AMD’s Financial Analyst Day last week answered some questions about the company’s long-term roadmap and plans for projects like Vega, Ryzen, and its new server platform, Epyc. Many of you still had questions, however, about the timing and nature of some of the updates we would see later on this year, when Vega might launch for consumer cards, and what’s next for AMD.
AMD CEO Lisa Su spoke at the JP Morgan Media Telecom Conference earlier this week, and gave some additional color on these questions. Let’s break it down:
Desktop, server, and future foundry plans
During AMD’s financial analyst day, it unveiled a roadmap for Zen that showed 14nm/14nm+, followed by a Zen 2 and Zen 3. Lisa Su confirmed that Zen 2 would be a 2018 or 2019 part (likely depending on GlobalFoundries’ own timetable for 7nm rollouts). This implies we’ll see a refreshed Zen point-update at some point on a second-generation 14nm process.
Exactly which benefits will be baked into the core are unclear. But yearly product cycles have become the norm for AMD as well as a number of other semiconductor firms, and we can safely predict what kind of chip they’ll launch — a modest improvement to Zen, possibly with some internal tweaks and combined with slightly higher clocks or improved TDP. Lisa Su did note they’ll be doing tapeouts for 7nm this year, and tapeout usually leads introduction by 12-18 months depending on the part.
Su also announced that going forward, the data center is the primary focus for AMD. I don’t think that means gamers should give up on Ryzen as a gaming chip — historically, in fact, focusing on the data center gave AMD the best CPU performance and overall market position that it ever enjoyed. Opteron and Athlon 64 / 64 X2 were AMD’s best cores at a time when Intel was performing poorly, with Opteron driving AMD up to nearly 25 percent of the server market over just a few short years.
This is the number one priority, the number one opportunity for us over the next three to four years. I think our priority really is a little bit around time and time it takes to ramp product. So our EPYC CPUs we’ll be launching later this month – I’m sorry, in June and our Vega GPUs also in a similar timeframe. And what we will be doing is working closely with our customers.
AMD intends to engage with cloud providers with gradual ramping across the entire server market. Lisa Su claims that the ramp will be somewhat slower than you might see in PCs, and AMD has set a goal for itself of capturing 10 percent of the data center market over the next few years. She also noted that moving to a multi-chip module (MCM) approach on Epyc has improved cost structures and yields because it doesn’t require AMD to hit a perfect 32-core microprocessor in a single die. Connecting four eight-core chips via MCM has been the better strategy.
Licensing and IP monetization
Su also spoke to AMD’s plans for licensing agreements and its own IP monetization, and indirectly confirmed what Intel has already said — there’s no plan to license AMD graphics technology to a major competitor in that kind of licensing agreement. She explicitly states, “We’re not trying to become an ARM or an Imagination…And so it’s about how do we sell more products or how do we have our IP in markets that we’re not currently selling products. And so that’s kind of how we think about it. We’re not looking at enabling a competitor to compete against our products.”
Between her statements and Intel’s, it’s safe to dismiss this rumor. An AMD-designed Intel integrated GPU might have been interesting, but it clearly isn’t happening. She did note that AMD had a new, non-game console, semi-custom design it expected to produce revenue from 2018 – 2020, and that its license agreement with THATIC (the joint venture it formed last year) was also going well and producing revenue on-schedule.
Vega’s consumer launch date
Lisa Su didn’t give as much detail as some might like regarding AMD’s Vega plans, but she did speak to them, saying:
Yes, so we’re very excited about Vega. Vega is a brand new architecture and actually has really the world’s most advanced memory architecture for CPUs. What we did announce at our Analyst Day is that the first shipping Vega will be the frontier editions, which will ship with 16 gigabytes of memory and it will ship towards the latter half of June. You will see the enthusiast gaming platform, the machine learning platform, the professional graphics platform very soon thereafter. And so we will be launching Vega across all of the market segments over the next couple of months.
Here’s how I’d parse that statement. As per AMD’s previous comments, the only Vega shipping in June will be the Frontier edition. Lisa Su is saying that Vega will launch across all platforms against the next couple of months, which could be read several different ways. If she means “The next couple of months after June,” then we’re looking at consumer launches in August and September (assuming “couple” is read loosely). If we read it more tightly, then we could see launches earlier, in July – August. But there’s very much a question as to where AMD will focus its efforts, and how it will price Vega.
All of the following is conjecture and based on no inside information from within AMD, but the company has sent some significant signs that its focus with Vega will not be on the consumer market. First, there’s the fact that Lisa Su has specifically identified data centers and machine learning as major foci for AMD. These are the markets where high-end GPUs truly make bank — a consumer high-end card may sell for $800, but the professional version of that same GPU could be a $1,600 – $3,200 part. Seriously, check Nvidia’s Quadro pricing against the same GPUs in consumer parts, or the gap between AMD’s FirePro and Radeon. Workstation cards do sometimes field certain capabilities that consumer cards lack, but nothing that remotely justifies the huge price premium between the two.
But given the issues that seem to be piling up around HBM2, it’s just not clear how low AMD’s Vega pricing will be, or where the company will introduce the GPU. Much, of course, depends on performance. Here, I think the naysayers are selling the company short. We watched doubts pile up around Ryzen for years before its introduction, only for the final chip to be a far stronger overall performer than many (including myself), expected — and I expected Ryzen to be a vast improvement over AMD’s Bulldozer-derived architectures. I’m not going to make any assertions or claims about how Pascal or Volta will compare, because there’s just not enough information to speak to that point.
If AMD starts pushing Vega into consumer markets in July and can keep the card on store shelves at reasonable prices, it’ll be a sign that whatever production issues may have occurred, they’ve been ironed out. But if it continues to focus on the highest end of the market, or to emphasize data centers and machine learning, it could mean the GPU will only see limited release to consumers. We won’t know which way the situation will break until AMD releases more concrete information or launches the GPU.