crowd as the chip, and more importantly the technology behind the chip, matures.

In some ways, I don’t know what to draw from this launch, because if you add this somewhat underwhelming launch to what has happened to AMD in the past eighteen months or so, there’s been quite a bizarre series of almost unfortunate events...to the point that all I could conclude was that AMD’s not had the rub of the green for much of this year. Some of this is the company's own doing, and the rest is down to the fact that both Intel and Nvidia have continually executed their roadmaps almost to perfection.

I just hope that there's a light at the end of this tunnel, because my magic eight ball keeps saying "outlook hazy." It's hardly the phenomenal launch that we were expecting, and I hope that AMD isn't looking for it to be considered that way - no matter what they named the chip.

Anyway, regardless of this last event, we're eagerly looking forward to testing Phenom on our own terms and we'll be back to report more experiences with the processors as soon as we can get one (or more) of the chips into our labs. Additionally, our slightly delayed Radeon HD 3870 and 3850 reviews will follow very shortly too, so stay tuned for those.