When Will Nvidia's New GPUs Arrive?

Summary

• Probably not in time for Nvidia's GPU Technology Conference.
• Nvidia maximizes profit at the expense of market share.
• Jon Peddie gets it wrong on crypto GPU sales.
• Revisiting the crypto GPU market split.

Rethink Technology business briefs for March 2, 2018.

Nvidia's GTC page, source: Nvidia.

Nvidia's (NASDAQ:NVDA) GPU (Graphics Processing Unit) Technology Conference is coming up later this month, on March 26. GTC has usually been the occasion to announce new GPU architectures, even if new products are months away. There has been much speculation that Nvidia would unveil a successor to Pascal, the current generation of GPUs used in Nvidia's consumer products.

The latest and perhaps most authoritative report comes from Tom's Hardware. Igor Wallassek reports that "nothing concrete" is likely to be revealed. This is unfortunately consistent with what CEO Jensen Huang stated at the fiscal 2018 Q4 conference call on February 8:

...we haven't announced anything for April or July. And so the best way to think about that is, Pascal is the best gaming platform on the planet... we expect Pascal to continue to be world's best gaming platform for the foreseeable future.

One might wonder just how long the "foreseeable future" is, but it was clear that the rollout of a new consumer graphics architecture wasn't imminent. The Tom's report left open the possibility that next-generation cards might be unveiled sometime in mid summer.

The report also indicated a continued divergence between datacenter/professional and consumer cards, with Ampere being the successor to Volta and Turing the successor to Pascal. If true, the nomenclature is somewhat ironic given Turing's interest in artificial intelligence.

I've been expecting this divergence as a necessary optimization of the GPU architecture for machine learning applications in order to fend off challenges from alternatives such as ASICs and competing GPUs.

Nvidia maximizes profit at the expense of market share

Delaying a new architecture rollout makes sense in terms of maximizing profit on the existing Pascal platform, given the lack of supply/demand balance. During the Q4 call, Nvidia indicated that it was still working to achieve that balance. As long as demand is so high from cryptomining and gaming, it makes sense to simply feed that demand with Pascal cards.

The strategy has worked so far, with the Gaming market segment (which includes some crypto) revenue growing 29% y/y to $1.739 billion in fiscal Q4. Assuming that Gaming had at least the same operating margin as the corporate average, then Gaming operating profit was $641 million, growing sequentially by 21%.

The strategy also makes sense in light of the roadmap that Advanced Micro Devices (AMD) laid out at its Tech Day press conference at CES:

Source: AnandTech

AnandTech, in its coverage of the event, confirmed that there would be no 12 nm Vega this year:

AMD's GPU line will be bypassing the 12nm process used by the Zen+ CPUs, and moving straight to 7nm with Vega. This is a distinct change from previous AMD roadmaps, which had confirmed that we would be seeing Vega on a "14nm+" process node, or Vega on the 12nm process. It would appear that the "12nm" node is the new name for 14+, but AMD's GPU product line will now be bypassing this completely, despite being "confirmed" before.

As mentioned above, AMD plans to start sampling a machine learning focused version of Vega on 7nm much later in 2018, where "sampling" in this case means pre-testing to select partners.

And even Ryzen APUs that include Vega graphics will be fabricated only on 14 nm this year. Thus, the lack of new competition from AMD also seems to mandate a go-slow approach to a Pascal successor.

The downside of the strategy appears to be that it opens the door to AMD to make further gains as it did in calendar Q4, as recently reported by Jon Peddie research. Although there was a sequential decline in Q4, total desktop add-in board sales were still a healthy 15.54 million.

AMD's market share performance was the best since 2014 Q1:

Jon Peddie gets it wrong on crypto GPU sales

Although I regard Peddie as generally reliable, Peddie makes a statement about add-in board sales for mining that's simply not supportable:

Over three million add-in boards (AIBs) were sold to cryptocurrency miners worth $776 million in 2017.

Those of you who have been following the discussion, including my own article on Nvidia's cryptomining exposure, will recognize that Peddie has to be a gross underestimate. Although I have my points of disagreement with SA contributor Akram's Razor, his relating of mining network expansion to GPU demand is both creative and insightful. Razor's estimate that "5.17 million GTX equivalents" had come online in mining networks in the past three months indicates the magnitude of the Peddie error.

My own analysis of cryptomining GPU demand indicates that about 2.736 million cards were added in calendar Q4 based on hash rate growth in the Ethereum mining network alone.

Growth in hash rate capacity for Ethereum, Monero, and Zcash. Source: bitinfocharts.

Between October 1, 2017, and December 31, 2017, the hash rate grew by 55 Terahash/sec. Assuming an average mining card capability of about 20 Megahash/sec results in the 2.736 million card number. And by Q4, it's likely that most of that growth was supplied by new cards, since suitably used cards already in circulation would have been committed to mining in the previous quarters.

There's probably room for debate about whether the 20 Mh/s number is the right one, but it's pretty typical for popular mining cards. In a recent review by Tom's Hardware of 15 popular mining cards (both Nvidia and AMD based), the average hash rate was 20.6 for the 14 cards for which Tom's provided complete data.

My estimate is based on taking only the top ten cards in terms of total cost of ownership from Tom's group of 15, assuming that these would be the most popular. TCO, in this case, was estimated from purchase price listed by Tom's, along with electricity cost for a year. If you've done any mining, and I have, you know that electricity is a major ongoing expense that directly affects mining profitability.

Revisiting the crypto GPU market split

It still remains an interesting question how the mining card market was split between Nvidia and AMD. The top ten cards from the Tom's review were the ones that offered the numerically lowest ratio of TCO/(Mh/s), i.e. the best value. Six of these were AMD based and four were Nvidia based. This is how I concluded that unit sales are split approximately 60% for AMD and 40% for Nvidia. Four out of the top five were AMD, but I assume that in a scarcity scenario, people will accept less desirable cards just to have something that works. Thus I believe that the 60/40 split is approximately correct.

The retail cost of the added Ethereum capacity in Q4 I estimate to be $852.9 million. This is just the added hash capacity multiplied by the average retail price/(hash capacity) for the top ten cards. This was calculated to be $15.49/(Mh/s). Once again, this is clearly larger than the Peddie estimate for the entire year.

Based on the respective prices for the AMD and Nvidia cards in the top ten, Nvidia's retail revenue share is estimated to be $420.9 million and AMD's to be $432 million. Once again, assuming that Nvidia and AMD only receive about 35% of the retail price as chip suppliers to the add-in board manufacturers, then Nvidia's direct revenue was $147.3 million and AMD's was $151.2 million in Q4.

Nvidia's Ethereum mining revenue was about 8.5% of its total gaming segment revenue for its fiscal Q4, while AMD's Ethereum mining revenue was about 31.5% of its Q4 GPU revenue, assuming that GPUs made up about half of the revenue of its Computing and Graphics segment.

Some have inferred that this conclusion must be at odds with Lisa Su's statement regarding cryptomining revenue in the Q4 conference call. Here's what Su said:

So, look on the Computing and Graphic segment, we grew about $140 million sequentially. And if I look at that growth, it was across Ryzen and Radeon. If you look at block chain in particular, our estimates are that it was about a third of the growth, a third of the $140 million. And then the rest of the two thirds are around the GPUs, the other segments of GPUs and Ryzen.

Su is only referring to the $139 million sequential change in revenue for the Computing and Graphics segment. She makes no statement regarding the percentage of total C&G revenue for the quarter, which was $958 million. It's not reasonable to assume that the balance of revenue for Q4 less the increase, $819 million, has no component due to cryptomining.

In fact, I believe my estimate that Ethereum crypto-mining contributed about 15% of total C&G revenue is probably conservative. First, the estimate doesn't include contributions from other currencies such as Zcash and Monero, although Ethereum is the biggest contributor to GPU mining demand and probably accounts for 80% of it.

Second, I probably have underestimated the popularity of AMD cards for mining. AMD supporters have extolled the superior value offered by AMD cards for cryptomining, with some justification. But at the same time, AMD supporters have chosen to minimize the potential exposure of AMD to a collapse in mining GPU demand. This amounts to having your cake and eating it, too.

Analyst’s Disclosure: I am/we are long NVDA. I wrote this article myself, and it expresses my own opinions. I am not receiving compensation for it (other than from Seeking Alpha). I have no business relationship with any company whose stock is mentioned in this article.